Merge branch 'upstream' into concedo_experimental

# Conflicts:
#	.github/ISSUE_TEMPLATE/010-bug-compilation.yml
#	.github/ISSUE_TEMPLATE/011-bug-results.yml
#	.github/labeler.yml
#	.github/workflows/build.yml
#	.github/workflows/release.yml
#	.gitmodules
#	CMakeLists.txt
#	ggml/CMakeLists.txt
#	ggml/src/CMakeLists.txt
#	ggml/src/ggml-cann/aclnn_ops.cpp
#	ggml/src/ggml-cann/ggml-cann.cpp
#	ggml/src/ggml-opencl/ggml-opencl.cpp
#	ggml/src/ggml-opencl/kernels/softmax_4_f16.cl
#	ggml/src/ggml-opencl/kernels/softmax_4_f32.cl
#	ggml/src/ggml-opencl/kernels/softmax_f16.cl
#	ggml/src/ggml-opencl/kernels/softmax_f32.cl
#	ggml/src/ggml-sycl/element_wise.cpp
#	ggml/src/ggml-sycl/element_wise.hpp
#	ggml/src/ggml-sycl/ggml-sycl.cpp
#	scripts/sync-ggml-am.sh
#	scripts/sync-ggml.last
#	scripts/sync-ggml.sh
#	tests/test-backend-ops.cpp
#	tests/test-c.c

convert_hf_to_gguf.py

@@ -4408,9 +4408,6 @@ class Gemma3NModel(Gemma3Model):
         ]
 
     def set_vocab(self):
-        with open(self.dir_model / "chat_template.jinja") as f:
-            # quick hack to make sure chat template is added
-            self.gguf_writer.add_chat_template(f.read())
         super().set_vocab()
 
     def set_gguf_parameters(self):
@@ -4781,6 +4778,14 @@ class ARwkv7Model(Rwkv7Model):
 class MambaModel(TextModel):
     model_arch = gguf.MODEL_ARCH.MAMBA
 
+    def __init__(self, dir_model: Path, *args, **kwargs):
+        # Avoid using AutoConfig for hparams
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+                hparams = json.load(f)
+        super().__init__(dir_model, *args, hparams=hparams, **kwargs)
+
     def set_vocab(self):
         vocab_size = self.hparams["vocab_size"]
         # Round vocab size to next multiple of 8
@@ -4855,6 +4860,100 @@ class MambaModel(TextModel):
         return [(new_name, data_torch)]
 
 
+@ModelBase.register("Mamba2ForCausalLM")
+class Mamba2Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.MAMBA2
+
+    def __init__(self, dir_model: Path, *args, **kwargs):
+        # Avoid using AutoConfig for hparams
+        # It wrongly assumes all Mamba2 models are Mamba-Codestral-7B-v0.1
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+                hparams = json.load(f)
+        super().__init__(dir_model, *args, hparams=hparams, **kwargs)
+
+    def set_vocab(self):
+        vocab_size = self.hparams["vocab_size"]
+        # Round vocab size to next multiple of 16
+        pad_vocab = self.hparams.get("pad_vocab_size_multiple", 16)
+        # pad using ceiling division
+        # ref: https://stackoverflow.com/a/17511341/22827863
+        vocab_size = -(vocab_size // -pad_vocab) * pad_vocab
+        self.hparams["vocab_size"] = vocab_size
+
+        if (self.dir_model / "tokenizer.model").is_file():
+            self._set_vocab_sentencepiece()
+        elif (self.dir_model / "tokenizer.model.v3").is_file():
+            # mamba-codestral
+            raise NotImplementedError(f"Please rename {self.dir_model / 'tokenizer.model.v3'} to {self.dir_model / 'tokenizer.model'}")
+        elif (self.dir_model / "tokenizer.json").is_file():
+            self._set_vocab_gpt2()
+        else:
+            # Use the GPT-NeoX tokenizer when no tokenizer files are present
+            self._set_vocab_builtin("gpt-neox", vocab_size)
+
+    def set_gguf_parameters(self):
+        d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
+        d_conv  = self.find_hparam(["conv_kernel",       "d_conv"],  optional=True) or 4
+        d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model
+        d_state = self.find_hparam(["state_size",        "d_state"], optional=True) or 128
+        head_dim = self.find_hparam(["head_dim"],                    optional=True) or 64
+        n_group = self.find_hparam(["n_groups"],                     optional=True) or 1
+
+        rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
+
+        # Fail early for models which don't have a block expansion factor of 2
+        # TODO: does this really matter?
+        assert d_inner == 2 * d_model
+        assert d_inner % head_dim == 0
+
+        self.gguf_writer.add_context_length(2**20)  # arbitrary value; for those who use the default
+        self.gguf_writer.add_embedding_length(d_model)
+        self.gguf_writer.add_feed_forward_length(0)  # unused, but seemingly required when loading
+        self.gguf_writer.add_head_count(0)  # unused, but seemingly required when loading
+        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_ssm_conv_kernel(d_conv)
+        self.gguf_writer.add_ssm_inner_size(d_inner)
+        self.gguf_writer.add_ssm_state_size(d_state)
+        self.gguf_writer.add_ssm_time_step_rank(d_inner // head_dim)
+        self.gguf_writer.add_ssm_group_count(n_group)
+        self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+
+        if name.startswith("model.backbone") or name.startswith("model.lm_head"):
+            # map Mamba-Codestral-7B-v0.1 tensor names to the names used by Mamba-2
+            name = name.removeprefix("model.")
+
+        if name.endswith(".dt_bias"):
+            name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"
+
+        new_name = self.map_tensor_name(name)
+
+        if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_CONV1D, bid):
+            data_torch = data_torch.squeeze()
+        elif any(self.match_model_tensor_name(new_name, t, bid, suffix="") for t in [
+            gguf.MODEL_TENSOR.SSM_A,
+            gguf.MODEL_TENSOR.SSM_D,
+        ]):
+            # unsqueeze A to use similar shape semantics as Mamba-1
+            # (D is also unsqueezed, but for more straightforward broadcast internally)
+            data_torch = data_torch.reshape((*data_torch.shape, 1))
+        elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_NORM, bid):
+            d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
+            d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model
+            n_group = self.hparams.get("n_groups", 1)
+            data_torch = data_torch.reshape((n_group, d_inner // n_group))
+
+        if name.endswith(".A_log"):
+            logger.debug("A_log --> A ==> " + new_name)
+            data_torch = -torch.exp(data_torch)
+
+        yield (new_name, data_torch)
+
+
 @ModelBase.register("CohereForCausalLM")
 class CommandR2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.COMMAND_R
@@ -6615,12 +6714,20 @@ def get_model_architecture(hparams: dict[str, Any], model_type: ModelType) -> st
     # maybe we should fallback to text model's arch in that case, since not many models have both
     text_config = hparams.get("text_config", {})
     vision_config = hparams.get("vision_config", {})
-    arch = hparams["architectures"][0]
+    arch = None
+    if (arches := hparams.get("architectures")) is not None and len(arches) > 0:
+        arch = arches[0]
+    elif "ssm_cfg" in hparams:
+        # For non-hf Mamba and Mamba2 models
+        arch = hparams["ssm_cfg"].get("layer", "Mamba") + "ForCausalLM"
+
     # if "architectures" is found in the sub-config, use that instead
     if model_type == ModelType.TEXT and text_config.get("architectures") is not None:
         arch = text_config["architectures"][0]
     elif model_type == ModelType.MMPROJ and vision_config.get("architectures") is not None:
         arch = vision_config["architectures"][0]
+    if arch is None:
+        raise ValueError("Failed to detect model architecture")
     return arch
 
 

ggml/include/ggml-kompute.h

@@ -1,50 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#include <stdbool.h>
-#include <stddef.h>
-#include <stdint.h>
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#define GGML_KOMPUTE_MAX_DEVICES 16
-
-struct ggml_vk_device {
-    int index;
-    int type; // same as VkPhysicalDeviceType
-    size_t heapSize;
-    const char * name;
-    const char * vendor;
-    int subgroupSize;
-    uint64_t bufferAlignment;
-    uint64_t maxAlloc;
-};
-
-struct ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count);
-bool ggml_vk_get_device(struct ggml_vk_device * device, size_t memoryRequired, const char * name);
-bool ggml_vk_has_vulkan(void);
-bool ggml_vk_has_device(void);
-struct ggml_vk_device ggml_vk_current_device(void);
-
-//
-// backend API
-//
-
-// forward declaration
-typedef struct ggml_backend * ggml_backend_t;
-
-GGML_BACKEND_API ggml_backend_t ggml_backend_kompute_init(int device);
-
-GGML_BACKEND_API bool ggml_backend_is_kompute(ggml_backend_t backend);
-
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
-
-GGML_BACKEND_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
-
-#ifdef __cplusplus
-}
-#endif

ggml/include/ggml.h

@@ -563,6 +563,8 @@ extern "C" {
         GGML_GLU_OP_REGLU,
         GGML_GLU_OP_GEGLU,
         GGML_GLU_OP_SWIGLU,
+        GGML_GLU_OP_GEGLU_ERF,
+        GGML_GLU_OP_GEGLU_QUICK,
 
         GGML_GLU_OP_COUNT,
     };
@@ -659,6 +661,9 @@ extern "C" {
 
     // misc
 
+    GGML_API const char * ggml_version(void);
+    GGML_API const char * ggml_commit(void);
+
     GGML_API void    ggml_time_init(void); // call this once at the beginning of the program
     GGML_API int64_t ggml_time_ms(void);
     GGML_API int64_t ggml_time_us(void);
@@ -1157,6 +1162,22 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_geglu_erf(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_erf_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_quick(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_quick_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // A: n columns, r rows,
     // B: n columns, r rows,
     GGML_API struct ggml_tensor * ggml_glu_split(
@@ -1180,6 +1201,16 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_geglu_erf_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_geglu_quick_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     // normalize along rows
     GGML_API struct ggml_tensor * ggml_norm(
             struct ggml_context * ctx,
@@ -1523,8 +1554,14 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // a    [ne0, ne01, ne02, ne03]
+    // mask [ne0, ne11, ne12, ne13] | ne11 >= ne01, F16 or F32, optional
+    //
+    // broadcast:
+    //   ne02 % ne12 == 0
+    //   ne03 % ne13 == 0
+    //
     // fused soft_max(a*scale + mask*(ALiBi slope))
-    // mask is optional
     // max_bias = 0.0f for no ALiBi
     GGML_API struct ggml_tensor * ggml_soft_max_ext(
             struct ggml_context * ctx,
@@ -1987,11 +2024,17 @@ extern "C" {
 
 #define GGML_KQ_MASK_PAD 64
 
-    // q:    [n_embd_k, n_batch,     n_head,    1]
+    // q:    [n_embd_k, n_batch,     n_head,    ne3 ]
-    // k:    [n_embd_k, n_kv,        n_head_kv, 1]
+    // k:    [n_embd_k, n_kv,        n_head_kv, ne3 ]
-    // v:    [n_embd_v, n_kv,        n_head_kv, 1] !! not transposed !!
+    // v:    [n_embd_v, n_kv,        n_head_kv, ne3 ] !! not transposed !!
-    // mask: [n_kv,     n_batch_pad, 1,         1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
+    // mask: [n_kv,     n_batch_pad, ne32,      ne33] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
-    // res:  [n_embd_v, n_head,      n_batch,   1] !! permuted !!
+    // res:  [n_embd_v, n_head,      n_batch,   ne3 ] !! permuted !!
+    //
+    // broadcast:
+    //   n_head % n_head_kv == 0
+    //   n_head % ne32      == 0
+    //   ne3    % ne33      == 0
+    //
     GGML_API struct ggml_tensor * ggml_flash_attn_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * q,
@@ -2030,7 +2073,8 @@ extern "C" {
             struct ggml_tensor  * dt,
             struct ggml_tensor  * A,
             struct ggml_tensor  * B,
-            struct ggml_tensor  * C);
+            struct ggml_tensor  * C,
+            struct ggml_tensor  * ids);
 
     // partition into non-overlapping windows with padding if needed
     // example:

ggml/src/ggml-backend-reg.cpp

@@ -61,10 +61,6 @@
 #include "ggml-cann.h"
 #endif
 
-#ifdef GGML_USE_KOMPUTE
-#include "ggml-kompute.h"
-#endif
-
 // disable C++17 deprecation warning for std::codecvt_utf8
 #if defined(__clang__)
 #    pragma clang diagnostic push
@@ -189,9 +185,6 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_RPC
         register_backend(ggml_backend_rpc_reg());
 #endif
-#ifdef GGML_USE_KOMPUTE
-        register_backend(ggml_backend_kompute_reg());
-#endif
 #ifdef GGML_USE_CPU
         register_backend(ggml_backend_cpu_reg());
 #endif
@@ -576,7 +569,6 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
     ggml_backend_load_best("cann", silent, dir_path);
     ggml_backend_load_best("cuda", silent, dir_path);
     ggml_backend_load_best("hip", silent, dir_path);
-    ggml_backend_load_best("kompute", silent, dir_path);
     ggml_backend_load_best("metal", silent, dir_path);
     ggml_backend_load_best("rpc", silent, dir_path);
     ggml_backend_load_best("sycl", silent, dir_path);

ggml/src/ggml-cpu/ggml-cpu.c

@@ -2186,6 +2186,8 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 case GGML_GLU_OP_REGLU:
                 case GGML_GLU_OP_GEGLU:
                 case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
                     {
                         n_tasks = n_threads;
                     } break;

ggml/src/ggml-cpu/ops.cpp

@@ -3614,6 +3614,292 @@ static void ggml_compute_forward_swiglu(
     }
 }
 
+// ggml_compute_forward_geglu_erf
+
+static void ggml_compute_forward_geglu_erf_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        float * src0_p = (float *) (src0_d + i1*src0_o);
+        float * src1_p = (float *) (src1_d + i1*src1_o);
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        ggml_vec_geglu_erf_f32(nc, (float *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);
+
+#ifndef NDEBUG
+        for (int k = 0; k < nc; k++) {
+            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            GGML_UNUSED(x);
+            assert(!isnan(x));
+            assert(!isinf(x));
+        }
+#endif
+    }
+}
+
+static void ggml_compute_forward_geglu_erf_f16(
+    const ggml_compute_params * params,
+    ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        ggml_fp16_t * src0_p = (ggml_fp16_t *) (src0_d + i1*src0_o);
+        ggml_fp16_t * src1_p = (ggml_fp16_t *) (src1_d + i1*src1_o);
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        ggml_vec_geglu_erf_f16(nc, (ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);
+
+#ifndef NDEBUG
+        for (int k = 0; k < nc; k++) {
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float v = GGML_FP16_TO_FP32(x);
+            GGML_UNUSED(v);
+            assert(!isnan(v));
+            assert(!isinf(v));
+        }
+#endif
+    }
+}
+
+static void ggml_compute_forward_geglu_erf(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_geglu_erf_f32(params, dst);
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_geglu_erf_f16(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+// ggml_compute_forward_geglu_quick
+
+static void ggml_compute_forward_geglu_quick_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        float * src0_p = (float *) (src0_d + i1*src0_o);
+        float * src1_p = (float *) (src1_d + i1*src1_o);
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        ggml_vec_geglu_quick_f32(nc, (float *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);
+
+#ifndef NDEBUG
+        for (int k = 0; k < nc; k++) {
+            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            GGML_UNUSED(x);
+            assert(!isnan(x));
+            assert(!isinf(x));
+        }
+#endif
+    }
+}
+
+static void ggml_compute_forward_geglu_quick_f16(
+    const ggml_compute_params * params,
+    ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        ggml_fp16_t * src0_p = (ggml_fp16_t *) (src0_d + i1*src0_o);
+        ggml_fp16_t * src1_p = (ggml_fp16_t *) (src1_d + i1*src1_o);
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        ggml_vec_geglu_quick_f16(nc, (ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);
+
+#ifndef NDEBUG
+        for (int k = 0; k < nc; k++) {
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float v = GGML_FP16_TO_FP32(x);
+            GGML_UNUSED(v);
+            assert(!isnan(v));
+            assert(!isinf(v));
+        }
+#endif
+    }
+}
+
+static void ggml_compute_forward_geglu_quick(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_geglu_quick_f32(params, dst);
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_geglu_quick_f16(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_norm
 
 static void ggml_compute_forward_norm_f32(
@@ -5232,14 +5518,17 @@ static void ggml_compute_forward_soft_max_f32(
     memcpy(&scale,    (float *) dst->op_params + 0, sizeof(float));
     memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
 
-    // TODO: handle transposed/permuted matrices
-
     const int ith = params->ith;
     const int nth = params->nth;
 
     GGML_TENSOR_UNARY_OP_LOCALS
 
-    //const int64_t ne11 = src1 ? src1->ne[1] : 1;
+    const int64_t nb11 = src1 ? src1->nb[1] : 1;
+    const int64_t nb12 = src1 ? src1->nb[2] : 1;
+    const int64_t nb13 = src1 ? src1->nb[3] : 1;
+
+    const int64_t ne12 = src1 ? src1->ne[2] : 1;
+    const int64_t ne13 = src1 ? src1->ne[3] : 1;
 
     // TODO: is this supposed to be ceil instead of floor?
     //       https://huggingface.co/mosaicml/mpt-7b/blob/main/attention.py#L370
@@ -5249,69 +5538,67 @@ static void ggml_compute_forward_soft_max_f32(
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
-    const int nc = src0->ne[0];
+    float * wp = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;
-    const int nr = ggml_nrows(src0);
-
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
-
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-
-    float * wp = (float *) params->wdata + (nc + CACHE_LINE_SIZE_F32) * ith;
 
     const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
+                const int64_t i11 = i01;
+                const int64_t i12 = i02%ne12;
+                const int64_t i13 = i03%ne13;
+
                 // ALiBi
-        const uint32_t h = (i1/ne01)%ne02; // head
+                const uint32_t h = i02; // head
                 const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f;
 
-        float * sp = (float *)((char *) src0->data + i1*src0->nb[1]);
+                float * sp = (float *)((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
-        float * dp = (float *)((char *)  dst->data +  i1*dst->nb[1]);
+                float * dp = (float *)((char *)  dst->data + i01*nb1  + i02*nb2  + i03*nb3);
 
                 // broadcast the mask across rows
-        ggml_fp16_t * mp_f16 = src1 ? (ggml_fp16_t *)((char *) src1->data) + (i1%ne01)*ne00 : NULL;
+                ggml_fp16_t * mp_f16 = src1 ? (ggml_fp16_t *)((char *) src1->data + i11*nb11 + i12*nb12 + i13*nb13) : NULL;
-        float       * mp_f32 = src1 ? (float       *)((char *) src1->data) + (i1%ne01)*ne00 : NULL;
+                float       * mp_f32 = src1 ? (float       *)((char *) src1->data + i11*nb11 + i12*nb12 + i13*nb13) : NULL;
 
-        ggml_vec_cpy_f32  (nc, wp, sp);
+                ggml_vec_cpy_f32  (ne00, wp, sp);
-        ggml_vec_scale_f32(nc, wp, scale);
+                ggml_vec_scale_f32(ne00, wp, scale);
                 if (mp_f32) {
                     if (use_f16) {
-                for (int i = 0; i < nc; ++i) {
+                        for (int i = 0; i < ne00; ++i) {
                             wp[i] += slope*GGML_CPU_FP16_TO_FP32(mp_f16[i]);
                         }
                     } else {
-                for (int i = 0; i < nc; ++i) {
+                        for (int i = 0; i < ne00; ++i) {
                             wp[i] += slope*mp_f32[i];
                         }
                     }
                 }
 
 #ifndef NDEBUG
-        for (int i = 0; i < nc; ++i) {
+                for (int i = 0; i < ne00; ++i) {
                     //printf("p[%d] = %f\n", i, p[i]);
                     assert(!isnan(wp[i]));
                 }
 #endif
 
                 float max = -INFINITY;
-        ggml_vec_max_f32(nc, &max, wp);
+                ggml_vec_max_f32(ne00, &max, wp);
 
-        ggml_float sum = ggml_vec_soft_max_f32(nc, dp, wp, max);
+                ggml_float sum = ggml_vec_soft_max_f32(ne00, dp, wp, max);
                 assert(sum > 0.0);
 
                 sum = 1.0/sum;
-        ggml_vec_scale_f32(nc, dp, sum);
+                ggml_vec_scale_f32(ne00, dp, sum);
 
 #ifndef NDEBUG
-        for (int i = 0; i < nc; ++i) {
+                for (int i = 0; i < ne00; ++i) {
                     assert(!isnan(dp[i]));
                     assert(!isinf(dp[i]));
                 }
 #endif
             }
+        }
+    }
 }
 
 void ggml_compute_forward_soft_max(
@@ -7798,7 +8085,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
             memset(VKQ32, 0, DV*sizeof(float));
         }
 
-        const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1]) : NULL;
+        const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1] + (iq2%mask->ne[2])*mask->nb[2] + (iq3%mask->ne[3])*mask->nb[3]) : NULL;
 
         // k indices
         const int ik3 = iq3 / rk3;
@@ -8336,120 +8623,210 @@ void ggml_compute_forward_ssm_conv(
 static void ggml_compute_forward_ssm_scan_f32(
         const ggml_compute_params * params,
         ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0]; // s
+    const ggml_tensor * src0 = dst->src[0]; // s  {d_state, dim, n_head, n_seqs+}
-    const ggml_tensor * src1 = dst->src[1]; // x
+    const ggml_tensor * src1 = dst->src[1]; // x  {dim, n_head, n_seq_tokens, n_seqs}
-    const ggml_tensor * src2 = dst->src[2]; // dt
+    const ggml_tensor * src2 = dst->src[2]; // dt {n_head, n_seq_tokens, n_seqs}
-    const ggml_tensor * src3 = dst->src[3]; // A
+    const ggml_tensor * src3 = dst->src[3]; // A  {d_state, n_head} or {1, n_head}
-    const ggml_tensor * src4 = dst->src[4]; // B
+    const ggml_tensor * src4 = dst->src[4]; // B  {d_state, n_group, n_seq_tokens, n_seqs}
-    const ggml_tensor * src5 = dst->src[5]; // C
+    const ggml_tensor * src5 = dst->src[5]; // C  {d_state, n_group, n_seq_tokens, n_seqs}
+    const ggml_tensor * src6 = dst->src[6]; // ids {n_seqs}
 
     const int ith = params->ith;
     const int nth = params->nth;
 
     const int64_t nc = src0->ne[0]; // d_state
-    const int64_t nr  = src0->ne[1]; // d_inner
+    const int64_t nr = src0->ne[1]; // dim
-    const int64_t n_t = src1->ne[1]; // number of tokens per sequence
+    const int64_t nh = src1->ne[1]; // n_head
-    const int64_t n_s = src0->ne[2]; // number of sequences in the batch
+    const int64_t ng = src4->ne[1];
+    const int64_t nt = src1->ne[2]; // number of tokens per sequence
+    const int64_t ns = src1->ne[3]; // number of sequences in the batch
 
-    GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst));
+    // can't use ggml_nbytes because src1 is not necessarily contiguous
+    const int64_t s_off = ggml_nelements(src1) * ggml_element_size(src1);
+
+    GGML_ASSERT(ggml_nelements(src1) + nc*nr*nh*ns == ggml_nelements(dst));
     GGML_ASSERT(src0->nb[0] == sizeof(float));
     GGML_ASSERT(src1->nb[0] == sizeof(float));
     GGML_ASSERT(src2->nb[0] == sizeof(float));
     GGML_ASSERT(src3->nb[0] == sizeof(float));
     GGML_ASSERT(src4->nb[0] == sizeof(float));
     GGML_ASSERT(src5->nb[0] == sizeof(float));
-    // required for the dot product between s and C
+    GGML_ASSERT(src6->nb[0] == sizeof(int32_t));
-    GGML_ASSERT(src0->nb[1] == src0->ne[0]*sizeof(float));
+    // allows optimizing the modulo since n_group should be a power of 2
-    // required for per-sequence offsets for states
+    GGML_ASSERT((ng & -ng) == ng);
-    GGML_ASSERT(src0->nb[2] == src0->ne[0]*src0->ne[1]*sizeof(float));
-    // required to get correct offset for state destination (i.e. src1->nb[3])
-    GGML_ASSERT(src1->nb[3] == src1->ne[0]*src1->ne[1]*src1->ne[2]*sizeof(float));
 
-    // rows per thread
+    // heads per thread
-    const int dr = (nr + nth - 1)/nth;
+    const int dh = (nh + nth - 1)/nth;
 
-    // row range for this thread
+    // head range for this thread
-    const int ir0 = dr*ith;
+    const int ih0 = dh*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
+    const int ih1 = MIN(ih0 + dh, nh);
-    const int ir  = ir1 - ir0;
 
-    #ifdef __ARM_FEATURE_SVE
+    const int32_t * ids = (const int32_t *) src6->data;
-        for (int i3 = 0; i3 < n_s; ++i3) {
-            for (int i2 = 0; i2 < n_t; ++i2) {
-                const float * s0 = (const float *) ((const char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2])); // {d_state, d_inner, n_s}
-                const float * x  = (const float *) ((const char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
-                const float * dt = (const float *) ((const char *) src2->data + ir0*(src2->nb[0]) + i2*(src2->nb[1]) + i3*(src2->nb[2])); // {d_inner, n_t, n_s}
-                const float * A  = (const float *) ((const char *) src3->data + ir0*(src3->nb[1])); // {d_state, d_inner}
-                const float * B  = (const float *) ((const char *) src4->data +  i2*(src4->nb[1]) + i3*(src4->nb[2])); // {d_state, n_t, n_s}
-                const float * C  = (const float *) ((const char *) src5->data +  i2*(src5->nb[1]) + i3*(src5->nb[2])); // {d_state, n_t, n_s}
-                    float * y  = (      float *) ((      char *) dst->data  + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
-                    float * s  = (      float *) ((      char *) dst->data  + ir0*(src0->nb[1]) + i3*(src0->nb[2]) +     src1->nb[3]);  // {d_state, d_inner, n_s}
 
-                // use the output as the source for the next token-wise iterations
+    for (int i3 = 0; i3 < ns; ++i3) {
-                if (i2 > 0) { s0 = s; }
+        const float * s0 = (const float *) ((const char *) src0->data + ids[i3]*(src0->nb[3])); // {d_state, dim, nh, ns}
+              float * s  = (      float *) ((      char *) dst->data  + i3*(src0->nb[3]) + s_off); // {d_state, dim, nh, ns}
 
-                // d_inner
+        for (int i2 = 0; i2 < nt; ++i2) {
-                for (int i1 = 0; i1 < ir; ++i1) {
+            const float * x  = (const float *) ((const char *) src1->data + i2*(src1->nb[2]) + i3*(src1->nb[3])); // {dim, nh, nt, ns}
-                    float dt_soft_plus = dt[i1] <= 20.0f ? log1pf(expf(dt[i1])) : dt[i1];
+            const float * dt = (const float *) ((const char *) src2->data + i2*(src2->nb[1]) + i3*(src2->nb[2])); // {nh, nt, ns}
-                    float x_dt = x[i1] * dt_soft_plus;
+            const float * A  = (const float *) ((const char *) src3->data); // {d_state, nh} or {1, nh}
+            const float * B  = (const float *) ((const char *) src4->data + i2*(src4->nb[2]) + i3*(src4->nb[3])); // {d_state, ng, nt, ns}
+            const float * C  = (const float *) ((const char *) src5->data + i2*(src5->nb[2]) + i3*(src5->nb[3])); // {d_state, ng, nt, ns}
+                  float * y  = (      float *) ((      char *) dst->data + i2*(nh*nr*sizeof(float)) + i3*(nt*nh*nr*sizeof(float))); // {dim, nh, nt, ns}
+
+            if (src3->ne[0] == 1) {
+                // Mamba-2 has a scalar decay factor per head; dA can be outside the state-wise loop
+
+                // n_head
+                for (int h = ih0; h < ih1; ++h) {
+                    // ref: https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/ops/triton/softplus.py#L16
+                    const float dt_soft_plus = dt[h] <= 20.0f ? log1pf(expf(dt[h])) : dt[h];
+                    const float dA = expf(dt_soft_plus * A[h]);
+
+                    // dim
+                    for (int i1 = 0; i1 < nr; ++i1) {
+                        const int ii = i1 + h*nr;
+                        const float x_dt = x[ii] * dt_soft_plus;
+                        float sumf = 0.0f;
+#if defined(GGML_SIMD)
+    #if defined(__ARM_FEATURE_SVE)
+                        const int ggml_f32_epr = svcntw();
+                        const int ggml_f32_step = 1 * ggml_f32_epr;
+
+                        const int np = (nc & ~(ggml_f32_step - 1));
+
+                        GGML_F32_VEC sum = GGML_F32_VEC_ZERO;
+
+                        GGML_F32_VEC adA = GGML_F32_VEC_SET1(dA);
+                        GGML_F32_VEC axdt = GGML_F32_VEC_SET1(x_dt);
+
+                        for (int i = 0; i < np; i += ggml_f32_step) {
+                            // TODO: maybe unroll more?
+                            for (int j = 0; j < 1; j++) {
+                                GGML_F32_VEC t0 = GGML_F32_VEC_LOAD(s0 + i + j*ggml_f32_epr + ii*nc);
+                                GGML_F32_VEC t1 = GGML_F32_VEC_LOAD(B + i + j*ggml_f32_epr + (h & (ng - 1))*nc);
+                                GGML_F32_VEC t2 = GGML_F32_VEC_LOAD(C + i + j*ggml_f32_epr + (h & (ng - 1))*nc);
+
+                                t0 = GGML_F32_VEC_MUL(t0, adA);
+                                t1 = GGML_F32_VEC_MUL(t1, axdt);
+
+                                t0 = GGML_F32_VEC_ADD(t0, t1);
+
+                                sum = GGML_F32_VEC_FMA(sum, t0, t2);
+
+                                GGML_F32_VEC_STORE(s + i + j*ggml_f32_epr + ii*nc, t0);
+                            }
+                        }
+
+                        sumf = GGML_F32xt_REDUCE_ONE(sum);
+    #else
+                        const int np = (nc & ~(GGML_F32_STEP - 1));
+
+                        GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
+
+                        GGML_F32_VEC adA = GGML_F32_VEC_SET1(dA);
+                        GGML_F32_VEC axdt = GGML_F32_VEC_SET1(x_dt);
+
+                        GGML_F32_VEC ax[GGML_F32_ARR];
+                        GGML_F32_VEC ay[GGML_F32_ARR];
+                        GGML_F32_VEC az[GGML_F32_ARR];
+
+                        for (int i = 0; i < np; i += GGML_F32_STEP) {
+                            for (int j = 0; j < GGML_F32_ARR; j++) {
+                                ax[j] = GGML_F32_VEC_LOAD(s0 + i + j*GGML_F32_EPR + ii*nc);
+                                ay[j] = GGML_F32_VEC_LOAD(B + i + j*GGML_F32_EPR + (h & (ng - 1))*nc);
+                                az[j] = GGML_F32_VEC_LOAD(C + i + j*GGML_F32_EPR + (h & (ng - 1))*nc);
+
+                                ax[j] = GGML_F32_VEC_MUL(ax[j], adA);
+                                ay[j] = GGML_F32_VEC_MUL(ay[j], axdt);
+
+                                ax[j] = GGML_F32_VEC_ADD(ax[j], ay[j]);
+
+                                sum[j] = GGML_F32_VEC_FMA(sum[j], ax[j], az[j]);
+
+                                GGML_F32_VEC_STORE(s + i + j*GGML_F32_EPR + ii*nc, ax[j]);
+                            }
+                        }
+
+                        // reduce sum0..sum3 to sum0
+                        GGML_F32_VEC_REDUCE(sumf, sum);
+    #endif
+#else
+                        const int np = 0;
+#endif
+                        // d_state
+                        for (int i0 = np; i0 < nc; ++i0) {
+                            const int i = i0 + ii*nc;
+                            const int ig = i0 + (h & (ng - 1))*nc;
+                            // state = prev_state * dA + dB * x
+                            const float state = (s0[i] * dA) + (B[ig] * x_dt);
+                            // y = rowwise_dotprod(state, C)
+                            sumf += state * C[ig];
+                            s[i] = state;
+                        }
+                        y[ii] = sumf;
+                    }
+                }
+            } else {
+                // Mamba-1 has an element-wise decay factor for the states
+
+                // n_head
+                for (int h = ih0; h < ih1; ++h) {
+                    // ref: https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/ops/triton/softplus.py#L16
+                    const float dt_soft_plus = dt[h] <= 20.0f ? log1pf(expf(dt[h])) : dt[h];
+
+                    // dim
+                    for (int i1 = 0; i1 < nr; ++i1) {
+                        const int ii = i1 + h*nr;
+                        const float x_dt = x[ii] * dt_soft_plus;
+#if defined(__ARM_FEATURE_SVE)
                         svfloat32_t vx_dt = GGML_F32_VEC_SET1(x_dt);
                         svfloat32_t vdt_soft_plus = GGML_F32_VEC_SET1(dt_soft_plus);
                         svfloat32_t r1_vector = GGML_F32_VEC_ZERO;
 
+                        // d_state
+                        // TODO: what happens when (d_state % svcntw()) != 0?
                         for (int64_t k = 0; k < nc; k += svcntw()) {
-                        svfloat32_t vA = GGML_F32_VEC_LOAD(&A[i1*nc + k]);
+                            svfloat32_t vA = GGML_F32_VEC_LOAD(&A[h*nc + k]);
-                        svfloat32_t vB = GGML_F32_VEC_LOAD(&B[k]);
+                            svfloat32_t vB = GGML_F32_VEC_LOAD(&B[k + (h & (ng - 1))*nc]);
-                        svfloat32_t vC = GGML_F32_VEC_LOAD(&C[k]);
+                            svfloat32_t vC = GGML_F32_VEC_LOAD(&C[k + (h & (ng - 1))*nc]);
-                        svfloat32_t vs0 = GGML_F32_VEC_LOAD(&s0[i1*nc + k]);
+                            svfloat32_t vs0 = GGML_F32_VEC_LOAD(&s0[ii*nc + k]);
 
                             svfloat32_t t1 = GGML_F32_VEC_MUL(vdt_soft_plus, vA);
                             t1 = exp_ps_sve(svptrue_b32(), t1);
                             svfloat32_t t2 = GGML_F32_VEC_MUL(vx_dt, vB);
 
-                        vs0 = GGML_F32_VEC_FMA(vs0, t1, t2);
+                            vs0 = GGML_F32_VEC_FMA(t2, vs0, t1);
                             r1_vector = GGML_F32_VEC_ADD(GGML_F32_VEC_MUL(vs0, vC), r1_vector);
 
-                        GGML_F32_VEC_STORE(&s[i1*nc + k], vs0);
+                            GGML_F32_VEC_STORE(&s[ii*nc + k], vs0);
                         }
-                    y[i1] = GGML_F32xt_REDUCE_ONE(r1_vector);
+                        y[ii] = GGML_F32xt_REDUCE_ONE(r1_vector);
-                }
+#else
-            }
-        }
-    #else
-        for (int i3 = 0; i3 < n_s; ++i3) {
-            for (int i2 = 0; i2 < n_t; ++i2) {
-                const float * s0 = (const float *) ((const char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2])); // {d_state, d_inner, n_s}
-                const float * x  = (const float *) ((const char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
-                const float * dt = (const float *) ((const char *) src2->data + ir0*(src2->nb[0]) + i2*(src2->nb[1]) + i3*(src2->nb[2])); // {d_inner, n_t, n_s}
-                const float * A  = (const float *) ((const char *) src3->data + ir0*(src3->nb[1])); // {d_state, d_inner}
-                const float * B  = (const float *) ((const char *) src4->data +  i2*(src4->nb[1]) + i3*(src4->nb[2])); // {d_state, n_t, n_s}
-                const float * C  = (const float *) ((const char *) src5->data +  i2*(src5->nb[1]) + i3*(src5->nb[2])); // {d_state, n_t, n_s}
-                    float * y  = (      float *) ((      char *) dst->data  + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
-                    float * s  = (      float *) ((      char *) dst->data  + ir0*(src0->nb[1]) + i3*(src0->nb[2]) +     src1->nb[3]);  // {d_state, d_inner, n_s}
-
-                // use the output as the source for the next token-wise iterations
-                if (i2 > 0) { s0 = s; }
-
-                // d_inner
-                for (int i1 = 0; i1 < ir; ++i1) {
-                    // ref: https://github.com/state-spaces/mamba/blob/34076d664838588a3c97727b263478ab9f621a07/mamba_ssm/ops/triton/selective_state_update.py#L78
-                    float dt_soft_plus = dt[i1] <= 20.0f ? log1pf(expf(dt[i1])) : dt[i1];
-                    float x_dt = x[i1] * dt_soft_plus;
                         float sumf = 0.0f;
+                        // NOTE: can't really use GGML_SIMD here because d_state is usually 16
+                        //       and also because expf is used within the loop.
                         // d_state
                         for (int i0 = 0; i0 < nc; ++i0) {
-                        int i = i0 + i1*nc;
+                            const int i = i0 + ii*nc;
+                            const int ig = i0 + (h & (ng - 1))*nc;
                             // state = prev_state * dA + dB * x
-                        float state = (s0[i] * expf(dt_soft_plus * A[i])) + (B[i0] * x_dt);
+                            const float state = (s0[i] * expf(dt_soft_plus * A[i0 + h*nc])) + (B[ig] * x_dt);
                             // y = rowwise_dotprod(state, C)
-                        sumf += state * C[i0];
+                            sumf += state * C[ig];
                             s[i] = state;
                         }
-                    y[i1] = sumf;
+                        y[ii] = sumf;
+#endif
                     }
                 }
             }
-    #endif
+            // use the output as the source when it's not the first token-wise iteration
+            s0 = s;
+        }
+    }
 }
 
 void ggml_compute_forward_ssm_scan(
@@ -8688,6 +9065,14 @@ void ggml_compute_forward_glu(
             {
                 ggml_compute_forward_swiglu(params, dst);
             } break;
+        case GGML_GLU_OP_GEGLU_ERF:
+            {
+                ggml_compute_forward_geglu_erf(params, dst);
+            } break;
+        case GGML_GLU_OP_GEGLU_QUICK:
+            {
+                ggml_compute_forward_geglu_quick(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");

ggml/src/ggml-cpu/simd-mappings.h

@@ -189,7 +189,7 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 #define GGML_F32xt_LOAD(...)              GGML_F32xt_LOAD_IMPL(DEFAULT_PG, __VA_ARGS__)
 #define GGML_F32xt_STORE_IMPL(pg,a,b)     svst1_f32(pg, a, b)
 #define GGML_F32xt_STORE(...)             GGML_F32xt_STORE_IMPL(DEFAULT_PG, __VA_ARGS__)
-#define GGML_F32xt_FMA_IMPL(pg, a, b, c)  svmad_f32_m(pg, a, b, c)
+#define GGML_F32xt_FMA_IMPL(pg, a, b, c)  svmad_f32_m(pg, b, c, a)
 #define GGML_F32xt_FMA(...)               GGML_F32xt_FMA_IMPL(DEFAULT_PG, __VA_ARGS__)
 #define GGML_F32xt_ADD_IMPL(pg, a, b)     svadd_f32_m(pg, a, b)
 #define GGML_F32xt_ADD(...)               GGML_F32xt_ADD_IMPL(DEFAULT_PG, __VA_ARGS__)

ggml/src/ggml-cpu/vec.cpp

@@ -37,35 +37,35 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
         for (int i = 0; i < np; i += ggml_f32_step) {
             ax1 = GGML_F32_VEC_LOAD(x + i);
             ay1 = GGML_F32_VEC_LOAD(y + i);
-            sum1 = GGML_F32_VEC_FMA(ax1, ay1, sum1);
+            sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);
 
             ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
             ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
-            sum2 = GGML_F32_VEC_FMA(ax2, ay2, sum2);
+            sum2 = GGML_F32_VEC_FMA(sum2, ax2, ay2);
 
             ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr);
             ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr);
-            sum3 = GGML_F32_VEC_FMA(ax3, ay3, sum3);
+            sum3 = GGML_F32_VEC_FMA(sum3, ax3, ay3);
 
             ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr);
             ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr);
-            sum4 = GGML_F32_VEC_FMA(ax4, ay4, sum4);
+            sum4 = GGML_F32_VEC_FMA(sum4, ax4, ay4);
 
             ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr);
             ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr);
-            sum5 = GGML_F32_VEC_FMA(ax5, ay5, sum5);
+            sum5 = GGML_F32_VEC_FMA(sum5, ax5, ay5);
 
             ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr);
             ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr);
-            sum6 = GGML_F32_VEC_FMA(ax6, ay6, sum6);
+            sum6 = GGML_F32_VEC_FMA(sum6, ax6, ay6);
 
             ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr);
             ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr);
-            sum7 = GGML_F32_VEC_FMA(ax7, ay7, sum7);
+            sum7 = GGML_F32_VEC_FMA(sum7, ax7, ay7);
 
             ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr);
             ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr);
-            sum8 = GGML_F32_VEC_FMA(ax8, ay8, sum8);
+            sum8 = GGML_F32_VEC_FMA(sum8, ax8, ay8);
         }
         // leftovers
         // Since 8 unrolls are done in above loop, leftovers lie in range [0, ggml_f32_step] which is handled in below loop
@@ -73,7 +73,7 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
         for (int i = np; i < np2; i += ggml_f32_epr) {
             ax1 = GGML_F32_VEC_LOAD(x + i);
             ay1 = GGML_F32_VEC_LOAD(y + i);
-            sum1 = GGML_F32_VEC_FMA(ax1, ay1, sum1);
+            sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);
         }
         // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
         if (np2 < n) {

ggml/src/ggml-cpu/vec.h

@@ -163,49 +163,49 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
 
             ax1 = GGML_F32_VEC_LOAD(x + i);
             ay1 = GGML_F32_VEC_LOAD(y + i);
-            ay1 = GGML_F32_VEC_FMA(ax1, vx, ay1);
+            ay1 = GGML_F32_VEC_FMA(ay1, ax1, vx);
 
             GGML_F32_VEC_STORE(y + i, ay1);
 
             ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
             ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
-            ay2 = GGML_F32_VEC_FMA(ax2, vx, ay2);
+            ay2 = GGML_F32_VEC_FMA(ay2, ax2, vx);
 
             GGML_F32_VEC_STORE(y + i + 1*ggml_f32_epr, ay2);
 
             ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr);
             ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr);
-            ay3 = GGML_F32_VEC_FMA(ax3, vx, ay3);
+            ay3 = GGML_F32_VEC_FMA(ay3, ax3, vx);
 
             GGML_F32_VEC_STORE(y + i + 2*ggml_f32_epr, ay3);
 
             ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr);
             ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr);
-            ay4 = GGML_F32_VEC_FMA(ax4, vx, ay4);
+            ay4 = GGML_F32_VEC_FMA(ay4, ax4, vx);
 
             GGML_F32_VEC_STORE(y + i + 3*ggml_f32_epr, ay4);
 
             ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr);
             ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr);
-            ay5 = GGML_F32_VEC_FMA(ax5, vx, ay5);
+            ay5 = GGML_F32_VEC_FMA(ay5, ax5, vx);
 
             GGML_F32_VEC_STORE(y + i + 4*ggml_f32_epr, ay5);
 
             ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr);
             ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr);
-            ay6 = GGML_F32_VEC_FMA(ax6, vx, ay6);
+            ay6 = GGML_F32_VEC_FMA(ay6, ax6, vx);
 
             GGML_F32_VEC_STORE(y + i + 5*ggml_f32_epr, ay6);
 
             ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr);
             ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr);
-            ay7 = GGML_F32_VEC_FMA(ax7, vx, ay7);
+            ay7 = GGML_F32_VEC_FMA(ay7, ax7, vx);
 
             GGML_F32_VEC_STORE(y + i + 6*ggml_f32_epr, ay7);
 
             ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr);
             ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr);
-            ay8 = GGML_F32_VEC_FMA(ax8, vx, ay8);
+            ay8 = GGML_F32_VEC_FMA(ay8, ax8, vx);
 
             GGML_F32_VEC_STORE(y + i + 7*ggml_f32_epr, ay8);
         }
@@ -215,7 +215,7 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
         for (int i = np; i < np2; i += ggml_f32_epr) {
             ax1 = GGML_F32_VEC_LOAD(x + i);
             ay1 = GGML_F32_VEC_LOAD(y + i);
-            ay1 = GGML_F32_VEC_FMA(ax1, vx, ay1);
+            ay1 = GGML_F32_VEC_FMA(ay1, ax1, vx);
 
             GGML_F32_VEC_STORE(y + i, ay1);
         }
@@ -959,6 +959,46 @@ inline static void ggml_vec_swiglu_f16(const int n, ggml_fp16_t * y, const ggml_
     }
 }
 
+inline static void ggml_vec_geglu_erf_f32(const int n, float * y, const float * x, const float * g) {
+    for (int i = 0; i < n; ++i) {
+        float xi = x[i];
+        y[i] = 0.5f * xi * (1.0f + erff(xi*SQRT_2_INV)) * g[i];
+    }
+}
+
+inline static void ggml_vec_geglu_erf_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+    for (int i = 0; i < n; ++i) {
+        float xi = GGML_CPU_FP16_TO_FP32(x[i]);
+        float gi = GGML_CPU_FP16_TO_FP32(g[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(0.5f * xi * (1.0f + erff(xi*SQRT_2_INV)) * gi);
+    }
+}
+
+#ifdef GGML_GELU_QUICK_FP16
+inline static void ggml_vec_geglu_quick_f32(const int n, float * y, const float * x, const float * g) {
+    uint16_t t;
+    for (int i = 0; i < n; ++i) {
+        ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
+        memcpy(&t, &fp16, sizeof(uint16_t));
+        y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]) * g[i];
+    }
+}
+#else
+inline static void ggml_vec_geglu_quick_f32(const int n, float * y, const float * x, const float * g) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = ggml_gelu_quick_f32(x[i]) * g[i];
+    }
+}
+#endif
+
+inline static void ggml_vec_geglu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+    const uint16_t * i16 = (const uint16_t *) x;
+    for (int i = 0; i < n; ++i) {
+        float v = GGML_CPU_FP16_TO_FP32(g[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(ggml_table_gelu_quick_f16[i16[i]]) * v);
+    }
+}
+
 inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) {
 #ifndef GGML_USE_ACCELERATE
     ggml_float sum = 0.0;

ggml/src/ggml-cuda/common.cuh

@@ -179,6 +179,20 @@ static const char * cu_get_error_str(CUresult err) {
 #define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str)
 #endif
 
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#define CUDA_SET_SHARED_MEMORY_LIMIT(kernel, nbytes) \
+    do { \
+        static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false}; \
+        const int id = ggml_cuda_get_device(); \
+        if (!shared_memory_limit_raised[id]) { \
+            CUDA_CHECK(cudaFuncSetAttribute(kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes)); \
+            shared_memory_limit_raised[id] = true; \
+        } \
+    } while (0)
+#else
+#define CUDA_SET_SHARED_MEMORY_LIMIT(kernel, nbytes) do {} while (0)
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+
 #if CUDART_VERSION >= 11010 || defined(GGML_USE_MUSA)
 #define GGML_CUDA_ASSUME(x) __builtin_assume(x)
 #else

ggml/src/ggml-cuda/cross-entropy-loss.cu

@@ -123,13 +123,7 @@ void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor *
     ggml_cuda_pool_alloc<float> dst_tmp(pool, blocks_num.x);
 
     if (nbytes_shared <= smpbo) {
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+        CUDA_SET_SHARED_MEMORY_LIMIT((cross_entropy_loss_f32<true>), smpbo);
-        static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
-        if (!shared_memory_limit_raised[id]) {
-            CUDA_CHECK(cudaFuncSetAttribute(cross_entropy_loss_f32<true>, cudaFuncAttributeMaxDynamicSharedMemorySize, smpbo));
-            shared_memory_limit_raised[id] = true;
-        }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
         cross_entropy_loss_f32<true><<<blocks_num, blocks_dim, nbytes_shared, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
     } else {
         cross_entropy_loss_f32<false><<<blocks_num, blocks_dim, 0, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
@@ -175,13 +169,7 @@ void ggml_cuda_cross_entropy_loss_back(ggml_backend_cuda_context & ctx, ggml_ten
     const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
 
     if (nbytes_shared <= smpbo) {
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+        CUDA_SET_SHARED_MEMORY_LIMIT((cross_entropy_loss_back_f32<true>), smpbo);
-        static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
-        if (!shared_memory_limit_raised[id]) {
-            CUDA_CHECK(cudaFuncSetAttribute(cross_entropy_loss_back_f32<true>, cudaFuncAttributeMaxDynamicSharedMemorySize, smpbo));
-            shared_memory_limit_raised[id] = true;
-        }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
         cross_entropy_loss_back_f32<true><<<blocks_num, blocks_dim, nbytes_shared, stream>>>(grad_d, src0f_d, src1f_d, dst_d, ne00);
     } else {
         cross_entropy_loss_back_f32<false><<<blocks_num, blocks_dim, 0, stream>>>(grad_d, src0f_d, src1f_d, dst_d, ne00);

ggml/src/ggml-cuda/fattn-common.cuh

@@ -32,7 +32,9 @@ typedef void (* fattn_kernel_t)(
         const int ne12,
         const int ne13,
         const int ne31,
+        const int ne32,
         const int nb31,
+        const int nb32,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -851,7 +853,8 @@ void launch_fattn(
         scale, max_bias, m0, m1, n_head_log2, logit_softcap,
         Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
         K->ne[0], K->ne[1], K->ne[2], K->ne[3],
-        mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
+        mask ? mask->ne[1] : 0, mask ? mask->ne[2] : 0,
+        mask ? mask->nb[1] : 0, mask ? mask->nb[2] : 0,
         Q->nb[1], Q->nb[2], Q->nb[3],
         nb11, nb12, nb13,
         nb21, nb22, nb23,

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -1223,7 +1223,9 @@ static __global__ void flash_attn_ext_f16(
         const int ne12,
         const int ne13,
         const int ne31,
+        const int ne32,
         const int nb31,
+        const int nb32,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -1288,7 +1290,8 @@ static __global__ void flash_attn_ext_f16(
 
         const float2 * Q_f2    = (const float2 *) (Q + nb02* channel*ncols2);
         const half2  * K_h2    = (const half2  *) (K + nb12*(channel*ncols2 / gqa_ratio));
-        const half2  * mask_h2 = ncols2 > 1 || mask ? (const half2  *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr;
+        const half2  * mask_h2 = ncols2 == 1 && !mask ? nullptr :
+            (const half2  *) (mask + nb32*(channel % ne32) + nb31*jt*ncols1);
         float2       * dstk    = ((float2 *) dst) + channel*(ncols2 * DV/2);
 
         const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio));
@@ -1327,7 +1330,8 @@ static __global__ void flash_attn_ext_f16(
 
     const float2 * Q_f2    = (const float2 *) (Q + nb02* channel*ncols2);
     const half2  * K_h2    = (const half2  *) (K + nb12*(channel*ncols2 / gqa_ratio));
-    const half2  * mask_h2 = ncols2 > 1 || mask ? (const half2  *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr;
+    const half2  * mask_h2 = ncols2 == 1 && !mask ? nullptr :
+        (const half2  *) (mask + nb32*(channel % ne32) + nb31*jt*ncols1);
     float2       * dstk    = ((float2 *) dst) + channel*(ncols2 * DV/2);
 
     const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio));
@@ -1348,8 +1352,8 @@ static __global__ void flash_attn_ext_f16(
     GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); GGML_UNUSED(ne00);
     GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03); GGML_UNUSED(ne10);
-    GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
+    GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
     GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13); GGML_UNUSED(nb21);
     GGML_UNUSED(nb22); GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
     GGML_UNUSED(ne2); GGML_UNUSED(ne3);

ggml/src/ggml-cuda/fattn-tile-f16.cu

@@ -6,7 +6,7 @@
 
 template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
-__launch_bounds__(nwarps*WARP_SIZE, 1)
+__launch_bounds__(nwarps*WARP_SIZE, 2)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_tile_ext_f16(
         const char * __restrict__ Q,
@@ -30,7 +30,9 @@ static __global__ void flash_attn_tile_ext_f16(
         const int ne12,
         const int ne13,
         const int ne31,
+        const int ne32,
         const int nb31,
+        const int nb32,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -64,7 +66,7 @@ static __global__ void flash_attn_tile_ext_f16(
     const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.z              + nb01*ic0);
     const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.z / gqa_ratio));
     const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
-    const half   * maskh = (const half   *)  mask + ne11*ic0;
+    const half   * maskh = (const half   *) (mask + nb32*(blockIdx.z % ne32)      + nb31*ic0);
 
     const int stride_KV2 = nb11 / sizeof(half2);
 
@@ -288,8 +290,8 @@ static __global__ void flash_attn_tile_ext_f16(
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
     GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
     GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
+    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
     GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
     GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
     GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);

ggml/src/ggml-cuda/fattn-tile-f32.cu

@@ -6,7 +6,7 @@
 
 template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
-__launch_bounds__(nwarps*WARP_SIZE, 1)
+__launch_bounds__(nwarps*WARP_SIZE, 2)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_tile_ext_f32(
         const char * __restrict__ Q,
@@ -30,7 +30,9 @@ static __global__ void flash_attn_tile_ext_f32(
         const int ne12,
         const int ne13,
         const int ne31,
+        const int ne32,
         const int nb31,
+        const int nb32,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -58,8 +60,8 @@ static __global__ void flash_attn_tile_ext_f32(
         GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
         GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
         GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-        GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
+        GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
-        GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+        GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
         GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
         GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
         GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
@@ -76,7 +78,7 @@ static __global__ void flash_attn_tile_ext_f32(
     const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.z              + nb01*ic0);
     const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.z / gqa_ratio));
     const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
-    const half   * maskh = (const half   *)  mask + ne11*ic0;
+    const half   * maskh = (const half   *) (mask + nb32*(blockIdx.z % ne32)      + nb31*ic0);
 
     const int stride_KV2 = nb11 / sizeof(half2);
 

ggml/src/ggml-cuda/fattn-vec-f16.cuh

@@ -27,7 +27,9 @@ static __global__ void flash_attn_vec_ext_f16(
         const int ne12,
         const int ne13,
         const int ne31,
+        const int ne32,
         const int nb31,
+        const int nb32,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -68,7 +70,7 @@ static __global__ void flash_attn_vec_ext_f16(
     K += nb12*(blockIdx.z / gqa_ratio);
     V += nb22*(blockIdx.z / gqa_ratio);
 
-    const half * maskh = (const half   *)  mask + ne11*ic0;
+    const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
 
     const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
     const half  slopeh = __float2half(slopef);
@@ -342,8 +344,8 @@ static __global__ void flash_attn_vec_ext_f16(
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
     GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
     GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
+    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
     GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
     GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
     GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);

ggml/src/ggml-cuda/fattn-vec-f32.cuh

@@ -27,7 +27,9 @@ static __global__ void flash_attn_vec_ext_f32(
         const int ne12,
         const int ne13,
         const int ne31,
+        const int ne32,
         const int nb31,
+        const int nb32,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -51,8 +53,8 @@ static __global__ void flash_attn_vec_ext_f32(
         GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
         GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
         GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-        GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
+        GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
-        GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+        GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
         GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
         GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
         GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
@@ -79,7 +81,8 @@ static __global__ void flash_attn_vec_ext_f32(
     Q += nb02* blockIdx.z              + nb01*ic0;
     K += nb12*(blockIdx.z / gqa_ratio);
     V += nb22*(blockIdx.z / gqa_ratio); // K and V have same shape
-    const half * maskh = (const half   *)  mask + ne11*ic0;
+
+    const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
 
     const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
 

ggml/src/ggml-cuda/fattn-wmma-f16.cu

@@ -46,7 +46,9 @@ static __global__ void flash_attn_ext_f16(
         const int ne12,
         const int ne13,
         const int ne31,
+        const int ne32,
         const int nb31,
+        const int nb32,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -97,8 +99,8 @@ static __global__ void flash_attn_ext_f16(
     const float * Q_f   = (const float *) (Q    + nb02* blockIdx.z              + nb01*ic0);
     const half  * K_h   = (const half  *) (K    + nb12*(blockIdx.z / gqa_ratio));
     const half  * V_h   = (const half  *) (V    + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
-    const half  * maskh = (const half  *)  mask + (nb31/sizeof(half))* ic0;
+    const half  * maskh = (const half  *) (mask + nb32*(blockIdx.z % ne32)      + nb31*ic0);
-    const half2 * mask2 = (const half2 *)  mask + (nb31/sizeof(half))*(ic0/2);
+    const half2 * mask2 = (const half2 *)  maskh;
 
     const int stride_Q  = nb01 / sizeof(float);
     const int stride_KV = nb11 / sizeof(half);
@@ -440,7 +442,7 @@ static __global__ void flash_attn_ext_f16(
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
     GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
     GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-    GGML_UNUSED(ne31); GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
     GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
     GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
     GGML_UNUSED(ne0); GGML_UNUSED(ne1); GGML_UNUSED(ne2); GGML_UNUSED(ne3);

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -2319,6 +2319,12 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                 case GGML_GLU_OP_SWIGLU:
                     ggml_cuda_op_swiglu(ctx, dst);
                     break;
+                case GGML_GLU_OP_GEGLU_ERF:
+                    ggml_cuda_op_geglu_erf(ctx, dst);
+                    break;
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    ggml_cuda_op_geglu_quick(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -3121,6 +3127,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_GLU_OP_REGLU:
                 case GGML_GLU_OP_GEGLU:
                 case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
                     return ggml_is_contiguous_1(op->src[0]);
                 default:
                     return false;
@@ -3326,12 +3334,26 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
         case GGML_OP_LOG:
-        case GGML_OP_SSM_SCAN:
-        case GGML_OP_SSM_CONV:
             return true;
+        case GGML_OP_SSM_SCAN: {
+            if (op->src[3]->ne[0] == 1) {
+                // Mamba2
+                // (kernel only supports d_state == 128 && d_head % 16 == 0)
+                return op->src[0]->ne[0] == 128 && op->src[0]->ne[1] % 16 == 0;
+            } else {
+                // Mamba
+                // (kernel only supports d_state == 16, d_head == 1, n_head % 128 == 0, n_group == 1)
+                return op->src[0]->ne[0] == 16 && op->src[0]->ne[1] == 1 && op->src[0]->ne[2] % 128 == 0 && op->src[4]->ne[1] == 1;
+            }
+        }
+        case GGML_OP_SSM_CONV: {
+            // assumes d_inner % threads == 0
+            return op->src[0]->ne[1] % 128 == 0;
+        }
         case GGML_OP_CONT:
             return op->src[0]->type != GGML_TYPE_BF16;
         case GGML_OP_DIAG_MASK_INF:
+            return true;
         case GGML_OP_SOFT_MAX:
             return true;
         case GGML_OP_SOFT_MAX_BACK: {
@@ -3380,6 +3402,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             if (op->src[0]->ne[0] == 192) {
                 return false;
             }
+            // TODO: support broadcast
+            // note: this was initially implemented in https://github.com/ggml-org/llama.cpp/pull/14500, but
+            //       the interface of ggml_flash_attn_ext() changed in https://github.com/ggml-org/llama.cpp/pull/14505
             if (op->src[0]->ne[3] != 1) {
                 return false;
             }

ggml/src/ggml-cuda/mmq.cuh

@@ -3017,14 +3017,8 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 
     const int nbytes_shared = mmq_get_nbytes_shared<type>(mmq_x, mmq_y, cc);
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+    CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q<type, mmq_x, MMQ_NWARPS, false>), nbytes_shared);
-    static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
+    CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q<type, mmq_x, MMQ_NWARPS, true>),  nbytes_shared);
-    if (!shared_memory_limit_raised[id]) {
-        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared));
-        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>,  cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared));
-        shared_memory_limit_raised[id] = true;
-    }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
 
     const int nty  = (args.nrows_x   + mmq_y - 1) / mmq_y;
     const int ntx  = (args.ncols_dst + mmq_x - 1) / mmq_x;

ggml/src/ggml-cuda/softmax.cu

@@ -2,6 +2,7 @@
 #include "ggml.h"
 #include "softmax.cuh"
 #include <cstdint>
+#include <utility>
 
 template <typename T>
 static __device__ __forceinline__ float t2f32(T val) {
@@ -13,6 +14,29 @@ __device__ float __forceinline__ t2f32<half>(half val) {
     return __half2float(val);
 }
 
+struct soft_max_params {
+
+    int64_t nheads;
+    uint32_t n_head_log2;
+    int64_t ncols;
+    int64_t nrows_x;
+    int64_t nrows_y;
+    int64_t ne00;
+    int64_t ne01;
+    int64_t ne02;
+    int64_t ne03;
+    int64_t nb11;
+    int64_t nb12;
+    int64_t nb13;
+
+    int64_t ne12;
+    int64_t ne13;
+    float scale;
+    float max_bias;
+    float m0;
+    float m1;
+};
+
 // When ncols_template == 0 the bounds for the loops in this function are not known and can't be unrolled.
 // As we want to keep pragma unroll for all other cases we supress the clang transformation warning here.
 #ifdef __clang__
@@ -21,16 +45,24 @@ __device__ float __forceinline__ t2f32<half>(half val) {
 #endif // __clang__
 template <bool use_shared, int ncols_template, int block_size_template, typename T>
 static __global__ void soft_max_f32(
-        const float * x, const T * mask, float * dst, const int ncols_par, const int nrows_y,
+        const float * x, const T * mask, float * dst, const soft_max_params p) {
-        const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) {
+    const int ncols = ncols_template == 0 ? p.ncols : ncols_template;
-    const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
 
     const int tid  = threadIdx.x;
-    const int rowx = blockIdx.x;
+
-    const int rowy = rowx % nrows_y; // broadcast the mask in the row dimension
+    const int64_t i03 = blockIdx.z;
+    const int64_t i02 = blockIdx.y;
+    const int64_t i01 = blockIdx.x;
+
+    //TODO: noncontigous inputs/outputs
+    const int rowx = blockIdx.x + blockIdx.y * gridDim.x + blockIdx.z * gridDim.x * gridDim.y;
+
+    const int64_t i11 = i01;
+    const int64_t i12 = i02 % p.ne12;
+    const int64_t i13 = i03 % p.ne13;
 
     x    += int64_t(rowx)*ncols;
-    mask += int64_t(rowy)*ncols * (mask != nullptr);
+    mask += (i11*p.nb11 + i12*p.nb12 + i13*p.nb13) / sizeof(T) * (mask != nullptr);
     dst  += int64_t(rowx)*ncols;
 
     const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
@@ -38,7 +70,7 @@ static __global__ void soft_max_f32(
     const int warp_id = threadIdx.x / WARP_SIZE;
     const int lane_id = threadIdx.x % WARP_SIZE;
 
-    const float slope = get_alibi_slope(max_bias, rowx/nrows_y, n_head_log2, m0, m1);
+    const float slope = get_alibi_slope(p.max_bias, i02, p.n_head_log2, p.m0, p.m1);
 
     extern __shared__ float data_soft_max_f32[];
     float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication
@@ -55,7 +87,7 @@ static __global__ void soft_max_f32(
             break;
         }
 
-        const float val = x[col]*scale + (mask ? slope*t2f32(mask[col]) : 0.0f);
+        const float val = x[col]*p.scale + (mask ? slope*t2f32(mask[col]) : 0.0f);
 
         vals[col] = val;
         max_val = max(max_val, val);
@@ -150,64 +182,58 @@ static __global__ void soft_max_back_f32(
     }
 }
 
+template<int... Ns, typename T>
+static void launch_soft_max_kernels(const float * x, const T * mask, float * dst,
+                             const soft_max_params & p, cudaStream_t stream, dim3 block_dims, dim3 block_nums, size_t nbytes_shared)
+{
+    const int id       = ggml_cuda_get_device();
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
+
+    auto launch_kernel = [=](auto I) -> bool {
+        constexpr int ncols = decltype(I)::value;
+        constexpr int block = (ncols > 1024 ? 1024 : ncols);
+
+        if (p.ncols == ncols) {
+            CUDA_SET_SHARED_MEMORY_LIMIT((soft_max_f32<true, ncols, block, T>), smpbo);
+            soft_max_f32<true, ncols, block><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, mask, dst, p);
+            return true;
+        }
+        return false;
+    };
+
+    // unary fold over launch_kernel
+    if ((launch_kernel(std::integral_constant<int, Ns>{}) || ...)) {
+        return;
+    }
+
+    //default case
+    CUDA_SET_SHARED_MEMORY_LIMIT((soft_max_f32<true, 0, 0, T>), smpbo);
+    soft_max_f32<true, 0, 0><<<block_nums, block_dims, nbytes_shared, stream>>>(x, mask, dst, p);
+}
+
+
 template<typename T>
-static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) {
+static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, const soft_max_params & params, cudaStream_t stream) {
     int nth = WARP_SIZE;
+    const int64_t ncols_x = params.ncols;
+
     while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
     const dim3 block_dims(nth,     1, 1);
-    const dim3 block_nums(nrows_x, 1, 1);
+    const dim3 block_nums(params.ne01, params.ne02, params.ne03);
     const size_t nbytes_shared = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float);
     static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
 
-    const uint32_t n_head      = nrows_x/nrows_y;
-    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
 
-    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const int id       = ggml_cuda_get_device();
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
 
-    // FIXME: this limit could be raised by ~2-4x on Ampere or newer
+
-    if (nbytes_shared < ggml_cuda_info().devices[ggml_cuda_get_device()].smpb) {
+    if (nbytes_shared <= smpbo) {
-        switch (ncols_x) {
+        launch_soft_max_kernels<32, 64, 128, 256, 512, 1024, 2048, 4096>(x, mask, dst, params, stream, block_dims, block_nums, nbytes_shared);
-            case 32:
-                soft_max_f32<true,   32,   32><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 64:
-                soft_max_f32<true,   64,   64><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 128:
-                soft_max_f32<true,  128,  128><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 256:
-                soft_max_f32<true,  256,  256><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 512:
-                soft_max_f32<true,  512,  512><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 1024:
-                soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 2048:
-                soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 4096:
-                soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            default:
-                soft_max_f32<true,    0,    0><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-        }
     } else {
         const size_t nbytes_shared_low = WARP_SIZE*sizeof(float);
-        soft_max_f32<false, 0, 0><<<block_nums, block_dims, nbytes_shared_low, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+        soft_max_f32<false, 0, 0><<<block_nums, block_dims, nbytes_shared_low, stream>>>(x, mask, dst, params);
     }
 }
 
@@ -235,10 +261,11 @@ void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional
 
-    const int64_t ne00    = src0->ne[0];
     const int64_t nrows_x = ggml_nrows(src0);
     const int64_t nrows_y = src0->ne[1];
 
+    const int64_t ne00 = src0->ne[0];
+
     float scale    = 1.0f;
     float max_bias = 0.0f;
 
@@ -247,10 +274,44 @@ void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
 
+    const int64_t nb11 = src1 ? src1->nb[1] : 1;
+    const int64_t nb12 = src1 ? src1->nb[2] : 1;
+    const int64_t nb13 = src1 ? src1->nb[3] : 1;
+
+    const int64_t ne12 = src1 ? src1->ne[2] : 1;
+    const int64_t ne13 = src1 ? src1->ne[3] : 1;
+
+    const uint32_t n_head      = src0->ne[2];
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+
+    soft_max_params params = {};
+    params.nheads = src0->ne[2];
+    params.n_head_log2 = n_head_log2;
+    params.ncols = ne00;
+    params.nrows_x = nrows_x;
+    params.nrows_y = nrows_y;
+    params.ne00 = src0->ne[0];
+    params.ne01 = src0->ne[1];
+    params.ne02 = src0->ne[2];
+    params.ne03 = src0->ne[3];
+    params.nb11 = nb11;
+    params.nb12 = nb12;
+    params.nb13 = nb13;
+    params.ne12 = ne12;
+    params.ne13 = ne13;
+    params.scale = scale;
+    params.max_bias = max_bias;
+    params.m0 = m0;
+    params.m1 = m1;
+
     if (use_f16) {
-        soft_max_f32_cuda(src0_d, (const half  *) src1_d, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
+        soft_max_f32_cuda(src0_d, (const half  *) src1_d, dst_d, params, stream);
     } else {
-        soft_max_f32_cuda(src0_d, (const float *) src1_d, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
+        soft_max_f32_cuda(src0_d, (const float *) src1_d, dst_d, params, stream);
     }
 }
 

ggml/src/ggml-cuda/ssm-scan.cu

@@ -4,16 +4,15 @@ template <size_t splitD, size_t N>
 __global__ void __launch_bounds__(splitD, 2)
     ssm_scan_f32(const float * __restrict__ src0, const float * __restrict__ src1, const float * __restrict__ src2,
                  const float * __restrict__ src3, const float * __restrict__ src4, const float * __restrict__ src5,
-                 const int src0_nb1, const int src0_nb2, const int src1_nb0, const int src1_nb1, const int src1_nb2,
+                 const int32_t * __restrict__ src6, float * __restrict__ dst,
-                 const int src1_nb3, const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1,
+                 const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3,
-                 const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2,
+                 const int src2_nb1, const int src2_nb2, const int src3_nb1,
-                 float * __restrict__ dst, const int64_t L) {
+                 const int src4_nb2, const int src4_nb3, const int src5_nb2, const int src5_nb3,
-    GGML_UNUSED(src1_nb0);
+                 const int64_t s_off, const int64_t d_inner, const int64_t L) {
-    GGML_UNUSED(src2_nb0);
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
-    const int bidx = blockIdx.x;  // split along B
+    const int bidx = blockIdx.x;  // split along B (sequences)
-    const int bidy = blockIdx.y;  // split along D
+    const int bidy = blockIdx.y;  // split along D (d_inner)
     const int tid  = threadIdx.x;
     const int wid  = tid / 32;
     const int wtid = tid % 32;
@@ -24,23 +23,23 @@ __global__ void __launch_bounds__(splitD, 2)
     float *                 smem_A     = smem;
     float *                 smem_s0    = smem_A + splitD * stride_sA;
 
-    const float * s0_block = (const float *) ((const char *) src0 + bidx * src0_nb2 + bidy * splitD * src0_nb1);
+    const float * s0_block = (const float *) ((const char *) src0 + src6[bidx] * src0_nb3 + bidy * splitD * src0_nb2);
-    const float * x_block  = (const float *) ((const char *) src1 + (bidx * src1_nb2) + bidy * splitD * sizeof(float));
+    const float * x_block  = (const float *) ((const char *) src1 + (bidx * src1_nb3) + bidy * splitD * sizeof(float));
     const float * dt_block = (const float *) ((const char *) src2 + (bidx * src2_nb2) + bidy * splitD * sizeof(float));
     const float * A_block  = (const float *) ((const char *) src3 + bidy * splitD * src3_nb1);
-    const float * B_block  = (const float *) ((const char *) src4 + (bidx * src4_nb2));
+    const float * B_block  = (const float *) ((const char *) src4 + (bidx * src4_nb3));
-    const float * C_block  = (const float *) ((const char *) src5 + (bidx * src5_nb2));
+    const float * C_block  = (const float *) ((const char *) src5 + (bidx * src5_nb3));
-    float *       y_block  = (float *) ((char *) dst + (bidx * src1_nb2) + bidy * splitD * sizeof(float));
+    float *       y_block  = (float *) ((char *) dst + (bidx * d_inner * L * sizeof(float)) + bidy * splitD * sizeof(float));
-    float *       s_block  = (float *) ((char *) dst + src1_nb3 + bidx * src0_nb2 + bidy * splitD * src0_nb1);
+    float *       s_block  = (float *) ((char *) dst + s_off + bidx * src0_nb3 + bidy * splitD * src0_nb2);
 
-    const int stride_s0 = src0_nb1 / sizeof(float);
+    const int stride_s0 = src0_nb2 / sizeof(float);
-    const int stride_x  = src1_nb1 / sizeof(float);
+    const int stride_x  = src1_nb2 / sizeof(float);
     const int stride_dt = src2_nb1 / sizeof(float);
     const int stride_A  = src3_nb1 / sizeof(float);
-    const int stride_B  = src4_nb1 / sizeof(float);
+    const int stride_B  = src4_nb2 / sizeof(float);
-    const int stride_C  = src5_nb1 / sizeof(float);
+    const int stride_C  = src5_nb2 / sizeof(float);
     const int stride_s  = stride_s0;
-    const int stride_y  = stride_x;
+    const int stride_y  = d_inner;
 
     // can N not be 16? for example 32?
     if (N == 16) {
@@ -84,24 +83,156 @@ __global__ void __launch_bounds__(splitD, 2)
     }
 }
 
+// assumes as many threads as d_state
+template <int splitH, int d_state>
+__global__ void __launch_bounds__(d_state, 1)
+    ssm_scan_f32_group(
+        const float * __restrict__ src0, const float * __restrict__ src1, const float * __restrict__ src2,
+        const float * __restrict__ src3, const float * __restrict__ src4, const float * __restrict__ src5,
+        const int32_t * __restrict__ src6, float * __restrict__ dst,
+        const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3,
+        const int src2_nb1, const int src2_nb2, const int src3_nb1,
+        const int src4_nb2, const int src4_nb3, const int src5_nb2, const int src5_nb3,
+        const int64_t s_off, const int64_t n_head, const int64_t d_head, const int64_t n_group, const int64_t n_tok) {
+
+    const int head_idx = (blockIdx.x * splitH) / d_head;
+    const int head_off = ((blockIdx.x * splitH) % d_head) * sizeof(float);
+    const int seq_idx = blockIdx.y;
+
+    const int group_off = (head_idx & (n_group - 1)) * d_state * sizeof(float);
+
+    const float * s0_block = (const float *) ((const char *) src0 + src6[seq_idx] * src0_nb3 + head_idx * src0_nb2 + head_off * d_state);
+    const float * x_block  = (const float *) ((const char *) src1 + (seq_idx * src1_nb3) + blockIdx.x * splitH * sizeof(float));
+    const float * dt_block = (const float *) ((const char *) src2 + (seq_idx * src2_nb2) + head_idx * sizeof(float));
+    const float * A_block  = (const float *) ((const char *) src3 + head_idx * src3_nb1);
+    const float * B_block  = (const float *) ((const char *) src4 + (seq_idx * src4_nb3) + (group_off));
+    const float * C_block  = (const float *) ((const char *) src5 + (seq_idx * src5_nb3) + (group_off));
+    float *       y_block  = dst + (seq_idx * n_tok * n_head * d_head) + blockIdx.x * splitH;
+    float *       s_block  = (float *) ((char *) dst + s_off + seq_idx * src0_nb3 + head_idx * src0_nb2 + head_off * d_state);
+
+    // strides across n_seq_tokens
+    const int stride_x  = src1_nb2 / sizeof(float);
+    const int stride_dt = src2_nb1 / sizeof(float);
+    const int stride_B  = src4_nb2 / sizeof(float);
+    const int stride_C  = src5_nb2 / sizeof(float);
+    const int stride_y  = n_head * d_head;
+
+    float state[splitH];
+    // for the parallel accumulation
+    __shared__ float stateC[splitH * d_state];
+
+#pragma unroll
+    for (int j = 0; j < splitH; j++) {
+        state[j] = s0_block[j * d_state + threadIdx.x];
+    }
+
+    for (int64_t i = 0; i < n_tok; i++) {
+        // TODO: only calculate dA and dt_soft_plus once per head instead of every splitH head elements
+        // TODO: only calculate B and C once per head group
+        // NOTE: dt_soft_plus, dA and x_dt have the same value across threads here.
+        float dt_soft_plus = dt_block[i * stride_dt];
+        if (dt_soft_plus <= 20.0f) {
+            dt_soft_plus = log1pf(expf(dt_soft_plus));
+        }
+        const float dA = expf(dt_soft_plus * A_block[0]);
+        const float B = B_block[i * stride_B + threadIdx.x];
+        const float C = C_block[i * stride_C + threadIdx.x];
+
+        // across d_head
+#pragma unroll
+        for (int j = 0; j < splitH; j++) {
+            const float x_dt = x_block[i * stride_x + j] * dt_soft_plus;
+
+            state[j] = (state[j] * dA) + (B * x_dt);
+
+            stateC[j * d_state + threadIdx.x] = state[j] * C;
+        }
+
+        __syncthreads();
+
+        // parallel accumulation for stateC
+        // TODO: simplify
+        {
+            static_assert((d_state & -d_state) == d_state, "the state size has to be a power of 2");
+            static_assert((splitH & -splitH) == splitH, "splitH has to be a power of 2");
+
+            // reduce until w matches the warp size
+            // TODO: does this work even when the physical warp size is 64?
+#pragma unroll
+            for (int w = d_state; w > WARP_SIZE; w >>= 1) {
+                // (assuming there are d_state threads)
+#pragma unroll
+                for (int j = 0; j < ((w >> 1) * splitH + d_state - 1) / d_state; j++) {
+                    // TODO: check for bank conflicts
+                    const int k = (threadIdx.x % (w >> 1)) + (d_state * (threadIdx.x / (w >> 1))) + j * d_state * (d_state / (w >> 1));
+                    stateC[k] += stateC[k + (w >> 1)];
+
+                }
+                __syncthreads();
+            }
+
+            static_assert(splitH >= d_state / WARP_SIZE);
+
+#pragma unroll
+            for (int j = 0; j < splitH / (d_state / WARP_SIZE); j++) {
+                float y = stateC[(threadIdx.x % WARP_SIZE) + d_state * (threadIdx.x / WARP_SIZE) + j * d_state * (d_state / WARP_SIZE)];
+                y = warp_reduce_sum(y);
+
+                // store the above accumulations
+                if (threadIdx.x % WARP_SIZE == 0) {
+                    const int k = threadIdx.x / WARP_SIZE + j * (d_state / WARP_SIZE);
+                    y_block[i * stride_y + k] = y;
+                }
+            }
+        }
+    }
+
+    // write back the state
+#pragma unroll
+    for (int j = 0; j < splitH; j++) {
+        s_block[j * d_state + threadIdx.x] = state[j];
+    }
+}
+
 static void ssm_scan_f32_cuda(const float * src0, const float * src1, const float * src2, const float * src3,
-                              const float * src4, const float * src5, const int src0_nb1, const int src0_nb2,
+                              const float * src4, const float * src5, const int32_t * src6, float * dst,
-                              const int src1_nb0, const int src1_nb1, const int src1_nb2, const int src1_nb3,
+                              const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3, const int src2_nb1,
-                              const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1,
+                              const int src2_nb2, const int src3_nb1, const int src4_nb2, const int src4_nb3, const int src5_nb2,
-                              const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2,
+                              const int src5_nb3, const int64_t s_off, const int64_t d_state, const int64_t head_dim,
-                              float * dst, const int64_t N, const int64_t D, const int64_t L, const int64_t B,
+                              const int64_t n_head, const int64_t n_group, const int64_t n_tok, const int64_t n_seq,
                               cudaStream_t stream) {
     const int threads = 128;
-    // todo: consider D cannot be divided,does this situation exist?
+    // NOTE: if you change conditions here, be sure to update the corresponding supports_op condition!
-    GGML_ASSERT(D % threads == 0);
+    if (src3_nb1 == sizeof(float)) {
-    const dim3 blocks(B, (D + threads - 1) / threads, 1);
+        // Mamba-2
-    const int  smem_size = (threads * (N + 1) * 2) * sizeof(float);
+        if (d_state == 128) {
-    if (N == 16) {
+            GGML_ASSERT(d_state % threads == 0);
-        ssm_scan_f32<128, 16><<<blocks, threads, smem_size, stream>>>(
+            // NOTE: can be any power of two between 4 and 64
-            src0, src1, src2, src3, src4, src5, src0_nb1, src0_nb2, src1_nb0, src1_nb1, src1_nb2, src1_nb3, src2_nb0,
+            const int splitH = 16;
-            src2_nb1, src2_nb2, src3_nb1, src4_nb1, src4_nb2, src5_nb1, src5_nb2, dst, L);
+            GGML_ASSERT(head_dim % splitH == 0);
+            const dim3 blocks((n_head * head_dim + (splitH - 1)) / splitH, n_seq, 1);
+            ssm_scan_f32_group<16, 128><<<blocks, threads, 0, stream>>>(
+                    src0, src1, src2, src3, src4, src5, src6, dst,
+                    src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2, src3_nb1,
+                    src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, head_dim, n_group, n_tok);
         } else {
-        GGML_ABORT("doesn't support N!=16.");
+            GGML_ABORT("doesn't support d_state!=128.");
+        }
+    } else {
+        // Mamba-1
+        GGML_ASSERT(n_head % threads == 0);
+        GGML_ASSERT(head_dim == 1);
+        GGML_ASSERT(n_group == 1);
+        const dim3 blocks(n_seq, (n_head + threads - 1) / threads, 1);
+        const int  smem_size = (threads * (d_state + 1) * 2) * sizeof(float);
+        if (d_state == 16) {
+            ssm_scan_f32<128, 16><<<blocks, threads, smem_size, stream>>>(
+                src0, src1, src2, src3, src4, src5, src6, dst,
+                src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
+                src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
+        } else {
+            GGML_ABORT("doesn't support d_state!=16.");
+        }
     }
 }
 
@@ -112,30 +243,25 @@ void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const struct ggml_tensor * src3 = dst->src[3];  // A
     const struct ggml_tensor * src4 = dst->src[4];  // B
     const struct ggml_tensor * src5 = dst->src[5];  // C
-
+    const struct ggml_tensor * src6 = dst->src[6];  // ids
-    //   const int64_t d_state = src0->ne[0];
-    //   const int64_t d_inner = src0->ne[1];
-    //   const int64_t l = src1->ne[1];
-    //   const int64_t b = src0->ne[2];
 
     const int64_t nc  = src0->ne[0];  // d_state
-    const int64_t nr  = src0->ne[1];  // d_inner
+    const int64_t nr  = src0->ne[1];  // head_dim or 1
-    const int64_t n_t = src1->ne[1];  // number of tokens per sequence
+    const int64_t nh  = src1->ne[1];  // n_head
-    const int64_t n_s = src0->ne[2];  // number of sequences in the batch
+    const int64_t ng  = src4->ne[1];  // n_group
+    const int64_t n_t = src1->ne[2];  // number of tokens per sequence
+    const int64_t n_s = src1->ne[3];  // number of sequences in the batch
 
-    GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst));
+    const int64_t s_off = ggml_nelements(src1) * sizeof(float);
+
+    GGML_ASSERT(ggml_nelements(src1) + nc*nr*nh*n_s == ggml_nelements(dst));
     GGML_ASSERT(src0->nb[0] == sizeof(float));
     GGML_ASSERT(src1->nb[0] == sizeof(float));
     GGML_ASSERT(src2->nb[0] == sizeof(float));
     GGML_ASSERT(src3->nb[0] == sizeof(float));
     GGML_ASSERT(src4->nb[0] == sizeof(float));
     GGML_ASSERT(src5->nb[0] == sizeof(float));
-    // required for the dot product between s and C
+    GGML_ASSERT(src6->nb[0] == sizeof(int32_t));
-    GGML_ASSERT(src0->nb[1] == src0->ne[0] * sizeof(float));
-    // required for per-sequence offsets for states
-    GGML_ASSERT(src0->nb[2] == src0->ne[0] * src0->ne[1] * sizeof(float));
-    // required to get correct offset for state destination (i.e. src1->nb[3])
-    GGML_ASSERT(src1->nb[3] == src1->ne[0] * src1->ne[1] * src1->ne[2] * sizeof(float));
 
     const float * src0_d = (const float *) src0->data;
     const float * src1_d = (const float *) src1->data;
@@ -143,13 +269,16 @@ void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const float * src3_d = (const float *) src3->data;
     const float * src4_d = (const float *) src4->data;
     const float * src5_d = (const float *) src5->data;
+    const int32_t * src6_d = (const int32_t *) src6->data;
     float *       dst_d  = (float *) dst->data;
     cudaStream_t  stream = ctx.stream();
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src6->type == GGML_TYPE_I32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
-    ssm_scan_f32_cuda(src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src0->nb[1], src0->nb[2], src1->nb[0],
+    ssm_scan_f32_cuda(src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src6_d, dst_d,
-                      src1->nb[1], src1->nb[2], src1->nb[3], src2->nb[0], src2->nb[1], src2->nb[2], src3->nb[1],
+                      src0->nb[2], src0->nb[3], src1->nb[2], src1->nb[3], src2->nb[1], src2->nb[2],
-                      src4->nb[1], src4->nb[2], src5->nb[1], src5->nb[2], dst_d, nc, nr, n_t, n_s, stream);
+                      src3->nb[1], src4->nb[2], src4->nb[3], src5->nb[2], src5->nb[3],
+                      s_off, nc, nr, nh, ng, n_t, n_s, stream);
 }

ggml/src/ggml-cuda/unary.cu

@@ -285,6 +285,14 @@ void ggml_cuda_op_swiglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     ggml_cuda_op_unary_gated<op_silu>(ctx, dst);
 }
 
+void ggml_cuda_op_geglu_erf(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary_gated<op_gelu_erf>(ctx, dst);
+}
+
+void ggml_cuda_op_geglu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary_gated<op_gelu_quick>(ctx, dst);
+}
+
 /* silu_back */
 
 static __device__ __forceinline__ float op_silu_back(float grad, float x) {

ggml/src/ggml-cuda/unary.cuh

@@ -64,3 +64,7 @@ void ggml_cuda_op_reglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 void ggml_cuda_op_geglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_swiglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_geglu_erf(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_geglu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -229,7 +229,11 @@ typedef struct {
     uint64_t nb21;
     uint64_t nb22;
     uint64_t nb23;
+    int32_t  ne32;
+    int32_t  ne33;
     uint64_t nb31;
+    uint64_t nb32;
+    uint64_t nb33;
     int32_t  ne1;
     int32_t  ne2;
     float    scale;
@@ -461,9 +465,21 @@ typedef struct {
 } ggml_metal_kargs_sum_rows;
 
 typedef struct {
-    int64_t  ne00;
+    int32_t  ne00;
-    int64_t  ne01;
+    int32_t  ne01;
-    int64_t  ne02;
+    int32_t  ne02;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne11;
+    int32_t  ne12;
+    int32_t  ne13;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
     float    scale;
     float    max_bias;
     float    m0;
@@ -499,26 +515,25 @@ typedef struct {
 typedef struct {
     int64_t  d_state;
     int64_t  d_inner;
+    int64_t  n_head;
+    int64_t  n_group;
     int64_t  n_seq_tokens;
     int64_t  n_seqs;
-    uint64_t nb00;
     uint64_t nb01;
     uint64_t nb02;
-    uint64_t nb10;
+    uint64_t nb03;
     uint64_t nb11;
     uint64_t nb12;
     uint64_t nb13;
-    uint64_t nb20;
     uint64_t nb21;
     uint64_t nb22;
-    uint64_t nb30;
     uint64_t nb31;
-    uint64_t nb40;
     uint64_t nb41;
     uint64_t nb42;
-    uint64_t nb50;
+    uint64_t nb43;
     uint64_t nb51;
     uint64_t nb52;
+    uint64_t nb53;
 } ggml_metal_kargs_ssm_scan;
 
 typedef struct {

ggml/src/ggml-metal/ggml-metal.m

@@ -217,6 +217,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_NORM,
     GGML_METAL_KERNEL_TYPE_SSM_CONV_F32,
     GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32,
+    GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32_GROUP,
     GGML_METAL_KERNEL_TYPE_RWKV_WKV6_F32,
     GGML_METAL_KERNEL_TYPE_RWKV_WKV7_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,
@@ -529,6 +530,8 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_REGLU,
     GGML_METAL_KERNEL_TYPE_GEGLU,
     GGML_METAL_KERNEL_TYPE_SWIGLU,
+    GGML_METAL_KERNEL_TYPE_GEGLU_ERF,
+    GGML_METAL_KERNEL_TYPE_GEGLU_QUICK,
     GGML_METAL_KERNEL_TYPE_SUM_ROWS,
     GGML_METAL_KERNEL_TYPE_MEAN,
     GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,
@@ -1196,6 +1199,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM,                            norm,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_CONV_F32,                    ssm_conv_f32,                    true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32,                    ssm_scan_f32,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32_GROUP,              ssm_scan_f32_group,              true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RWKV_WKV6_F32,                   rwkv_wkv6_f32,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RWKV_WKV7_F32,                   rwkv_wkv7_f32,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,                  mul_mv_f32_f32,                  has_simdgroup_reduction);
@@ -1508,6 +1512,8 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REGLU,                           reglu,                           true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GEGLU,                           geglu,                           true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SWIGLU,                          swiglu,                          true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GEGLU_ERF,                       geglu_erf,                       true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GEGLU_QUICK,                     geglu_quick,                     true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                        sum_rows,                        true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MEAN,                            mean,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGMAX,                          argmax,                          true);
@@ -1691,6 +1697,8 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
                 case GGML_GLU_OP_REGLU:
                 case GGML_GLU_OP_GEGLU:
                 case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
                     return ggml_is_contiguous_1(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
                default:
                     return false;
@@ -1725,7 +1733,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
         case GGML_OP_MEAN:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_GROUP_NORM:
-            return has_simdgroup_reduction && ggml_is_contiguous(op->src[0]);
+            return has_simdgroup_reduction && ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_RMS_NORM:
         case GGML_OP_L2_NORM:
             return has_simdgroup_reduction && (op->ne[0] % 4 == 0 && ggml_is_contiguous_1(op->src[0]));
@@ -2454,6 +2462,12 @@ static bool ggml_metal_encode_node(
                     case GGML_GLU_OP_SWIGLU:
                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SWIGLU].pipeline;
                         break;
+                    case GGML_GLU_OP_GEGLU_ERF:
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GEGLU_ERF].pipeline;
+                        break;
+                    case GGML_GLU_OP_GEGLU_QUICK:
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GEGLU_QUICK].pipeline;
+                        break;
                     default:
                         GGML_ABORT("fatal error");
                 }
@@ -2644,10 +2658,7 @@ static bool ggml_metal_encode_node(
                 memcpy(&scale,    ((const int32_t *) dst->op_params) + 0, sizeof(scale));
                 memcpy(&max_bias, ((const int32_t *) dst->op_params) + 1, sizeof(max_bias));
 
-                const int64_t nrows_x = ggml_nrows(src0);
+                const uint32_t n_head      = src0->ne[2];
-                const int64_t nrows_y = src0->ne[1];
-
-                const uint32_t n_head      = nrows_x/nrows_y;
                 const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
 
                 const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
@@ -2707,6 +2718,18 @@ static bool ggml_metal_encode_node(
                     /*.ne00        =*/ ne00,
                     /*.ne01        =*/ ne01,
                     /*.ne02        =*/ ne02,
+                    /*.nb01        =*/ nb01,
+                    /*.nb02        =*/ nb02,
+                    /*.nb03        =*/ nb03,
+                    /*.ne11        =*/ ne11,
+                    /*.ne12        =*/ ne12,
+                    /*.ne13        =*/ ne13,
+                    /*.nb11        =*/ nb11,
+                    /*.nb12        =*/ nb12,
+                    /*.nb13        =*/ nb13,
+                    /*.nb1         =*/ nb1,
+                    /*.nb2         =*/ nb2,
+                    /*.nb3         =*/ nb3,
                     /*.scale       =*/ scale,
                     /*.max_bias    =*/ max_bias,
                     /*.m0          =*/ m0,
@@ -2726,7 +2749,7 @@ static bool ggml_metal_encode_node(
 
                 [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
-                [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
             } break;
         case GGML_OP_DIAG_MASK_INF:
             {
@@ -2800,71 +2823,91 @@ static bool ggml_metal_encode_node(
                 struct ggml_tensor * src3 = node->src[3];
                 struct ggml_tensor * src4 = node->src[4];
                 struct ggml_tensor * src5 = node->src[5];
+                struct ggml_tensor * src6 = node->src[6];
 
                 GGML_ASSERT(src3);
                 GGML_ASSERT(src4);
                 GGML_ASSERT(src5);
+                GGML_ASSERT(src6);
 
                 size_t offs_src3 = 0;
                 size_t offs_src4 = 0;
                 size_t offs_src5 = 0;
+                size_t offs_src6 = 0;
 
                 id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil;
                 id<MTLBuffer> id_src4 = src4 ? ggml_metal_get_buffer(src4, &offs_src4) : nil;
                 id<MTLBuffer> id_src5 = src5 ? ggml_metal_get_buffer(src5, &offs_src5) : nil;
+                id<MTLBuffer> id_src6 = src6 ? ggml_metal_get_buffer(src6, &offs_src6) : nil;
 
-                const int64_t  ne30 = src3->ne[0]; GGML_UNUSED(ne30);
+                const int64_t  ne30 = src3->ne[0];
                 const int64_t  ne31 = src3->ne[1]; GGML_UNUSED(ne31);
 
-                const uint64_t nb30 = src3->nb[0];
+                const uint64_t nb30 = src3->nb[0]; GGML_UNUSED(nb30);
                 const uint64_t nb31 = src3->nb[1];
 
                 const int64_t  ne40 = src4->ne[0]; GGML_UNUSED(ne40);
-                const int64_t  ne41 = src4->ne[1]; GGML_UNUSED(ne41);
+                const int64_t  ne41 = src4->ne[1];
                 const int64_t  ne42 = src4->ne[2]; GGML_UNUSED(ne42);
+                const int64_t  ne43 = src4->ne[3]; GGML_UNUSED(ne43);
 
-                const uint64_t nb40 = src4->nb[0];
+                const uint64_t nb40 = src4->nb[0]; GGML_UNUSED(nb40);
                 const uint64_t nb41 = src4->nb[1];
                 const uint64_t nb42 = src4->nb[2];
+                const uint64_t nb43 = src4->nb[3];
 
                 const int64_t  ne50 = src5->ne[0]; GGML_UNUSED(ne50);
                 const int64_t  ne51 = src5->ne[1]; GGML_UNUSED(ne51);
                 const int64_t  ne52 = src5->ne[2]; GGML_UNUSED(ne52);
+                const int64_t  ne53 = src5->ne[3]; GGML_UNUSED(ne53);
 
-                const uint64_t nb50 = src5->nb[0];
+                const uint64_t nb50 = src5->nb[0]; GGML_UNUSED(nb50);
                 const uint64_t nb51 = src5->nb[1];
                 const uint64_t nb52 = src5->nb[2];
+                const uint64_t nb53 = src5->nb[3];
+
+                const int64_t  ne60 = src6->ne[0]; GGML_UNUSED(ne60);
+
+                const uint64_t nb60 = src6->nb[0]; GGML_UNUSED(nb60);
 
                 const int64_t d_state      = ne00;
                 const int64_t d_inner      = ne01;
-                const int64_t n_seq_tokens = ne11;
+                const int64_t n_head       = ne02;
-                const int64_t n_seqs       = ne02;
+                const int64_t n_group      = ne41;
+                const int64_t n_seq_tokens = ne12;
+                const int64_t n_seqs       = ne13;
 
-                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline;
+                id<MTLComputePipelineState> pipeline = nil;
+
+                if (ne30 == 1) {
+                    // Mamba-2
+                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32_GROUP].pipeline;
+                } else {
+                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline;
+                }
 
                 ggml_metal_kargs_ssm_scan args = {
                     /*.d_state      =*/ d_state,
                     /*.d_inner      =*/ d_inner,
+                    /*.n_head       =*/ n_head,
+                    /*.n_group      =*/ n_group,
                     /*.n_seq_tokens =*/ n_seq_tokens,
                     /*.n_seqs       =*/ n_seqs,
-                    /*.nb00 =*/ nb00,
                     /*.nb01         =*/ nb01,
                     /*.nb02         =*/ nb02,
-                    /*.nb10 =*/ nb10,
+                    /*.nb03         =*/ nb03,
                     /*.nb11         =*/ nb11,
                     /*.nb12         =*/ nb12,
                     /*.nb13         =*/ nb13,
-                    /*.nb20 =*/ nb20,
                     /*.nb21         =*/ nb21,
                     /*.nb22         =*/ nb22,
-                    /*.nb30 =*/ nb30,
                     /*.nb31         =*/ nb31,
-                    /*.nb40 =*/ nb40,
                     /*.nb41         =*/ nb41,
                     /*.nb42         =*/ nb42,
-                    /*.nb50 =*/ nb50,
+                    /*.nb43         =*/ nb43,
                     /*.nb51         =*/ nb51,
                     /*.nb52         =*/ nb52,
+                    /*.nb53         =*/ nb53,
                 };
 
                 [encoder setComputePipelineState:pipeline];
@@ -2874,10 +2917,17 @@ static bool ggml_metal_encode_node(
                 [encoder setBuffer:id_src3 offset:offs_src3 atIndex:3];
                 [encoder setBuffer:id_src4 offset:offs_src4 atIndex:4];
                 [encoder setBuffer:id_src5 offset:offs_src5 atIndex:5];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:6];
+                [encoder setBuffer:id_src6 offset:offs_src6 atIndex:6];
-                [encoder setBytes:&args    length:sizeof(args) atIndex:7];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:7];
+                [encoder setBytes:&args    length:sizeof(args) atIndex:8];
 
-                [encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                if (ne30 == 1) {
+                    // Mamba-2
+                    [encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_head, n_seqs) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } else {
+                    GGML_ASSERT(d_inner == 1);
+                    [encoder dispatchThreadgroups:MTLSizeMake(n_head, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                }
             } break;
         case GGML_OP_RWKV_WKV6:
             {
@@ -4979,7 +5029,11 @@ static bool ggml_metal_encode_node(
                     /*.nb21          =*/ nb21,
                     /*.nb22          =*/ nb22,
                     /*.nb23          =*/ nb23,
+                    /*.ne32          =*/ ne32,
+                    /*.ne33          =*/ ne33,
                     /*.nb31          =*/ nb31,
+                    /*.nb32          =*/ nb32,
+                    /*.nb33          =*/ nb33,
                     /*.ne1           =*/ ne1,
                     /*.ne2           =*/ ne2,
                     /*.scale         =*/ scale,

ggml/src/ggml-metal/ggml-metal.metal

@@ -109,6 +109,7 @@ void dequantize_q4_0_t4(device const block_q4_0 * xb, short il, thread type4 & r
 }
 
 void quantize_q4_0(device const float * src, device block_q4_0 & dst) {
+#pragma METAL fp math_mode(safe)
     float amax = 0.0f; // absolute max
     float max  = 0.0f;
 
@@ -167,6 +168,7 @@ void quantize_q4_1(device const float * src, device block_q4_1 & dst) {
 }
 
 void quantize_q5_0(device const float * src, device block_q5_0 & dst) {
+#pragma METAL fp math_mode(safe)
     float amax = 0.0f; // absolute max
     float max  = 0.0f;
 
@@ -461,6 +463,7 @@ void dequantize_q8_0_t4(device const block_q8_0 *xb, short il, thread type4 & re
 }
 
 void quantize_q8_0(device const float * src, device block_q8_0 & dst) {
+#pragma METAL fp math_mode(safe)
     float amax = 0.0f; // absolute max
 
     for (int j = 0; j < QK8_0; j++) {
@@ -1258,6 +1261,50 @@ kernel void kernel_swiglu(
     }
 }
 
+kernel void kernel_geglu_erf(
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant ggml_metal_kargs_glu & args,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const float * src0_row = (device const float *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const float * src1_row = (device const float *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       float * dst_row  = (device       float *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_erf = 0.5f*x0*(1.0f+erf_approx<float>(x0*SQRT_2_INV));
+
+        dst_row[i0] = gelu_erf*x1;
+    }
+}
+
+kernel void kernel_geglu_quick(
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant ggml_metal_kargs_glu & args,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const float * src0_row = (device const float *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const float * src1_row = (device const float *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       float * dst_row  = (device       float *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_quick = x0*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x0)));
+
+        dst_row[i0] = gelu_quick*x1;
+    }
+}
+
 template <bool norm>
 kernel void kernel_sum_rows(
         constant ggml_metal_kargs_sum_rows & args,
@@ -1320,24 +1367,28 @@ kernel void kernel_soft_max(
         device        char * dst,
         constant ggml_metal_kargs_soft_max & args,
         threadgroup  float * buf [[threadgroup(0)]],
-        uint  tgpig[[threadgroup_position_in_grid]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tpitg[[thread_position_in_threadgroup]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
         uint  sgitg[[simdgroup_index_in_threadgroup]],
         uint  tiisg[[thread_index_in_simdgroup]],
-        uint    ntg[[threads_per_threadgroup]]) {
+        uint3  tptg[[threads_per_threadgroup]]) {
-    const int64_t i03 = (tgpig) / (args.ne02*args.ne01);
+    const int32_t i03 = tgpig.z;
-    const int64_t i02 = (tgpig - i03*args.ne02*args.ne01) / args.ne01;
+    const int32_t i02 = tgpig.y;
-    const int64_t i01 = (tgpig - i03*args.ne02*args.ne01 - i02*args.ne01);
+    const int32_t i01 = tgpig.x;
 
-    device const float * psrc0 = (device const float *) src0 + (i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00);
+    const int32_t i13 = i03%args.ne13;
-    device const     T * pmask = src1 != src0 ? (device const    T *) src1         + i01*args.ne00 : nullptr;
+    const int32_t i12 = i02%args.ne12;
-    device       float * pdst  = (device       float *) dst  + (i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00);
+    const int32_t i11 = i01;
+
+    device const float * psrc0 =                (device const float *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+    device const     T * pmask = src1 != src0 ? (device const T *    ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
+    device       float * pdst  =                (device       float *) (dst  + i01*args.nb1  + i02*args.nb2  + i03*args.nb3);
 
     float slope = 1.0f;
 
     // ALiBi
     if (args.max_bias > 0.0f) {
-        const int64_t h = i02;
+        const int32_t h = i02;
 
         const float base = h < args.n_head_log2 ? args.m0 : args.m1;
         const int   exp  = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
@@ -1348,13 +1399,13 @@ kernel void kernel_soft_max(
     // parallel max
     float lmax = -INFINITY;
 
-    for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) {
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
         lmax = MAX(lmax, psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f));
     }
 
     // find the max value in the block
     float max_val = simd_max(lmax);
-    if (ntg > N_SIMDWIDTH) {
+    if (tptg.x > N_SIMDWIDTH) {
         if (sgitg == 0) {
             buf[tiisg] = -INFINITY;
         }
@@ -1373,7 +1424,7 @@ kernel void kernel_soft_max(
 
     // parallel sum
     float lsum = 0.0f;
-    for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) {
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
         const float exp_psrc0 = exp((psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f)) - max_val);
         lsum += exp_psrc0;
         pdst[i00] = exp_psrc0;
@@ -1385,7 +1436,7 @@ kernel void kernel_soft_max(
 
     float sum = simd_sum(lsum);
 
-    if (ntg > N_SIMDWIDTH) {
+    if (tptg.x > N_SIMDWIDTH) {
         if (sgitg == 0) {
             buf[tiisg] = 0.0f;
         }
@@ -1404,7 +1455,7 @@ kernel void kernel_soft_max(
 
     const float inv_sum = 1.0f/sum;
 
-    for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) {
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
         pdst[i00] *= inv_sum;
     }
 }
@@ -1416,23 +1467,27 @@ kernel void kernel_soft_max_4(
         device        char * dst,
         constant ggml_metal_kargs_soft_max & args,
         threadgroup  float * buf [[threadgroup(0)]],
-        uint  tgpig[[threadgroup_position_in_grid]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tpitg[[thread_position_in_threadgroup]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
         uint  sgitg[[simdgroup_index_in_threadgroup]],
         uint  tiisg[[thread_index_in_simdgroup]],
-        uint    ntg[[threads_per_threadgroup]]) {
+        uint3  tptg[[threads_per_threadgroup]]) {
-    const int64_t i03 = (tgpig) / (args.ne02*args.ne01);
+    const int32_t i03 = tgpig.z;
-    const int64_t i02 = (tgpig - i03*args.ne02*args.ne01) / args.ne01;
+    const int32_t i02 = tgpig.y;
-    const int64_t i01 = (tgpig - i03*args.ne02*args.ne01 - i02*args.ne01);
+    const int32_t i01 = tgpig.x;
 
-    device const float4 * psrc4 = (device const float4 *) src0 + (i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00)/4;
+    const int32_t i13 = i03%args.ne13;
-    device const      T * pmask = src1 != src0 ? (device const     T *) src1         + i01*args.ne00/4 : nullptr;
+    const int32_t i12 = i02%args.ne12;
-    device       float4 * pdst4 = (device       float4 *) dst  + (i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00)/4;
+    const int32_t i11 = i01;
+
+    device const float4 * psrc4 =                (device const float4 *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+    device const      T * pmask = src1 != src0 ? (device const T *     ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
+    device       float4 * pdst4 =                (device       float4 *) (dst  + i01*args.nb1  + i02*args.nb2  + i03*args.nb3);
 
     float slope = 1.0f;
 
     if (args.max_bias > 0.0f) {
-        const int64_t h = i02;
+        const int32_t h = i02;
 
         const float base = h < args.n_head_log2 ? args.m0 : args.m1;
         const int   exp  = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
@@ -1443,14 +1498,14 @@ kernel void kernel_soft_max_4(
     // parallel max
     float4 lmax4 = -INFINITY;
 
-    for (int i00 = tpitg; i00 < args.ne00/4; i00 += ntg) {
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
         lmax4 = fmax(lmax4, psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f)));
     }
 
     const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
 
     float max_val = simd_max(lmax);
-    if (ntg > N_SIMDWIDTH) {
+    if (tptg.x > N_SIMDWIDTH) {
         if (sgitg == 0) {
             buf[tiisg] = -INFINITY;
         }
@@ -1469,7 +1524,7 @@ kernel void kernel_soft_max_4(
 
     // parallel sum
     float4 lsum4 = 0.0f;
-    for (int i00 = tpitg; i00 < args.ne00/4; i00 += ntg) {
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
         const float4 exp_psrc4 = exp((psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f))) - max_val);
         lsum4 += exp_psrc4;
         pdst4[i00] = exp_psrc4;
@@ -1483,7 +1538,7 @@ kernel void kernel_soft_max_4(
 
     float sum = simd_sum(lsum);
 
-    if (ntg > N_SIMDWIDTH) {
+    if (tptg.x > N_SIMDWIDTH) {
         if (sgitg == 0) {
             buf[tiisg] = 0.0f;
         }
@@ -1502,7 +1557,7 @@ kernel void kernel_soft_max_4(
 
     const float inv_sum = 1.0f/sum;
 
-    for (int i00 = tpitg; i00 < args.ne00/4; i00 += ntg) {
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
         pdst4[i00] *= inv_sum;
     }
 }
@@ -1588,7 +1643,7 @@ kernel void kernel_ssm_conv_f32(
     x[0] = sumf;
 }
 
-// ref: ggml.c:ggml_compute_forward_ssm_scan_f32
+// ref: ggml.c:ggml_compute_forward_ssm_scan_f32, Mamba-1 part
 kernel void kernel_ssm_scan_f32(
         device const void * src0,
         device const void * src1,
@@ -1596,46 +1651,119 @@ kernel void kernel_ssm_scan_f32(
         device const void * src3,
         device const void * src4,
         device const void * src5,
+        device const void * src6,
         device      float * dst,
         constant ggml_metal_kargs_ssm_scan & args,
         uint3 tgpig[[threadgroup_position_in_grid]],
         uint3 tpitg[[thread_position_in_threadgroup]],
         uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t ir = tgpig.x;
+    const int64_t i1 = 0;
-    const int64_t i3 = tgpig.y;
+    const int64_t ir = tgpig.x; // current head
+    const int64_t i3 = tgpig.y; // current seq
+
+    const uint64_t nb00 = sizeof(float);
+    const uint64_t nb10 = sizeof(float);
+    const uint64_t nb20 = sizeof(float);
 
     const int64_t nc  = args.d_state;
-    // const int64_t nr  = args.d_inner;
+    const int64_t nr  = args.d_inner;
+    const int64_t nh  = args.n_head;
+    const int64_t ng  = args.n_group;
     const int64_t n_t = args.n_seq_tokens;
-    // const int64_t n_s = args.n_seqs;
+
+    const int64_t s_off = nr * nh * n_t * args.n_seqs * sizeof(float);
+
+    device const int32_t * ids = (device const int32_t *) src6;
+
+    device const float * s0 = (device const float *) ((device const char *) src0 + ir*args.nb02 + ids[i3]*args.nb03);
+    device       float * s  = (device       float *) ((device       char *) dst  + ir*args.nb02 +      i3*args.nb03 + s_off);
 
     for (int64_t i2 = 0; i2 < n_t; ++i2) {
-        device const float * s0 = (device const float *) ((device const char *) src0 + ir*args.nb01 + i3*args.nb02);
+        device const float * x  = (device const float *) ((device const char *) src1 + i1*nb10 + ir*args.nb11 + i2*args.nb12 + i3*args.nb13); // {dim, nh, nt, ns}
-        device const float * x  = (device const float *) ((device const char *) src1 + ir*args.nb10 + i2*args.nb11 + i3*args.nb12);
+        device const float * dt = (device const float *) ((device const char *) src2 + ir*nb20 + i2*args.nb21 + i3*args.nb22); // {nh, nt, ns}
-        device const float * dt = (device const float *) ((device const char *) src2 + ir*args.nb20 + i2*args.nb21 + i3*args.nb22);
+        device const float * A  = (device const float *) ((device const char *) src3 + ir*args.nb31); // {d_state, nh}
-        device const float * A  = (device const float *) ((device const char *) src3 + ir*args.nb31);
+        device const float * B  = (device const float *) ((device const char *) src4 + (ir & (ng - 1))*args.nb41 + i2*args.nb42 + i3*args.nb43); // {d_state, ng, nt, ns}
-        device const float * B  = (device const float *) ((device const char *) src4 + i2*args.nb41 + i3*args.nb42);
+        device const float * C  = (device const float *) ((device const char *) src5 + (ir & (ng - 1))*args.nb51 + i2*args.nb52 + i3*args.nb53); // {d_state, ng, nt, ns}
-        device const float * C  = (device const float *) ((device const char *) src5 + i2*args.nb51 + i3*args.nb52);
+        device       float * y  = (device       float *) ((device       char *) dst  + (i1 + ir*(nr) + i2*(nh*nr) + i3*(n_t*nh*nr))*nb00); // {dim, nh, nt, ns}
-        device       float * y  = (device       float *) ((device       char *) dst  + ir*args.nb10 + i2*args.nb11 + i3*args.nb12); // TODO: do not use src1 strides
-        device       float * s  = (device       float *) ((device       char *) dst  + ir*args.nb01 + i3*args.nb02 +    args.nb13);
 
-        if (i2 > 0) {
+        const float dt_soft_plus = dt[0] <= 20.0f ? log(1.0f + exp(dt[0])) : dt[0];
-            s0 = s;
+        const float x_dt = x[0] * dt_soft_plus;
-        }
-
-        // i1 == 0
-        float dt_soft_plus = dt[0] <= 20.0f ? log(1.0f + exp(dt[0])) : dt[0];
-        float x_dt = x[0] * dt_soft_plus;
         float sumf = 0.0f;
 
         for (int64_t i0 = 0; i0 < nc; ++i0) {
-            int64_t i = i0;
+            const int64_t i = i0 + i1*nc;
-            float state = (s0[i] * exp(dt_soft_plus * A[i])) + (B[i0] * x_dt);
+            const float state = (s0[i] * exp(dt_soft_plus * A[i0])) + (B[i0] * x_dt);
             sumf += state * C[i0];
             s[i] = state;
         }
 
         y[0] = sumf;
+
+        // recurse
+        s0 = s;
+    }
+}
+
+// ref: ggml.c:ggml_compute_forward_ssm_scan_f32, Mamba-2 part
+// TODO: optimize (e.g. by parallelizing over d_state)
+kernel void kernel_ssm_scan_f32_group(
+        device const void * src0,
+        device const void * src1,
+        device const void * src2,
+        device const void * src3,
+        device const void * src4,
+        device const void * src5,
+        device const void * src6,
+        device      float * dst,
+        constant ggml_metal_kargs_ssm_scan & args,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i1 = tgpig.x;
+    const int64_t ir = tgpig.y; // current head
+    const int64_t i3 = tgpig.z; // current seq
+
+    const uint64_t nb00 = sizeof(float);
+    const uint64_t nb10 = sizeof(float);
+    const uint64_t nb20 = sizeof(float);
+
+    const int64_t nc  = args.d_state;
+    const int64_t nr  = args.d_inner;
+    const int64_t nh  = args.n_head;
+    const int64_t ng  = args.n_group;
+    const int64_t n_t = args.n_seq_tokens;
+
+    const int64_t s_off = nr * nh * n_t * args.n_seqs * sizeof(float);
+
+    device const int32_t * ids = (device const int32_t *) src6;
+
+    device const float * s0 = (device const float *) ((device const char *) src0 + ir*args.nb02 + ids[i3]*args.nb03);
+    device       float * s  = (device       float *) ((device       char *) dst  + ir*args.nb02 +      i3*args.nb03 + s_off);
+
+    for (int64_t i2 = 0; i2 < n_t; ++i2) {
+        device const float * x  = (device const float *) ((device const char *) src1 + i1*nb10 + ir*args.nb11 + i2*args.nb12 + i3*args.nb13); // {dim, nh, nt, ns}
+        device const float * dt = (device const float *) ((device const char *) src2 + ir*nb20 + i2*args.nb21 + i3*args.nb22); // {nh, nt, ns}
+        device const float * A  = (device const float *) ((device const char *) src3 + ir*args.nb31); // {1, nh}
+        device const float * B  = (device const float *) ((device const char *) src4 + (ir & (ng - 1))*args.nb41 + i2*args.nb42 + i3*args.nb43); // {d_state, ng, nt, ns}
+        device const float * C  = (device const float *) ((device const char *) src5 + (ir & (ng - 1))*args.nb51 + i2*args.nb52 + i3*args.nb53); // {d_state, ng, nt, ns}
+        device       float * y  = (device       float *) ((device       char *) dst  + (i1 + ir*(nr) + i2*(nh*nr) + i3*(n_t*nh*nr))*nb00); // {dim, nh, nt, ns}
+
+        const float dt_soft_plus = dt[0] <= 20.0f ? log(1.0f + exp(dt[0])) : dt[0];
+        const float x_dt = x[0] * dt_soft_plus;
+        const float dA = exp(dt_soft_plus * A[0]);
+        float sumf = 0.0f;
+
+        for (int64_t i0 = 0; i0 < nc; ++i0) {
+            const int64_t i = i0 + i1*nc;
+            const float state = (s0[i] * dA) + (B[i0] * x_dt);
+            sumf += state * C[i0];
+            s[i] = state;
+        }
+
+        y[0] = sumf;
+
+        // recurse
+        s0 = s;
     }
 }
 
@@ -3776,7 +3904,7 @@ kernel void kernel_flash_attn_ext(
                 // load the mask in shared memory
                 #pragma unroll(Q)
                 for (short j = 0; j < Q; ++j) {
-                    device const half * pm = (device const half *) ((device const char *) mask + (iq1 + j)*args.nb31);
+                    device const half * pm = (device const half *) ((device const char *) mask + (iq1 + j)*args.nb31 + (iq2%args.ne32)*args.nb32 + (iq3%args.ne33)*args.nb33);
 
                     const float m = pm[ic + tiisg];
 
@@ -4262,7 +4390,7 @@ kernel void kernel_flash_attn_ext_vec(
         const bool has_mask = mask != q;
 
         // pointer to the mask
-        device const half * pm = (device const half *) (mask + iq1*args.nb31);
+        device const half * pm = (device const half *) (mask + iq1*args.nb31 + (iq2%args.ne32)*args.nb32 + (iq3%args.ne33)*args.nb33);
 
         float slope = 1.0f;
 

ggml/src/ggml-opencl/kernels/glu.cl

@@ -1,7 +1,9 @@
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
 
 #define GELU_COEF_A     0.044715f
+#define GELU_QUICK_COEF -1.702f
 #define SQRT_2_OVER_PI  0.79788456080286535587989211986876f
+#define SQRT_2_INV      0.70710678118654752440084436210484f
 
 //------------------------------------------------------------------------------
 // geglu
@@ -199,3 +201,137 @@ kernel void kernel_swiglu_f16(
         dst_row[i0] = silu*x1;
     }
 }
+
+//------------------------------------------------------------------------------
+// geglu_erf
+//------------------------------------------------------------------------------
+kernel void kernel_geglu_erf(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global float * dst_row  = (global float *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_erf = 0.5f*x0*(1.0f + erf(x0*SQRT_2_INV));
+
+        dst_row[i0] = gelu_erf*x1;
+    }
+}
+
+kernel void kernel_geglu_erf_f16(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global half * dst_row  = (global half *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const half x0 = src0_row[i0];
+        const half x1 = src1_row[i0];
+
+        const half gelu_erf = 0.5f*x0*(1.0f + erf(x0*SQRT_2_INV));
+
+        dst_row[i0] = gelu_erf*x1;
+    }
+}
+
+//------------------------------------------------------------------------------
+// geglu_quick
+//------------------------------------------------------------------------------
+kernel void kernel_geglu_quick(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global float * dst_row  = (global float *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_quick = x0*(1.0f/(1.0f + exp(GELU_QUICK_COEF*x0)));
+
+        dst_row[i0] = gelu_quick*x1;
+    }
+}
+
+kernel void kernel_geglu_quick_f16(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global half * dst_row  = (global half *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const half x0 = src0_row[i0];
+        const half x1 = src1_row[i0];
+
+        const half gelu_quick = x0*(1.0f/(1.0f + exp(GELU_QUICK_COEF*x0)));
+
+        dst_row[i0] = gelu_quick*x1;
+    }
+}

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -240,6 +240,21 @@ enum vk_device_architecture {
     INTEL_XE2,
 };
 
+// HSK x HSV
+enum FaHeadSizes {
+    FA_HEAD_SIZE_64,
+    FA_HEAD_SIZE_80,
+    FA_HEAD_SIZE_96,
+    FA_HEAD_SIZE_112,
+    FA_HEAD_SIZE_128,
+    FA_HEAD_SIZE_192,
+    FA_HEAD_SIZE_192_128,
+    FA_HEAD_SIZE_256,
+    FA_HEAD_SIZE_576_512,
+    FA_HEAD_SIZE_UNSUPPORTED,
+    FA_HEAD_SIZE_COUNT = FA_HEAD_SIZE_UNSUPPORTED,
+};
+
 static vk_device_architecture get_device_architecture(const vk::PhysicalDevice& device) {
     vk::PhysicalDeviceProperties props = device.getProperties();
 
@@ -457,6 +472,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_geglu[2];
     vk_pipeline pipeline_reglu[2];
     vk_pipeline pipeline_swiglu[2];
+    vk_pipeline pipeline_geglu_erf[2];
+    vk_pipeline pipeline_geglu_quick[2];
 
     vk_pipeline pipeline_leaky_relu_f32;
     vk_pipeline pipeline_silu_back_f32;
@@ -483,26 +500,11 @@ struct vk_device_struct {
     vk_pipeline pipeline_conv2d_dw_cwhn_f32;
 
     // [2][2][2] is for {f16acc,f32acc}x{large,small_rows}x{unaligned, aligned}
-    vk_pipeline pipeline_flash_attn_f32_f16_D64_cm2[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_cm2[GGML_TYPE_COUNT][FA_HEAD_SIZE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D80_cm2[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D96_cm2[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D112_cm2[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D128_cm2[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D256_cm2[GGML_TYPE_COUNT][2][2][2];
 
-    vk_pipeline pipeline_flash_attn_f32_f16_D64_cm1[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_cm1[GGML_TYPE_COUNT][FA_HEAD_SIZE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D80_cm1[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D96_cm1[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D112_cm1[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D128_cm1[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D256_cm1[GGML_TYPE_COUNT][2][2][2];
 
-    vk_pipeline pipeline_flash_attn_f32_f16_D64[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16[GGML_TYPE_COUNT][FA_HEAD_SIZE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D80[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D96[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D112[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D128[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D256[GGML_TYPE_COUNT][2][2][2];
 
     vk_pipeline pipeline_flash_attn_split_k_reduce;
 
@@ -649,6 +651,7 @@ struct vk_flash_attn_push_constants {
     uint32_t nev2;
     uint32_t nev3;
     uint32_t nem1;
+    uint32_t nem2;
 
     uint32_t nb01;
     uint32_t nb02;
@@ -659,7 +662,6 @@ struct vk_flash_attn_push_constants {
     uint32_t nb21;
     uint32_t nb22;
     uint32_t nb23;
-    uint32_t nb31;
 
     float scale;
     float max_bias;
@@ -674,6 +676,7 @@ struct vk_flash_attn_push_constants {
     uint32_t split_kv;
     uint32_t k_num;
 };
+static_assert(sizeof(vk_flash_attn_push_constants) <= 128, "sizeof(vk_flash_attn_push_constants) must be <= 128");
 
 struct vk_op_push_constants {
     uint32_t KX;
@@ -772,6 +775,14 @@ struct vk_op_rope_push_constants {
 struct vk_op_soft_max_push_constants {
     uint32_t KX;
     uint32_t KY;
+    uint32_t ne00;
+    uint32_t ne01;
+    uint32_t ne02;
+    uint32_t ne12;
+    uint32_t ne13;
+    uint32_t nb11;
+    uint32_t nb12;
+    uint32_t nb13;
     float scale;
     float max_bias;
     float m0;
@@ -1010,7 +1021,7 @@ struct ggml_backend_vk_context {
 
     // number of additional consecutive nodes that are being fused with the
     // node currently being processed
-    uint32_t num_additional_fused_ops {};
+    int num_additional_fused_ops {};
 };
 
 static void * const vk_ptr_base = (void *)(uintptr_t) 0x1000;  // NOLINT
@@ -1706,6 +1717,35 @@ enum FaCodePath {
     FA_COOPMAT2,
 };
 
+static FaHeadSizes fa_get_head_sizes(uint32_t hsk, uint32_t hsv) {
+    if (hsk != 192 && hsk != 576 && hsk != hsv) {
+        return FA_HEAD_SIZE_UNSUPPORTED;
+    }
+    switch (hsk) {
+    case 64: return FA_HEAD_SIZE_64;
+    case 80: return FA_HEAD_SIZE_80;
+    case 96: return FA_HEAD_SIZE_96;
+    case 112: return FA_HEAD_SIZE_112;
+    case 128: return FA_HEAD_SIZE_128;
+    case 192:
+        if (hsv == 192) {
+            return FA_HEAD_SIZE_192;
+        } else if (hsv == 128) {
+            return FA_HEAD_SIZE_192_128;
+        } else {
+            return FA_HEAD_SIZE_UNSUPPORTED;
+        }
+    case 256: return FA_HEAD_SIZE_256;
+    case 576:
+        if (hsv == 512) {
+            return FA_HEAD_SIZE_576_512;
+        } else {
+            return FA_HEAD_SIZE_UNSUPPORTED;
+        }
+    default: return FA_HEAD_SIZE_UNSUPPORTED;
+    }
+}
+
 // number of rows/cols for flash attention shader
 static constexpr uint32_t flash_attention_num_small_rows = 32;
 static constexpr uint32_t scalar_flash_attention_num_small_rows = 1;
@@ -1726,8 +1766,9 @@ static uint32_t get_fa_num_small_rows(FaCodePath path) {
     }
 }
 
-static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) {
+static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows) {
     GGML_UNUSED(clamp);
+    GGML_UNUSED(hsv);
 
     if (path == FA_SCALAR) {
         if (small_rows) {
@@ -1751,7 +1792,7 @@ static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t D, uint32_
     }
 
     // small cols to reduce register count
-    if (ggml_is_quantized(type) || D == 256) {
+    if (ggml_is_quantized(type) || hsk >= 256) {
         return {64, 32};
     }
     return {64, 64};
@@ -2044,19 +2085,21 @@ static void ggml_vk_load_shaders(vk_device& device) {
                                       parameter_count, wg_denoms, specialization_constants, disable_robustness, require_full_subgroups, required_subgroup_size));
     };
 
-    auto const &fa_wg_denoms = [&](FaCodePath path, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::array<uint32_t, 3> {
+    auto const &fa_wg_denoms = [&](FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows) -> std::array<uint32_t, 3> {
-        return {fa_rows_cols(path, D, clamp, type, small_rows)[0], 1, 1};
+        return {fa_rows_cols(path, hsk, hsv, clamp, type, small_rows)[0], 1, 1};
     };
 
-    auto const &fa_spec_constants = [&](FaCodePath path, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::vector<uint32_t> {
+    auto const &fa_spec_constants = [&](FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows) -> std::vector<uint32_t> {
         // For large number of rows, 128 invocations seems to work best.
         // For small number of rows (e.g. N==1), 256 works better. But matrix granularity for 256 is 32, so we
         // can't use 256 for D==80.
         // For scalar, use 128 (arbitrary)
+        // The same D_split value is used for both HSK and HSV, so just base it on the union of the LSBs.
+        const uint32_t D = (hsk|hsv);
         uint32_t wg_size = (path == FA_SCALAR || path == FA_COOPMAT1)
                             ? scalar_flash_attention_workgroup_size
                             : ((small_rows && (D % 32) == 0) ? 256 : 128);
-        auto rows_cols = fa_rows_cols(path, D, clamp, type, small_rows);
+        auto rows_cols = fa_rows_cols(path, hsk, hsv, clamp, type, small_rows);
 
         // D_split can't be larger than a subgroup because we use subgroupShuffle to reduce it.
         // D_split can't be larger than the LSB of D divided by 4 due to vectorization in the shader.
@@ -2065,26 +2108,29 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
         // mask dim1 is padded to 64, we rely on this to avoid clamping mask loads
         GGML_ASSERT((GGML_KQ_MASK_PAD % rows_cols[0]) == 0);
-        return {wg_size, rows_cols[0], rows_cols[1], (D), clamp, D_split};
+        return {wg_size, rows_cols[0], rows_cols[1], hsk, hsv, clamp, D_split};
     };
 
-#define CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, D) \
+#define CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, HSK, HSV, HEAD_SIZES) \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][0], "flash_attn_f32_f16_D" #D "_f16acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,false), fa_spec_constants(FAPATH, D,1,TYPE,false), 1,                                      true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][0][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f16acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,false), 1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,false), fa_spec_constants(FAPATH, D,0,TYPE,false), fa_rows_cols(FAPATH,D,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][0][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f16acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,false), fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][0], "flash_attn_f32_f16_D" #D "_f32acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,false), fa_spec_constants(FAPATH, D,1,TYPE,false), 1,                                      true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][0][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f32acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,false), 1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,false), fa_spec_constants(FAPATH, D,0,TYPE,false), fa_rows_cols(FAPATH,D,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][0][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f32acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,false), fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][0], "flash_attn_f32_f16_D" #D "_f16acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,true), fa_spec_constants(FAPATH, D,1,TYPE,true),   1,                                      true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][1][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f16acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,true),   1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,true), fa_spec_constants(FAPATH, D,0,TYPE,true),   fa_rows_cols(FAPATH,D,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][1][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,true),   fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][0], "flash_attn_f32_f16_D" #D "_f32acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,true), fa_spec_constants(FAPATH, D,1,TYPE,true),   1,                                      true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][1][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f32acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,true),   1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,true), fa_spec_constants(FAPATH, D,0,TYPE,true),   fa_rows_cols(FAPATH,D,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][1][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,true),   fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
 
 #define CREATE_FA(TYPE, NAMELC, FAPATH, SUFFIX) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 64) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 64, 64, 64) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 80) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 80, 80, 80) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 96) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 96, 96, 96) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 112) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 112, 112, 112) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 128) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 128, 128, 128) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 256)
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 192, 192, 192) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 192, 128, 192_128) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 256, 256, 256) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 576, 512, 576_512)
 
     CREATE_FA(GGML_TYPE_F16, f16, FA_SCALAR, )
     CREATE_FA(GGML_TYPE_Q4_0, q4_0, FA_SCALAR, )
@@ -2793,6 +2839,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
     CREATE_GLU(geglu)
     CREATE_GLU(reglu)
     CREATE_GLU(swiglu)
+    CREATE_GLU(geglu_erf)
+    CREATE_GLU(geglu_quick)
 #undef CREATE_GLU
 
     ggml_vk_create_pipeline(device, device->pipeline_leaky_relu_f32, "leaky_relu_f32", leaky_relu_f32_len, leaky_relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
@@ -3703,7 +3751,6 @@ static void ggml_vk_instance_init() {
 
     }
 
-    size_t num_available_devices = vk_instance.instance.enumeratePhysicalDevices().size();
     vk_perf_logger_enabled = getenv("GGML_VK_PERF_LOGGER") != nullptr;
 
     // Emulate behavior of CUDA_VISIBLE_DEVICES for Vulkan
@@ -6017,24 +6064,47 @@ static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx
     }
 }
 
-static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const uint32_t D, bool f32acc) {
+static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, const uint32_t hsk, uint32_t hsv) {
     // Needs to be kept up to date on shader changes
+    GGML_UNUSED(hsv);
     const uint32_t wg_size = scalar_flash_attention_workgroup_size;
     const uint32_t Br = scalar_flash_attention_num_large_rows;
     const uint32_t Bc = scalar_flash_attention_Bc;
 
+    const uint32_t tmpsh = wg_size * sizeof(float);
+    const uint32_t tmpshv4 = wg_size * 4 * sizeof(float);
+
+    const uint32_t masksh = Bc * Br * sizeof(float);
+
+    const uint32_t Qf = Br * (hsk / 4 + 2) * 4 * sizeof(float);
+
+    const uint32_t total_size = tmpsh + tmpshv4 + masksh + Qf;
+    const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize;
+
+    VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", total_size=" << total_size << ", supported=" << supported);
+
+    return supported;
+}
+
+static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const uint32_t hsk, uint32_t hsv, bool f32acc) {
+    // Needs to be kept up to date on shader changes
+    GGML_UNUSED(hsv);
+    const uint32_t wg_size = scalar_flash_attention_workgroup_size;
+    const uint32_t Br = coopmat1_flash_attention_num_large_rows;
+    const uint32_t Bc = scalar_flash_attention_Bc;
+
     const uint32_t acctype = f32acc ? 4 : 2;
     const uint32_t f16vec4 = 8;
 
     const uint32_t tmpsh = wg_size * sizeof(float);
     const uint32_t tmpshv4 = wg_size * 4 * acctype;
 
-    const uint32_t Qf = Br * (D / 4 + 2) * f16vec4;
+    const uint32_t Qf = Br * (hsk / 4 + 2) * f16vec4;
 
-    const uint32_t sfshstride = (D <= 128) ? (Br + 8) : Br;
+    const uint32_t sfshstride = (hsk <= 128) ? (Br + 8) : Br;
     const uint32_t sfsh = Bc * sfshstride * acctype;
 
-    const uint32_t kshstride = D / 4 + 2;
+    const uint32_t kshstride = hsk / 4 + 2;
     const uint32_t ksh = Bc * kshstride * f16vec4;
 
     const uint32_t slope = Br * sizeof(float);
@@ -6042,7 +6112,7 @@ static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, co
     const uint32_t total_size = tmpsh + tmpshv4 + Qf + sfsh + ksh + slope;
     const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize;
 
-    VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(D=" << D << ", f32acc=" << f32acc << ", total_size=" << total_size << ", supported=" << supported);
+    VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", f32acc=" << f32acc << ", total_size=" << total_size << ", supported=" << supported);
 
     return supported;
 }
@@ -6064,13 +6134,14 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
 
     const uint32_t nem1 = mask ? mask->ne[1] : 0;
-    const uint32_t nbm1 = mask ? mask->nb[1] : 0;
+    const uint32_t nem2 = mask ? mask->ne[2] : 0;
 
-    const uint32_t D = neq0;
+    const uint32_t HSK = nek0;
+    const uint32_t HSV = nev0;
     uint32_t N = neq1;
     const uint32_t KV = nek1;
 
-    GGML_ASSERT(ne0 == D);
+    GGML_ASSERT(ne0 == HSV);
     GGML_ASSERT(ne2 == N);
 
     // input tensor rows must be contiguous
@@ -6078,12 +6149,9 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     GGML_ASSERT(nbk0 == ggml_type_size(k->type));
     GGML_ASSERT(nbv0 == ggml_type_size(v->type));
 
-    GGML_ASSERT(neq0 == D);
+    GGML_ASSERT(neq0 == HSK);
-    GGML_ASSERT(nek0 == D);
-    GGML_ASSERT(nev0 == D);
 
     GGML_ASSERT(neq1 == N);
-    GGML_ASSERT(nev0 == D);
 
     GGML_ASSERT(nev1 == nek1);
 
@@ -6104,7 +6172,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         const bool coopmat_shape_supported = (dst->op_params[3] == GGML_PREC_F32 && ctx->device->coopmat_support_16x16x16_f32acc) ||
                                              (dst->op_params[3] != GGML_PREC_F32 && ctx->device->coopmat_support_16x16x16_f16acc);
 
-        const bool coopmat_shmem_supported = ggml_vk_flash_attn_coopmat_shmem_support(ctx->device, D, dst->op_params[3] == GGML_PREC_F32);
+        const bool coopmat_shmem_supported = ggml_vk_flash_attn_coopmat_shmem_support(ctx->device, HSK, HSV, dst->op_params[3] == GGML_PREC_F32);
 
         if (!coopmat_shape_supported || !coopmat_shmem_supported) {
             path = FA_SCALAR;
@@ -6157,47 +6225,25 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         path = FA_SCALAR;
     }
 
+    // with large hsk/hsv, scalar path may need to use small_rows to fit in shared memory
+    if (path == FA_SCALAR &&
+        !ggml_vk_flash_attn_scalar_shmem_support(ctx->device, HSK, HSV)) {
+        small_rows = true;
+    }
+
     bool f32acc = path == FA_SCALAR || dst->op_params[3] == GGML_PREC_F32;
 
+    FaHeadSizes head_sizes = fa_get_head_sizes(k->ne[0], v->ne[0]);
+
     switch (path) {
     case FA_SCALAR:
-        switch (D) {
+        pipelines = &ctx->device->pipeline_flash_attn_f32_f16[k->type][head_sizes][f32acc][small_rows][0];
-        case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64[k->type][f32acc][small_rows][0]; break;
-        case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80[k->type][f32acc][small_rows][0]; break;
-        case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96[k->type][f32acc][small_rows][0]; break;
-        case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112[k->type][f32acc][small_rows][0]; break;
-        case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128[k->type][f32acc][small_rows][0]; break;
-        case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256[k->type][f32acc][small_rows][0]; break;
-        default:
-            GGML_ASSERT(!"unsupported D value");
-            return;
-        }
         break;
     case FA_COOPMAT1:
-        switch (D) {
+        pipelines = &ctx->device->pipeline_flash_attn_f32_f16_cm1[k->type][head_sizes][f32acc][small_rows][0];
-        case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64_cm1[k->type][f32acc][small_rows][0]; break;
-        case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80_cm1[k->type][f32acc][small_rows][0]; break;
-        case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96_cm1[k->type][f32acc][small_rows][0]; break;
-        case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112_cm1[k->type][f32acc][small_rows][0]; break;
-        case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128_cm1[k->type][f32acc][small_rows][0]; break;
-        case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256_cm1[k->type][f32acc][small_rows][0]; break;
-        default:
-            GGML_ASSERT(!"unsupported D value");
-            return;
-        }
         break;
     case FA_COOPMAT2:
-        switch (D) {
+        pipelines = &ctx->device->pipeline_flash_attn_f32_f16_cm2[k->type][head_sizes][f32acc][small_rows][0];
-        case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64_cm2[k->type][f32acc][small_rows][0]; break;
-        case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80_cm2[k->type][f32acc][small_rows][0]; break;
-        case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96_cm2[k->type][f32acc][small_rows][0]; break;
-        case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112_cm2[k->type][f32acc][small_rows][0]; break;
-        case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128_cm2[k->type][f32acc][small_rows][0]; break;
-        case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256_cm2[k->type][f32acc][small_rows][0]; break;
-        default:
-            GGML_ASSERT(!"unsupported D value");
-            return;
-        }
         break;
     default:
         GGML_ASSERT(0);
@@ -6227,7 +6273,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     // Try to use split_k when KV is large enough to be worth the overhead
     if (workgroups_x == 1 && shader_core_count > 0 && KV >= 512) {
         // Try to run two workgroups per SM.
-        split_k = ctx->device->shader_core_count * 2 / workgroups_y;
+        split_k = shader_core_count * 2 / (workgroups_y * workgroups_z);
         if (split_k > 1) {
             // Try to evenly split KV into split_k chunks, but it needs to be a multiple
             // of "align", so recompute split_k based on that.
@@ -6237,9 +6283,9 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         }
     }
 
-    // Reserve space for split_k temporaries. For each split, we need to store the O matrix (D x ne1)
+    // Reserve space for split_k temporaries. For each split x batch, we need to store the O matrix (D x ne1)
-    // and the per-row m and L values (ne1 rows).
+    // and the per-row m and L values (ne1 rows). We store all the matrices first, followed by the rows.
-    const uint64_t split_k_size = split_k > 1 ? (D * ne1 * sizeof(float) + ne1 * sizeof(float) * 2) * split_k : 0;
+    const uint64_t split_k_size = split_k > 1 ? (HSV * ne1 * sizeof(float) + ne1 * sizeof(float) * 2) * split_k * ne3 : 0;
     if (split_k_size > ctx->device->max_memory_allocation_size) {
         GGML_ABORT("Requested preallocation size is too large");
     }
@@ -6331,11 +6377,10 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                               (uint32_t)neq2, (uint32_t)neq3,
                                               (uint32_t)nek2, (uint32_t)nek3,
                                               (uint32_t)nev2, (uint32_t)nev3,
-                                              nem1,
+                                              nem1, nem2,
                                               q_stride, (uint32_t)nbq2, (uint32_t)nbq3,
                                               k_stride, (uint32_t)nbk2, (uint32_t)nbk3,
                                               v_stride, (uint32_t)nbv2, (uint32_t)nbv3,
-                                              nbm1,
                                               scale, max_bias, logit_softcap,
                                               mask != nullptr, n_head_log2, m0, m1,
                                               gqa_ratio, split_kv, split_k };
@@ -6358,13 +6403,13 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                     pc, { workgroups_x * pipeline->wg_denoms[0], workgroups_y, workgroups_z });
 
         ggml_vk_sync_buffers(subctx);
-        const std::array<uint32_t, 3> pc2 = { D, (uint32_t)ne1, split_k };
+        const std::array<uint32_t, 4> pc2 = { HSV, (uint32_t)ne1, (uint32_t)ne3, split_k };
         ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_flash_attn_split_k_reduce,
                                     {
                                         vk_subbuffer{ctx->prealloc_split_k, 0, VK_WHOLE_SIZE},
                                         vk_subbuffer{d_D, d_buf_offset, VK_WHOLE_SIZE},
                                     },
-                                    pc2, { (uint32_t)ne1, 1, 1 });
+                                    pc2, { (uint32_t)ne1, 1, (uint32_t)ne3 });
     } else {
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
                                     {
@@ -6558,6 +6603,10 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
                 return ctx->device->pipeline_reglu[dst->type == GGML_TYPE_F16];
             case GGML_GLU_OP_SWIGLU:
                 return ctx->device->pipeline_swiglu[dst->type == GGML_TYPE_F16];
+            case GGML_GLU_OP_GEGLU_ERF:
+                return ctx->device->pipeline_geglu_erf[dst->type == GGML_TYPE_F16];
+            case GGML_GLU_OP_GEGLU_QUICK:
+                return ctx->device->pipeline_geglu_quick[dst->type == GGML_TYPE_F16];
             default:
                 break;
         }
@@ -7690,7 +7739,13 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
     const uint32_t nrows_x = (uint32_t)ggml_nrows(src0);
     const uint32_t nrows_y = (uint32_t)src0->ne[1];
 
-    const uint32_t n_head_kv   = nrows_x/nrows_y;
+    const uint32_t ne12 = src1 ? (uint32_t)(src1->ne[2]) : 0u;
+    const uint32_t ne13 = src1 ? (uint32_t)(src1->ne[3]) : 0u;
+    const uint32_t nb11 = src1 ? (uint32_t)(src1->nb[1] / src1->nb[0]) : 0u;
+    const uint32_t nb12 = src1 ? (uint32_t)(src1->nb[2] / src1->nb[0]) : 0u;
+    const uint32_t nb13 = src1 ? (uint32_t)(src1->nb[3] / src1->nb[0]) : 0u;
+
+    const uint32_t n_head_kv   = src0->ne[2];
     const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
 
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
@@ -7699,6 +7754,9 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
     ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SOFT_MAX, {
         ncols,
         src1 != nullptr ? nrows_y : (uint32_t)0,
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],
+        ne12, ne13,
+        nb11, nb12, nb13,
         scale, max_bias,
         m0, m1,
         n_head_log2,
@@ -8893,6 +8951,8 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_GLU_OP_GEGLU:
         case GGML_GLU_OP_REGLU:
         case GGML_GLU_OP_SWIGLU:
+        case GGML_GLU_OP_GEGLU_ERF:
+        case GGML_GLU_OP_GEGLU_QUICK:
             break;
         default:
             return false;
@@ -9140,6 +9200,8 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_GLU_OP_GEGLU:
         case GGML_GLU_OP_REGLU:
         case GGML_GLU_OP_SWIGLU:
+        case GGML_GLU_OP_GEGLU_ERF:
+        case GGML_GLU_OP_GEGLU_QUICK:
             ggml_vk_glu(ctx, compute_ctx, src0, src1, node, dryrun);
             break;
         default:
@@ -9358,6 +9420,8 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
         case GGML_GLU_OP_GEGLU:
         case GGML_GLU_OP_REGLU:
         case GGML_GLU_OP_SWIGLU:
+        case GGML_GLU_OP_GEGLU_ERF:
+        case GGML_GLU_OP_GEGLU_QUICK:
             buf = tensor->buffer;
             break;
         default:
@@ -10168,6 +10232,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 case GGML_GLU_OP_GEGLU:
                 case GGML_GLU_OP_REGLU:
                 case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
                     return ggml_is_contiguous(op->src[0]) &&
                            (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
                            (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) &&
@@ -10248,19 +10314,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
                 auto device = ggml_vk_get_device(ctx->device);
                 bool coopmat2 = device->coopmat2;
-                switch (op->src[0]->ne[0]) {
+                FaHeadSizes head_sizes = fa_get_head_sizes(op->src[1]->ne[0], op->src[2]->ne[0]);
-                case 64:
+                if (head_sizes == FA_HEAD_SIZE_UNSUPPORTED) {
-                case 80:
-                case 96:
-                case 112:
-                case 128:
-                case 256:
-                    break;
-                default:
-                    return false;
-                }
-                if (op->src[1]->ne[0] != op->src[2]->ne[0]) {
-                    // different head sizes of K and V are not supported yet
                     return false;
                 }
                 if (op->src[0]->type != GGML_TYPE_F32) {
@@ -10272,6 +10327,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 if (op->src[3] && op->src[3]->type != GGML_TYPE_F16) {
                     return false;
                 }
+                // TODO: support broadcast
+                // note: this was initially implemented in https://github.com/ggml-org/llama.cpp/pull/14449, but
+                //       the interface of ggml_flash_attn_ext() changed in https://github.com/ggml-org/llama.cpp/pull/14505
+                if (op->src[0]->ne[3] != 1 || (op->src[3] && op->src[3]->ne[2] != 1)) {
+                    return false;
+                }
                 // It's straightforward to support different K/V dequant, but would
                 // significantly increase the number of pipelines
                 if (op->src[1]->type != op->src[2]->type) {
@@ -10340,6 +10401,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                         return false;
                 }
             } break;
+        case GGML_OP_SET_ROWS:
+            {
+                // TODO: add support
+                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
+                return false;
+            } break;
         case GGML_OP_CONT:
         case GGML_OP_CPY:
         case GGML_OP_DUP:
@@ -10430,6 +10497,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_SCALE:
         case GGML_OP_PAD:
         case GGML_OP_DIAG_MASK_INF:
+            return true;
         case GGML_OP_SOFT_MAX:
         case GGML_OP_SOFT_MAX_BACK:
         case GGML_OP_ARGSORT:

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp

@@ -11,7 +11,8 @@
 #include "types.comp"
 #include "flash_attn_base.comp"
 
-const uint32_t D_per_thread = D / D_split;
+const uint32_t HSK_per_thread = HSK / D_split;
+const uint32_t HSV_per_thread = HSV / D_split;
 
 const uint32_t cols_per_iter = WorkGroupSize / D_split;
 const uint32_t cols_per_thread = Bc / cols_per_iter;
@@ -29,7 +30,7 @@ layout (binding = 3) readonly buffer M {float16_t data_m[];};
 // Rows index by Q's dimension 2, and the first N rows are valid.
 D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
 {
-    uint32_t offset = (iq2 + r) * D + c;
+    uint32_t offset = (iq2 + r) * HSV + c;
     data_o[o_offset + offset] = D_TYPE(elem);
     return elem;
 }
@@ -38,7 +39,7 @@ shared FLOAT_TYPE tmpsh[WorkGroupSize];
 shared vec4 tmpshv4[WorkGroupSize];
 
 shared float masksh[Bc][Br];
-shared vec4 Qf[Br][D / 4];
+shared vec4 Qf[Br][HSK / 4];
 
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
@@ -53,18 +54,18 @@ void main() {
 
     uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
 
-    [[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) {
+    [[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
-        uint32_t d = (idx + tid) % (D / 4);
+        uint32_t d = (idx + tid) % (HSK / 4);
-        uint32_t r = (idx + tid) / (D / 4);
+        uint32_t r = (idx + tid) / (HSK / 4);
-        if (r < Br && d < D / 4 &&
+        if (r < Br && d < HSK / 4 &&
             i * Br + r < N) {
             Qf[r][d] = vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d]) * p.scale;
         }
     }
     barrier();
 
-    vec4 Of[Br][D_per_thread / 4];
+    vec4 Of[Br][HSV_per_thread / 4];
-    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             Of[r][d] = vec4(0.0);
         }
@@ -99,6 +100,10 @@ void main() {
     uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2;
     uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
 #endif
+    uint32_t m_offset = 0;
+    if (p.nem2 != 1) {
+        m_offset = (iq3 % p.nem2) * p.nem1 * KV;
+    }
 
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
@@ -112,7 +117,7 @@ void main() {
 
 
         [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+            [[unroll]] for (uint32_t d = 0; d < HSK_per_thread / 4; ++d) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
                 uint ib = coord / BLOCK_SIZE;
@@ -150,7 +155,7 @@ void main() {
                 uint32_t c = (idx + tid) % Bc;
                 uint32_t r = (idx + tid) / Bc;
                 if (idx + tid < Bc * Br) {
-                    masksh[c][r] = float(data_m[(i * Br + r) * m_stride + (j * Bc + c)]);
+                    masksh[c][r] = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
                 }
             }
             barrier();
@@ -191,14 +196,14 @@ void main() {
             Lf[r] = eMf[r]*Lf[r] + rowsumf[r];
         }
 
-        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
             [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
                 Of[r][d] = eMf[r] * Of[r][d];
             }
         }
 
         [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+            [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
                 uint ib = coord / BLOCK_SIZE;
@@ -255,7 +260,7 @@ void main() {
         Lf[r] = tmpsh[d_tid];
         barrier();
 
-        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 
             Of[r][d] = eMf * Of[r][d];
             tmpshv4[tid] = Of[r][d];
@@ -277,11 +282,11 @@ void main() {
     // If there is split_k, then the split_k resolve shader does the final
     // division by L. Store the intermediate O value and per-row m and L values.
     if (p.k_num > 1) {
-        uint32_t o_offset = D * p.ne1 * split_k_index;
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
 
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
                     }
@@ -289,7 +294,7 @@ void main() {
             }
         }
 
-        o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
+        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
                 perElemOpStoreCol0(r, 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
@@ -305,18 +310,18 @@ void main() {
         Lfrcp[r] = 1.0 / Lf[r];
     }
 
-    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             Of[r][d] *= Lfrcp[r];
         }
     }
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1;
+    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
 
     if (p.gqa_ratio > 1) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
                     }
@@ -326,9 +331,9 @@ void main() {
     } else {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (i * Br + r < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        data_o[o_offset + iq2 * D + (i * Br + r) * p.ne1 * D + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
+                        data_o[o_offset + iq2 * HSV + (i * Br + r) * p.ne1 * HSV + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
                     }
                 }
             }

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp

@@ -4,10 +4,10 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 layout (constant_id = 0) const uint32_t WorkGroupSize = 128;
 layout (constant_id = 1) const uint32_t Br = 1;
 layout (constant_id = 2) const uint32_t Bc = 32;
-layout (constant_id = 3) const uint32_t D = 32;
+layout (constant_id = 3) const uint32_t HSK = 32;
-layout (constant_id = 4) const uint32_t Clamp = 0;
+layout (constant_id = 4) const uint32_t HSV = 32;
-layout (constant_id = 5) const uint32_t D_split = 16;
+layout (constant_id = 5) const uint32_t Clamp = 0;
-
+layout (constant_id = 6) const uint32_t D_split = 16;
 
 layout (push_constant) uniform parameter {
     uint32_t N;
@@ -24,6 +24,7 @@ layout (push_constant) uniform parameter {
     uint32_t nev2;
     uint32_t nev3;
     uint32_t nem1;
+    uint32_t nem2;
 
     uint32_t nb01;
     uint32_t nb02;
@@ -34,7 +35,6 @@ layout (push_constant) uniform parameter {
     uint32_t nb21;
     uint32_t nb22;
     uint32_t nb23;
-    uint32_t nb31;
 
     float scale;
     float max_bias;

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp

@@ -13,7 +13,9 @@
 #include "types.comp"
 #include "flash_attn_base.comp"
 
-const uint32_t D_per_thread = D / D_split;
+const uint32_t HSK_per_thread = HSK / D_split;
+const uint32_t HSV_per_thread = HSV / D_split;
+
 const uint32_t row_split = 4;
 const uint32_t rows_per_thread = Br / row_split;
 const uint32_t cols_per_iter = gl_WorkGroupSize.x / D_split / row_split;
@@ -32,7 +34,7 @@ layout (binding = 3) readonly buffer M {float16_t data_m[];};
 // Rows index by Q's dimension 2, and the first N rows are valid.
 D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
 {
-    uint32_t offset = (iq2 + r) * D + c;
+    uint32_t offset = (iq2 + r) * HSV + c;
     data_o[o_offset + offset] = D_TYPE(elem);
     return elem;
 }
@@ -44,14 +46,14 @@ const uint32_t MatBc = 16;
 shared FLOAT_TYPE tmpsh[gl_WorkGroupSize.x];
 shared ACC_TYPEV4 tmpshv4[gl_WorkGroupSize.x];
 
-const uint32_t qstride = D / 4 + 2; // in units of f16vec4
+const uint32_t qstride = HSK / 4 + 2; // in units of f16vec4
 shared f16vec4 Qf[Br * qstride];
 
-// Avoid padding for D==256 to make it fit in 48KB shmem.
+// Avoid padding for hsk==256 to make it fit in 48KB shmem.
-const uint32_t sfshstride = (D <= 128) ? (Br + 8) : Br;
+const uint32_t sfshstride = (HSK <= 128) ? (Br + 8) : Br;
 shared ACC_TYPE sfsh[Bc * sfshstride];
 
-const uint32_t kshstride = D / 4 + 2; // in units of f16vec4
+const uint32_t kshstride = HSK / 4 + 2; // in units of f16vec4
 shared f16vec4 ksh[Bc * kshstride];
 
 shared float slope[Br];
@@ -74,18 +76,18 @@ void main() {
 
     uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
 
-    [[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) {
+    [[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
-        uint32_t d = (idx + tid) % (D / 4);
+        uint32_t d = (idx + tid) % (HSK / 4);
-        uint32_t r = (idx + tid) / (D / 4);
+        uint32_t r = (idx + tid) / (HSK / 4);
-        if (r < Br && d < D / 4 &&
+        if (r < Br && d < HSK / 4 &&
             i * Br + r < N) {
             Qf[r * qstride + d] = f16vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d] * p.scale);
         }
     }
     barrier();
 
-    ACC_TYPEV4 Of[rows_per_thread][D_per_thread / 4];
+    ACC_TYPEV4 Of[rows_per_thread][HSV_per_thread / 4];
-    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             Of[r][d] = ACC_TYPEV4(0.0);
         }
@@ -123,14 +125,18 @@ void main() {
     uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2;
     uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
 #endif
+    uint32_t m_offset = 0;
+    if (p.nem2 != 1) {
+        m_offset = (iq3 % p.nem2) * p.nem1 * KV;
+    }
 
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
 
-        [[unroll]] for (uint32_t idx = 0; idx < Bc * D / 4; idx += gl_WorkGroupSize.x) {
+        [[unroll]] for (uint32_t idx = 0; idx < Bc * HSK / 4; idx += gl_WorkGroupSize.x) {
-            uint32_t d = (idx + tid) % (D / 4);
+            uint32_t d = (idx + tid) % (HSK / 4);
-            uint32_t c = (idx + tid) / (D / 4);
+            uint32_t c = (idx + tid) / (HSK / 4);
-            if (c < Bc && d < D / 4) {
+            if (c < Bc && d < HSK / 4) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d;
                 uint ib = coord / BLOCK_SIZE;
@@ -145,14 +151,14 @@ void main() {
         }
         barrier();
 
-        // K * Q^T -> S^T: Bc x D * D x Br -> Bc x Br
+        // K * Q^T -> S^T: Bc x HSK * HSK x Br -> Bc x Br
-        // Bc split across workgroup (four subgroups), loop over D in chunks of 16: 16 x 16 * 16 x 16 -> 16 x 16
+        // Bc split across workgroup (four subgroups), loop over HSK in chunks of 16: 16 x 16 * 16 x 16 -> 16 x 16
         // This is written transposed in order to allow for N being 8 if implementations need it
         coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator> SfMat = coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator>(0);
         coopmat<float16_t, gl_ScopeSubgroup, MatBc, 16, gl_MatrixUseA> KMat;
         coopmat<float16_t, gl_ScopeSubgroup, 16, MatBr, gl_MatrixUseB> QMat;
 
-        for (uint32_t d = 0; d < D / 16; ++d) {
+        for (uint32_t d = 0; d < HSK / 16; ++d) {
             coopMatLoad(QMat, Qf, d * 16 / 4, qstride, gl_CooperativeMatrixLayoutColumnMajor);
 
             uint coord = (gl_SubgroupID * MatBc) * kshstride + d * 16 / 4;
@@ -181,7 +187,7 @@ void main() {
                 uint32_t c = (idx + tid) % Bc;
                 uint32_t r = (idx + tid) / Bc;
                 if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) {
-                    sfsh[c * sfshstride + r] += ACC_TYPE(slope[r] * float(data_m[(i * Br + r) * m_stride + (j * Bc + c)]));
+                    sfsh[c * sfshstride + r] += ACC_TYPE(slope[r] * float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]));
                 }
             }
             barrier();
@@ -202,7 +208,7 @@ void main() {
             eMf[r] = exp(Moldf - Mf[r]);
         }
 
-        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
             [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
                 Of[r][d] = float16_t(eMf[r]) * Of[r][d];
             }
@@ -217,7 +223,7 @@ void main() {
                 Pf[r] = exp(sfsh[tile_row(r) + (c * cols_per_iter + col_tid) * sfshstride] - Mf[r]);
                 Lf[r] += Pf[r];
             }
-            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+            [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
                 uint ib = coord / BLOCK_SIZE;
@@ -280,7 +286,7 @@ void main() {
     }
 
     [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 
             Of[r][d] = float16_t(eMf[r]) * Of[r][d];
             tmpshv4[tid] = Of[r][d];
@@ -300,11 +306,11 @@ void main() {
     // If there is split_k, then the split_k resolve shader does the final
     // division by L. Store the intermediate O value and per-row m and L values.
     if (p.k_num > 1) {
-        uint32_t o_offset = D * p.ne1 * split_k_index;
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
 
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (tile_row(r) < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(tile_row(r), 4*(d * D_split + d_tid) + comp, float(Of[r][d][comp]), o_offset, iq2, N);
                     }
@@ -312,7 +318,7 @@ void main() {
             }
         }
 
-        o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
+        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (tile_row(r) < N) {
                 perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
@@ -328,18 +334,18 @@ void main() {
         Lfrcp[r] = 1.0 / Lf[r];
     }
 
-    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             Of[r][d] *= float16_t(Lfrcp[r]);
         }
     }
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1;
+    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
 
     if (p.gqa_ratio > 1) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (tile_row(r) < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(tile_row(r), 4*(d * D_split + d_tid) + comp, float(Of[r][d][comp]), o_offset, iq2, N);
                     }
@@ -349,9 +355,9 @@ void main() {
     } else {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (i * Br + tile_row(r) < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        data_o[o_offset + iq2 * D + (i * Br + tile_row(r)) * p.ne1 * D + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
+                        data_o[o_offset + iq2 * HSV + (i * Br + tile_row(r)) * p.ne1 * HSV + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
                     }
                 }
             }

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp

@@ -61,8 +61,8 @@ ACC_TYPE Max(const in uint32_t row, const in uint32_t col, const in ACC_TYPE ele
 // Rows index by Q's dimension 2, and the first N rows are valid.
 D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
 {
-    if (r < N && c < D) {
+    if (r < N && c < HSV) {
-        uint32_t offset = (iq2 + r) * D + c;
+        uint32_t offset = (iq2 + r) * HSV + c;
         data_o[o_offset + offset] = D_TYPE(elem);
     }
     return elem;
@@ -86,9 +86,9 @@ void main() {
     tensorLayoutV = setTensorLayoutBlockSizeNV(tensorLayoutV, 1, BLOCK_SIZE);
 #endif
 
-    tensorLayoutQ = setTensorLayoutDimensionNV(tensorLayoutQ, N, D);
+    tensorLayoutQ = setTensorLayoutDimensionNV(tensorLayoutQ, N, HSK);
-    tensorLayoutK = setTensorLayoutDimensionNV(tensorLayoutK, KV, D);
+    tensorLayoutK = setTensorLayoutDimensionNV(tensorLayoutK, KV, HSK);
-    tensorLayoutV = setTensorLayoutDimensionNV(tensorLayoutV, KV, D);
+    tensorLayoutV = setTensorLayoutDimensionNV(tensorLayoutV, KV, HSV);
 
     // hint to the compiler that strides are aligned for the aligned variant of the shader
     if (Clamp != gl_CooperativeMatrixClampModeConstantNV)
@@ -104,16 +104,16 @@ void main() {
     tensorLayoutK = setTensorLayoutStrideNV(tensorLayoutK, k_stride, 1);
     tensorLayoutV = setTensorLayoutStrideNV(tensorLayoutV, v_stride, 1);
 
-    coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> Q;
+    coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseAccumulator> Q;
-    coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA> Qf16;
+    coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseA> Qf16;
 
     uint32_t q_offset = iq2*p.nb02+iq3*p.nb03;
-    coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, D));
+    coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, HSK));
 
-    Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA>(Q);
+    Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseA>(Q);
     Qf16 *= float16_t(p.scale);
 
-    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(0);
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(0);
 
     coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> L, M;
 
@@ -130,15 +130,20 @@ void main() {
         coopMatPerElementNV(slopeMat, slopeMat, perElemOpComputeSlope, iq2);
     }
 
+    uint32_t m_offset = 0;
+    if (p.nem2 != 1) {
+        m_offset = (iq3 % p.nem2) * p.nem1 * KV * 2 /*sizeof(float16_t)*/;
+    }
+
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
 
         coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> S = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
 
-        coopmat<float16_t, gl_ScopeWorkgroup, D, Bc, gl_MatrixUseB> K_T;
+        coopmat<float16_t, gl_ScopeWorkgroup, HSK, Bc, gl_MatrixUseB> K_T;
 
         uint32_t k_offset = ik2*p.nb12 + ik3*p.nb13;
-        coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, D), tensorViewTranspose DECODEFUNC);
+        coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, HSK), tensorViewTranspose DECODEFUNC);
         S = coopMatMulAdd(Qf16, K_T, S);
 
         if (p.logit_softcap != 0.0f) {
@@ -155,7 +160,7 @@ void main() {
 
             coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv;
 
-            coopMatLoadTensorNV(mv, data_m, 0, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
+            coopMatLoadTensorNV(mv, data_m, m_offset, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
 
             S += slopeMat*coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(mv);
         }
@@ -203,42 +208,42 @@ void main() {
         rowsum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0.0);
         rowsum = coopMatMulAdd(P_A, One, rowsum);
 
-        coopmat<float16_t, gl_ScopeWorkgroup, Bc, D, gl_MatrixUseB> V;
+        coopmat<float16_t, gl_ScopeWorkgroup, Bc, HSV, gl_MatrixUseB> V;
         uint32_t v_offset = iv2*p.nb22 + iv3*p.nb23;
-        coopMatLoadTensorNV(V,  data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, D) DECODEFUNC);
+        coopMatLoadTensorNV(V,  data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, HSV) DECODEFUNC);
 
         L = eM*L + rowsum;
 
         // This is the "diagonal" matrix in the paper, but since we do componentwise
         // multiply rather than matrix multiply it has the diagonal element smeared
         // across the row
-        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> eMdiag;
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> eMdiag;
 
         // resize eM by using smear/reduce
         coopMatReduceNV(eMdiag, eM, gl_CooperativeMatrixReduceRowNV, smearReduce);
 
         // multiply with fp16 accumulation, then add to O.
-        coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(0);
+        coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(0);
         PV = coopMatMulAdd(P_A, V, PV);
 
-        O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(PV);
+        O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(PV);
     }
 
     // If there is split_k, then the split_k resolve shader does the final
     // division by L. Store the intermediate O value and per-row m and L values.
     if (p.k_num > 1) {
-        coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O);
+        coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(O);
 
-        uint32_t o_offset = D * p.ne1 * split_k_index;
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
         coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
 
-        o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
+        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
         coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N);
         coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N);
         return;
     }
 
-    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> Ldiag;
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> Ldiag;
 
     // resize L by using smear/reduce
     coopMatReduceNV(Ldiag, L, gl_CooperativeMatrixReduceRowNV, smearReduce);
@@ -250,18 +255,18 @@ void main() {
 
     O = Ldiag*O;
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1;
+    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
 
-    coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O);
+    coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(O);
     if (p.gqa_ratio > 1) {
         coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
     } else {
         tensorLayoutNV<3, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutD = createTensorLayoutNV(3, gl_CooperativeMatrixClampModeConstantNV);
-        tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.ne2, p.ne1, D);
+        tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.ne2, p.ne1, HSV);
 
         // permute dimensions
         tensorViewNV<3, false, 1, 0, 2> tensorViewPermute = createTensorViewNV(3, false, 1, 0, 2);
 
-        coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, D), tensorViewPermute);
+        coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, HSV), tensorViewPermute);
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp

@@ -12,6 +12,7 @@ layout (binding = 1) writeonly buffer D {float data_d[];};
 layout (push_constant) uniform parameter {
     uint D;
     uint N;
+    uint ne3;
     uint k_num;
 } p;
 
@@ -19,13 +20,14 @@ void main() {
     // Each workgroup handles a row
     const uint n = gl_WorkGroupID.x;
     const uint tid = gl_LocalInvocationID.x;
+    const uint iq3 = gl_WorkGroupID.z;
 
     uint D = p.D;
     uint N = p.N;
     uint k_num = p.k_num;
 
-    uint l_offset = D * N * k_num + n;
+    uint l_offset = D * N * p.ne3 * k_num + N * iq3 * k_num * 2 + n;
-    uint m_offset = D * N * k_num + N + n;
+    uint m_offset = D * N * p.ne3 * k_num + N * iq3 * k_num * 2 + N + n;
     uint lm_stride = N * 2;
 
     // Compute the max m value for the row
@@ -49,11 +51,11 @@ void main() {
     for (uint d = tid; d < D; d += BLOCK_SIZE) {
         float O = 0.0;
         [[unroll]] for (uint k = 0; k < k_num; ++k) {
-            uint o_offset = D * N * k + D * n + d;
+            uint o_offset = D * N * (k + iq3 * k_num) + D * n + d;
             float m = data_a[m_offset + k * lm_stride];
             O += exp(m - m_max) * data_a[o_offset];
         }
         O *= L;
-        data_d[D * n + d] = O;
+        data_d[iq3 * D * N + D * n + d] = O;
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/geglu_erf.comp

@@ -0,0 +1,27 @@
+#version 450
+
+#include "glu_head.comp"
+
+// based on Abramowitz and Stegun formula 7.1.26 or similar Hastings' approximation
+// ref: https://www.johndcook.com/blog/python_erf/
+const float p_erf  = 0.3275911f;
+const float a1_erf = 0.254829592f;
+const float a2_erf = -0.284496736f;
+const float a3_erf = 1.421413741f;
+const float a4_erf = -1.453152027f;
+const float a5_erf = 1.061405429f;
+
+const float SQRT_2_INV = 0.70710678118654752440084436210484f;
+
+float op(float a, float b) {
+    const float a_div_sqr2 = a * SQRT_2_INV;
+    const float sign_x = sign(a_div_sqr2);
+    const float x = abs(a_div_sqr2);
+    const float t = 1.0f / (1.0f + p_erf * x);
+    const float y = 1.0f - (((((a5_erf * t + a4_erf) * t) + a3_erf) * t + a2_erf) * t + a1_erf) * t * exp(-x * x);
+    const float erf_approx = sign_x * y;
+
+    return 0.5f * a * (1.0f + erf_approx) * b;
+}
+
+#include "glu_main.comp"

ggml/src/ggml-vulkan/vulkan-shaders/geglu_quick.comp

@@ -0,0 +1,11 @@
+#version 450
+
+#include "glu_head.comp"
+
+const float GELU_QUICK_COEF = -1.702f;
+
+float op(float a, float b) {
+    return a * (1.0f / (1.0f + exp(GELU_QUICK_COEF * a))) * b;
+}
+
+#include "glu_main.comp"

ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp

@@ -6,6 +6,14 @@ layout (push_constant) uniform parameter
 {
     uint KX;
     uint KY;
+    uint ne00;
+    uint ne01;
+    uint ne02;
+    uint ne12;
+    uint ne13;
+    uint nb11;
+    uint nb12;
+    uint nb13;
     float scale;
     float max_bias;
     float m0;
@@ -31,7 +39,15 @@ shared FLOAT_TYPE vals[BLOCK_SIZE];
 void soft_max(uint num_iters) {
     const uint tid = gl_LocalInvocationID.x;
     const uint rowx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
-    const uint rowy = (p.KY > 0) ? (rowx % p.KY) : 0;
+
+    const uint32_t i03 = rowx / (p.ne01 * p.ne02);
+    const uint32_t i02 = (rowx - i03 * p.ne01 * p.ne02) / p.ne01;
+    const uint32_t i01 = rowx % p.ne01;
+
+    uint rowy_start = 0;
+    if (p.KY > 0) {
+        rowy_start = i01 * p.nb11 + (i02 % p.ne12) * p.nb12 + (i03 % p.ne13) * p.nb13;
+    }
 
     if (rowx >= p.nrows_x) {
         return;
@@ -41,7 +57,7 @@ void soft_max(uint num_iters) {
 
     // ALiBi
     if (p.max_bias > 0.0f) {
-        const uint h = rowx/p.KY; // head index
+        const uint h = (rowx / p.ne01) % p.ne02; // head index
 
         const float base = h < p.n_head_log2 ? p.m0 : p.m1;
         const uint   exp  = h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1;
@@ -67,7 +83,7 @@ void soft_max(uint num_iters) {
 
         FLOAT_TYPE b = FLOAT_TYPE(0);
         if (p.KY > 0 && col < p.KX) {
-            b = data_b[rowy * p.KX + col];
+            b = data_b[rowy_start + col];
         }
 
         FLOAT_TYPE v = a * p.scale + slope * b;
@@ -111,7 +127,7 @@ void soft_max(uint num_iters) {
         if (idx < DATA_CACHE_SIZE) {
             val = exp(data_cache[idx] - max_val);
         } else {
-            val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) - max_val);
+            val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy_start + col]) : FLOAT_TYPE(0.0f)) - max_val);
         }
         sum += val;
         if (idx < DATA_CACHE_SIZE) {

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -607,6 +607,10 @@ void process_shaders() {
     string_to_spv("reglu_f32",      "reglu.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
     string_to_spv("swiglu_f16",     "swiglu.comp",      {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("swiglu_f32",     "swiglu.comp",      {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    string_to_spv("geglu_erf_f16",  "geglu_erf.comp",   {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("geglu_erf_f32",  "geglu_erf.comp",   {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    string_to_spv("geglu_quick_f16","geglu_quick.comp", {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("geglu_quick_f32","geglu_quick.comp", {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
 
     string_to_spv("leaky_relu_f32", "leaky_relu.comp",  {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("silu_back_f32",  "silu_back.comp",   {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});

ggml/src/ggml.c

@@ -21,6 +21,9 @@
 #include <alloca.h>
 #endif
 
+#define GGML_VERSION "0.0.1"
+#define GGML_COMMIT "KCPP"
+
 #include <assert.h>
 #include <errno.h>
 #include <time.h>
@@ -474,6 +477,14 @@ bool ggml_guid_matches(ggml_guid_t guid_a, ggml_guid_t guid_b) {
     return memcmp(guid_a, guid_b, sizeof(ggml_guid)) == 0;
 }
 
+const char * ggml_version(void) {
+    return GGML_VERSION;
+}
+
+const char * ggml_commit(void) {
+    return GGML_COMMIT;
+}
+
 //
 // timing
 //
@@ -1145,9 +1156,11 @@ static const char * GGML_GLU_OP_NAME[GGML_GLU_OP_COUNT] = {
     "REGLU",
     "GEGLU",
     "SWIGLU",
+    "GEGLU_ERF",
+    "GEGLU_QUICK",
 };
 
-static_assert(GGML_GLU_OP_COUNT == 3, "GGML_GLU_OP_COUNT != 3");
+static_assert(GGML_GLU_OP_COUNT == 5, "GGML_GLU_OP_COUNT != 5");
 
 
 static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN");
@@ -2773,6 +2786,48 @@ struct ggml_tensor * ggml_swiglu_split(
     return ggml_glu_impl(ctx, a, b, GGML_GLU_OP_SWIGLU, false);
 }
 
+// ggml_geglu_erf
+
+struct ggml_tensor * ggml_geglu_erf(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_GEGLU_ERF, false);
+}
+
+struct ggml_tensor * ggml_geglu_erf_swapped(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_GEGLU_ERF, true);
+}
+
+struct ggml_tensor * ggml_geglu_erf_split(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
+    return ggml_glu_impl(ctx, a, b, GGML_GLU_OP_GEGLU_ERF, false);
+}
+
+// ggml_geglu_quick
+
+struct ggml_tensor * ggml_geglu_quick(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_GEGLU_QUICK, false);
+}
+
+struct ggml_tensor * ggml_geglu_quick_swapped(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_GEGLU_QUICK, true);
+}
+
+struct ggml_tensor * ggml_geglu_quick_split(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
+    return ggml_glu_impl(ctx, a, b, GGML_GLU_OP_GEGLU_QUICK, false);
+}
+
 // ggml_norm
 
 static struct ggml_tensor * ggml_norm_impl(
@@ -3679,9 +3734,10 @@ static struct ggml_tensor * ggml_soft_max_impl(
     if (mask) {
         GGML_ASSERT(mask->type == GGML_TYPE_F16 || mask->type == GGML_TYPE_F32);
         GGML_ASSERT(ggml_is_contiguous(mask));
-        GGML_ASSERT(ggml_is_matrix(mask));
         GGML_ASSERT(mask->ne[0] == a->ne[0]);
         GGML_ASSERT(mask->ne[1] >= a->ne[1]);
+        GGML_ASSERT(a->ne[2]%mask->ne[2] == 0);
+        GGML_ASSERT(a->ne[3]%mask->ne[3] == 0);
     }
 
     if (max_bias > 0.0f) {
@@ -4702,13 +4758,17 @@ struct ggml_tensor * ggml_flash_attn_ext(
     GGML_ASSERT(ggml_can_mul_mat(k, q));
     // TODO: check if vT can be multiplied by (k*qT)
 
+    GGML_ASSERT(q->ne[3] == k->ne[3]);
+    GGML_ASSERT(q->ne[3] == v->ne[3]);
+
     if (mask) {
         GGML_ASSERT(ggml_is_contiguous(mask));
-        GGML_ASSERT(mask->ne[2] == 1);
-        GGML_ASSERT(mask->ne[3] == 1);
         GGML_ASSERT(mask->ne[1] >= GGML_PAD(q->ne[1], GGML_KQ_MASK_PAD) &&
                 "the Flash-Attention kernel requires the mask to be padded to GGML_KQ_MASK_PAD and at least n_queries big");
         //GGML_ASSERT(ggml_can_repeat_rows(mask, qk));
+
+        GGML_ASSERT(q->ne[2] % mask->ne[2] == 0);
+        GGML_ASSERT(q->ne[3] % mask->ne[3] == 0);
     }
 
     if (max_bias > 0.0f) {
@@ -4836,7 +4896,6 @@ struct ggml_tensor * ggml_ssm_conv(
     const int64_t n_s     = sx->ne[2];
 
     // TODO: maybe support other strides than 1?
-    // FIXME: this is always true?
     GGML_ASSERT(sx->ne[0] == d_conv - 1 + n_t);
     GGML_ASSERT(sx->ne[1] == d_inner);
     GGML_ASSERT(n_t >= 0);
@@ -4859,36 +4918,49 @@ struct ggml_tensor * ggml_ssm_scan(
         struct ggml_tensor  * dt,
         struct ggml_tensor  * A,
         struct ggml_tensor  * B,
-        struct ggml_tensor  * C) {
+        struct ggml_tensor  * C,
+        struct ggml_tensor  * ids) {
     GGML_ASSERT(ggml_is_contiguous(s));
-    GGML_ASSERT(ggml_is_contiguous(x));
     GGML_ASSERT(ggml_is_contiguous(dt));
     GGML_ASSERT(ggml_is_contiguous(A));
-    GGML_ASSERT(ggml_is_matrix(A));
+    GGML_ASSERT(x->nb[0] == ggml_type_size(x->type));
-    GGML_ASSERT(ggml_is_3d(B));
-    GGML_ASSERT(ggml_is_3d(s));
     GGML_ASSERT(B->nb[0] == ggml_type_size(B->type));
     GGML_ASSERT(C->nb[0] == ggml_type_size(C->type));
-    GGML_ASSERT(ggml_are_same_shape(x, dt));
+    GGML_ASSERT(x->nb[1] == x->ne[0]*x->nb[0]);
+    GGML_ASSERT(B->nb[1] == B->ne[0]*B->nb[0]);
+    GGML_ASSERT(C->nb[1] == C->ne[0]*C->nb[0]);
     GGML_ASSERT(ggml_are_same_shape(B, C));
+    GGML_ASSERT(ids->type == GGML_TYPE_I32);
 
     {
         const int64_t d_state      = s->ne[0];
-        const int64_t d_inner      = s->ne[1];
+        const int64_t head_dim     = x->ne[0];
-        const int64_t n_seq_tokens = x->ne[1];
+        const int64_t n_head       = x->ne[1];
-        const int64_t n_seqs       = x->ne[2];
+        const int64_t n_seq_tokens = x->ne[2];
+        const int64_t n_seqs       = x->ne[3];
 
-        GGML_ASSERT(s->ne[2] == n_seqs);
+        GGML_ASSERT(dt->ne[0] == n_head);
-        GGML_ASSERT(x->ne[0] == d_inner);
+        GGML_ASSERT(dt->ne[1] == n_seq_tokens);
-        GGML_ASSERT(A->ne[0] == d_state);
+        GGML_ASSERT(dt->ne[2] == n_seqs);
-        GGML_ASSERT(A->ne[1] == d_inner);
+        GGML_ASSERT(ggml_is_3d(dt));
+        GGML_ASSERT(s->ne[1] == head_dim);
+        GGML_ASSERT(s->ne[2] == n_head);
         GGML_ASSERT(B->ne[0] == d_state);
-        GGML_ASSERT(B->ne[1] == n_seq_tokens);
+        GGML_ASSERT(B->ne[2] == n_seq_tokens);
-        GGML_ASSERT(B->ne[2] == n_seqs);
+        GGML_ASSERT(B->ne[3] == n_seqs);
+        GGML_ASSERT(ids->ne[0] == n_seqs);
+        GGML_ASSERT(ggml_is_vector(ids));
+        GGML_ASSERT(A->ne[1] == n_head);
+        GGML_ASSERT(ggml_is_matrix(A));
+
+        if (A->ne[0] != 1) {
+            // Mamba-1 has more granular decay factors
+            GGML_ASSERT(A->ne[0] == d_state);
+        }
     }
 
     // concatenated y + ssm_states
-    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, ggml_nelements(x) + ggml_nelements(s));
+    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, ggml_nelements(x) + s->ne[0]*s->ne[1]*s->ne[2]*ids->ne[0]);
 
     result->op   = GGML_OP_SSM_SCAN;
     result->src[0] = s;
@@ -4897,6 +4969,7 @@ struct ggml_tensor * ggml_ssm_scan(
     result->src[3] = A;
     result->src[4] = B;
     result->src[5] = C;
+    result->src[6] = ids;
 
     return result;
 }
@@ -6037,13 +6110,28 @@ static void ggml_compute_backward(
             }
             GGML_ASSERT(!src1_needs_grads && "backward pass for labels not implemented");
         } break;
+        case GGML_OP_GLU: {
+            switch (ggml_get_glu_op(tensor)) {
+                case GGML_GLU_OP_SWIGLU: {
+                    if (src0_needs_grads) {
+                        GGML_ASSERT(src1 && "backward pass only implemented for split swiglu");
+                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_silu_back(ctx, ggml_mul(ctx, grad, src1), src0));
+                    }
+                    if (src1_needs_grads) {
+                        ggml_add_or_set(ctx, cgraph, isrc1, ggml_mul(ctx, ggml_silu(ctx, src0), grad));
+                    }
+                } break;
+                default: {
+                    GGML_ABORT("unsupported glu op for backward pass: %s", ggml_glu_op_name(ggml_get_glu_op(tensor)));
+                } //break;
+            }
+        } break;
         case GGML_OP_NONE: {
             // noop
         } break;
         case GGML_OP_COUNT:
         default: {
-            fprintf(stderr, "%s: unsupported ggml op for backward pass: %s\n", __func__, ggml_op_name(tensor->op));
+            GGML_ABORT("%s: unsupported ggml op for backward pass: %s\n", __func__, ggml_op_name(tensor->op));
-            GGML_ABORT("fatal error");
         } //break;
     }
 

gguf-py/gguf/constants.py

@@ -170,6 +170,7 @@ class Keys:
         INNER_SIZE     = "{arch}.ssm.inner_size"
         STATE_SIZE     = "{arch}.ssm.state_size"
         TIME_STEP_RANK = "{arch}.ssm.time_step_rank"
+        GROUP_COUNT    = "{arch}.ssm.group_count"
         DT_B_C_RMS     = "{arch}.ssm.dt_b_c_rms"
 
     class WKV:
@@ -327,6 +328,7 @@ class MODEL_ARCH(IntEnum):
     RWKV7            = auto()
     ARWKV7           = auto()
     MAMBA            = auto()
+    MAMBA2           = auto()
     XVERSE           = auto()
     COMMAND_R        = auto()
     COHERE2          = auto()
@@ -429,6 +431,7 @@ class MODEL_TENSOR(IntEnum):
     SSM_DT               = auto()
     SSM_A                = auto()
     SSM_D                = auto()
+    SSM_NORM             = auto()
     SSM_OUT              = auto()
     TIME_MIX_W0          = auto()
     TIME_MIX_W1          = auto()
@@ -628,6 +631,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.RWKV7:            "rwkv7",
     MODEL_ARCH.ARWKV7:           "arwkv7",
     MODEL_ARCH.MAMBA:            "mamba",
+    MODEL_ARCH.MAMBA2:           "mamba2",
     MODEL_ARCH.XVERSE:           "xverse",
     MODEL_ARCH.COMMAND_R:        "command-r",
     MODEL_ARCH.COHERE2:          "cohere2",
@@ -730,6 +734,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.SSM_DT:                    "blk.{bid}.ssm_dt",
     MODEL_TENSOR.SSM_A:                     "blk.{bid}.ssm_a",
     MODEL_TENSOR.SSM_D:                     "blk.{bid}.ssm_d",
+    MODEL_TENSOR.SSM_NORM:                  "blk.{bid}.ssm_norm",
     MODEL_TENSOR.SSM_OUT:                   "blk.{bid}.ssm_out",
     MODEL_TENSOR.TIME_MIX_W0:               "blk.{bid}.time_mix_w0",
     MODEL_TENSOR.TIME_MIX_W1:               "blk.{bid}.time_mix_w1",
@@ -1714,6 +1719,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.SSM_D,
         MODEL_TENSOR.SSM_OUT,
     ],
+    MODEL_ARCH.MAMBA2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.SSM_IN,
+        MODEL_TENSOR.SSM_CONV1D,
+        MODEL_TENSOR.SSM_DT,
+        MODEL_TENSOR.SSM_A,
+        MODEL_TENSOR.SSM_D,
+        MODEL_TENSOR.SSM_NORM,
+        MODEL_TENSOR.SSM_OUT,
+    ],
     MODEL_ARCH.XVERSE: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -2497,6 +2515,7 @@ KEY_SSM_CONV_KERNEL    = Keys.SSM.CONV_KERNEL
 KEY_SSM_INNER_SIZE     = Keys.SSM.INNER_SIZE
 KEY_SSM_STATE_SIZE     = Keys.SSM.STATE_SIZE
 KEY_SSM_TIME_STEP_RANK = Keys.SSM.TIME_STEP_RANK
+KEY_SSM_GROUP_COUNT    = Keys.SSM.GROUP_COUNT
 KEY_SSM_DT_B_C_RMS     = Keys.SSM.DT_B_C_RMS
 
 # tokenization

gguf-py/gguf/gguf_writer.py

@@ -714,8 +714,8 @@ class GGUFWriter:
     def add_clamp_kqv(self, value: float) -> None:
         self.add_float32(Keys.Attention.CLAMP_KQV.format(arch=self.arch), value)
 
-    def add_shared_kv_layers(self, value: float) -> None:
+    def add_shared_kv_layers(self, value: int) -> None:
-        self.add_float32(Keys.Attention.SHARED_KV_LAYERS.format(arch=self.arch), value)
+        self.add_uint32(Keys.Attention.SHARED_KV_LAYERS.format(arch=self.arch), value)
 
     def add_sliding_window_pattern(self, value: Sequence[bool]) -> None:
         self.add_array(Keys.Attention.SLIDING_WINDOW_PATTERN.format(arch=self.arch), value)
@@ -861,6 +861,9 @@ class GGUFWriter:
     def add_ssm_time_step_rank(self, value: int) -> None:
         self.add_uint32(Keys.SSM.TIME_STEP_RANK.format(arch=self.arch), value)
 
+    def add_ssm_group_count(self, value: int) -> None:
+        self.add_uint32(Keys.SSM.GROUP_COUNT.format(arch=self.arch), value)
+
     def add_ssm_dt_b_c_rms(self, value: bool) -> None:
         self.add_bool(Keys.SSM.DT_B_C_RMS.format(arch=self.arch), value)
 

gguf-py/gguf/tensor_mapping.py

@@ -477,7 +477,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.norm2",                   # nomic-bert
             "transformer.decoder_layer.{bid}.rms_norm_3",   # Grok
             "encoder.layer.{bid}.mlp.layernorm",            # jina-bert-v2
-            "encoder.layer.{bid}.layer_norm_2"              # jina-v2-code
+            "encoder.layer.{bid}.layer_norm_2",             # jina-v2-code
         ),
 
         MODEL_TENSOR.PER_LAYER_TOKEN_EMBD: (
@@ -574,6 +574,10 @@ class TensorNameMap:
             "backbone.layers.{bid}.mixer.D",
         ),
 
+        MODEL_TENSOR.SSM_NORM: (
+            "backbone.layers.{bid}.mixer.norm",  # mamba2
+        ),
+
         MODEL_TENSOR.SSM_OUT: (
             "model.layers.{bid}.out_proj",
             "backbone.layers.{bid}.mixer.out_proj",

gguf-py/gguf/vocab.py

@@ -245,9 +245,18 @@ class SpecialVocab:
         if not tokenizer_config:
             return True
         chat_template_alt = None
-        chat_template_file = path / 'chat_template.json'
+        chat_template_json = path / 'chat_template.json'
-        if chat_template_file.is_file():
+        chat_template_jinja = path / 'chat_template.jinja'
-            with open(chat_template_file, encoding = 'utf-8') as f:
+        if chat_template_jinja.is_file():
+            with open(chat_template_jinja, encoding = 'utf-8') as f:
+                chat_template_alt = f.read()
+            if additional_templates := list((path / 'additional_chat_templates').glob('*.jinja')):
+                chat_template_alt = [{'name': 'default', 'template': chat_template_alt}]
+                for template_path in additional_templates:
+                    with open(template_path, encoding = 'utf-8') as fp:
+                        chat_template_alt.append({'name': template_path.stem, 'template': fp.read()})
+        elif chat_template_json.is_file():
+            with open(chat_template_json, encoding = 'utf-8') as f:
                 chat_template_alt = json.load(f).get('chat_template')
         chat_template = tokenizer_config.get('chat_template', chat_template_alt)
         if chat_template is None or isinstance(chat_template, (str, list)):

klite.embd

@@ -3208,6 +3208,7 @@ Current version indicated by LITEVER below.
 	var schedule_multiplayer_major_change = false;
 	var last_request_str = "No Requests Available"; //full context of last submitted request
 	var last_response_obj = null;
+	var last_response_streamlog = "";
 	var lastcheckgenkey = ""; //for checking polled-streaming unique id when generating in kcpp
 	var kai_poll_recoverykey = ""; //for recovering a lost polled streaming in case of disconnect.
 	var globalabortcontroller = null;
@@ -5634,6 +5635,10 @@ Current version indicated by LITEVER below.
 				},
 				transform(chunk, ctrl) {
 					ctrl.buf += chunk;
+					if(chunk)
+					{
+						last_response_streamlog += chunk;
+					}
 					let evs = [];
 					let m;
 					while ((m = /^data: ?(.*)(\r?\n){2}/m.exec(ctrl.buf)) !== null) {
@@ -9343,6 +9348,10 @@ Current version indicated by LITEVER below.
 		{
 			lr += "\n\nResponse:\n" + JSON.stringify(last_response_obj);
 		}
+		if(last_response_streamlog)
+		{
+			lr += "\n\nResponse:\n" + last_response_streamlog;
+		}
 		msgbox(lr,"Last Request Info",false);
 	}
 	function show_last_logprobs()
@@ -10401,7 +10410,7 @@ Current version indicated by LITEVER below.
 		desired_oai_ep = transform_oai_ep(desired_oai_ep);
 
 		let oaiheaders = {};
-		if(desired_oai_key!=""){
+		if(desired_oai_key!="" && !desired_oai_ep.toLowerCase().includes("pollinations.ai")){
 			oaiheaders["Authorization"] = "Bearer " + desired_oai_key;
 		};
 		if (desired_oai_ep.toLowerCase().includes("api.mistral.ai")) {
@@ -14041,6 +14050,7 @@ Current version indicated by LITEVER below.
 		redo_arr = [];
 		last_request_str = "No Requests Available";
 		last_response_obj = null;
+		last_response_streamlog = "";
 		retry_prev_text = [];
 		retry_preserve_last = false;
 		retry_in_progress = false;
@@ -16471,6 +16481,7 @@ Current version indicated by LITEVER below.
 
 				last_request_str = JSON.stringify(submit_payload);
 				last_response_obj = null;
+				last_response_streamlog = "";
 				if (localsettings.tokenstreammode==2 && is_using_kcpp_with_sse()) {
 					let sub_endpt = apply_proxy_url(custom_kobold_endpoint + kobold_custom_gen_stream_endpoint);
 					kobold_api_stream_sse(sub_endpt, submit_payload);
@@ -16643,11 +16654,15 @@ Current version indicated by LITEVER below.
 
 				last_request_str = JSON.stringify(oai_payload);
 				last_response_obj = null;
+				last_response_streamlog = "";
 				let oaiheaders = {
-					'Content-Type': 'application/json',
+					'Content-Type': 'application/json'
-					'Authorization': 'Bearer ' + custom_oai_key
 				};
 
+				if (!targetep.toLowerCase().includes("pollinations.ai")) {
+                    oaiheaders['Authorization'] = 'Bearer ' + custom_oai_key;
+                }
+
 				if(targetep.toLowerCase().includes("openrouter.ai"))
 				{
 					oaiheaders["HTTP-Referer"] = "https://lite.koboldai.net";
@@ -16780,6 +16795,7 @@ Current version indicated by LITEVER below.
 
 				last_request_str = JSON.stringify(claude_payload);
 				last_response_obj = null;
+				last_response_streamlog = "";
 
 				let claudeheaders = {
 					'Content-Type': 'application/json',
@@ -16970,6 +16986,7 @@ Current version indicated by LITEVER below.
 
 				last_request_str = JSON.stringify(payload);
 				last_response_obj = null;
+				last_response_streamlog = "";
 
 				let geminiheaders = { 'Content-Type': 'application/json' };
 				if(is_browser_supports_sse() && localsettings.tokenstreammode!=0)
@@ -17018,6 +17035,7 @@ Current version indicated by LITEVER below.
 
 				last_request_str = JSON.stringify(cohere_payload);
 				last_response_obj = null;
+				last_response_streamlog = "";
 				let cohere_headers = {
 					'Content-Type': 'application/json',
 					'Authorization': 'Bearer ' + custom_cohere_key
@@ -17093,6 +17111,7 @@ Current version indicated by LITEVER below.
 
 			last_request_str = JSON.stringify(submit_payload);
 			last_response_obj = null;
+			last_response_streamlog = "";
 
 			fetch(horde_submit_endpoint, {
 				method: 'POST', // or 'PUT'

koboldcpp.py

@@ -62,7 +62,7 @@ dry_seq_break_max = 128
 extra_images_max = 4
 
 # global vars
-KcppVersion = "1.95.1"
+KcppVersion = "1.96"
 showdebug = True
 kcpp_instance = None #global running instance
 global_memory = {"tunnel_url": "", "restart_target":"", "input_to_exit":False, "load_complete":False, "restart_override_config_target":""}

src/llama-arch.cpp

@@ -45,6 +45,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_GEMMA3N,          "gemma3n"          },
     { LLM_ARCH_STARCODER2,       "starcoder2"       },
     { LLM_ARCH_MAMBA,            "mamba"            },
+    { LLM_ARCH_MAMBA2,           "mamba2"           },
     { LLM_ARCH_XVERSE,           "xverse"           },
     { LLM_ARCH_COMMAND_R,        "command-r"        },
     { LLM_ARCH_COHERE2,          "cohere2"          },
@@ -170,6 +171,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_SSM_INNER_SIZE,     "%s.ssm.inner_size"     },
     { LLM_KV_SSM_STATE_SIZE,     "%s.ssm.state_size"     },
     { LLM_KV_SSM_TIME_STEP_RANK, "%s.ssm.time_step_rank" },
+    { LLM_KV_SSM_GROUP_COUNT,    "%s.ssm.group_count"    },
     { LLM_KV_SSM_DT_B_C_RMS,     "%s.ssm.dt_b_c_rms"     },
 
     { LLM_KV_WKV_HEAD_SIZE, "%s.wkv.head_size" },
@@ -1004,6 +1006,22 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_SSM_OUT,         "blk.%d.ssm_out" },
         },
     },
+    {
+        LLM_ARCH_MAMBA2,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_SSM_IN,          "blk.%d.ssm_in" },
+            { LLM_TENSOR_SSM_CONV1D,      "blk.%d.ssm_conv1d" },
+            { LLM_TENSOR_SSM_DT,          "blk.%d.ssm_dt" },
+            { LLM_TENSOR_SSM_A,           "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_D,           "blk.%d.ssm_d" },
+            { LLM_TENSOR_SSM_NORM,        "blk.%d.ssm_norm" },
+            { LLM_TENSOR_SSM_OUT,         "blk.%d.ssm_out" },
+        },
+    },
     {
         LLM_ARCH_XVERSE,
         {
@@ -1761,6 +1779,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_SSM_CONV1D,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
     {LLM_TENSOR_SSM_A,                      {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_SCAN}},
     {LLM_TENSOR_SSM_D,                      {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_SSM_NORM,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_TIME_MIX_LERP_X,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_TIME_MIX_LN,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_CHANNEL_MIX_LERP_K,         {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
@@ -1894,6 +1913,7 @@ const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor) {
 bool llm_arch_is_recurrent(const llm_arch & arch) {
     switch (arch) {
         case LLM_ARCH_MAMBA:
+        case LLM_ARCH_MAMBA2:
         case LLM_ARCH_RWKV6:
         case LLM_ARCH_RWKV6QWEN2:
         case LLM_ARCH_RWKV7:

src/llama-arch.h

@@ -49,6 +49,7 @@ enum llm_arch {
     LLM_ARCH_GEMMA3N,
     LLM_ARCH_STARCODER2,
     LLM_ARCH_MAMBA,
+    LLM_ARCH_MAMBA2,
     LLM_ARCH_XVERSE,
     LLM_ARCH_COMMAND_R,
     LLM_ARCH_COHERE2,
@@ -174,6 +175,7 @@ enum llm_kv {
     LLM_KV_SSM_CONV_KERNEL,
     LLM_KV_SSM_STATE_SIZE,
     LLM_KV_SSM_TIME_STEP_RANK,
+    LLM_KV_SSM_GROUP_COUNT,
     LLM_KV_SSM_DT_B_C_RMS,
 
     LLM_KV_WKV_HEAD_SIZE,
@@ -293,6 +295,7 @@ enum llm_tensor {
     LLM_TENSOR_SSM_DT,
     LLM_TENSOR_SSM_A,
     LLM_TENSOR_SSM_D,
+    LLM_TENSOR_SSM_NORM,
     LLM_TENSOR_SSM_OUT,
     LLM_TENSOR_TIME_MIX_W0,
     LLM_TENSOR_TIME_MIX_W1,

src/llama-batch.cpp

@@ -166,6 +166,8 @@ bool llama_batch_allocr::init(
 
                 // note: tracking the other way around is not necessary for now
                 //seq_cpl[s0][s1] = true;
+
+                has_cpl = true;
             }
         }
     }
@@ -404,6 +406,10 @@ uint32_t llama_batch_allocr::get_n_outputs() const {
     return n_outputs;
 }
 
+uint32_t llama_batch_allocr::get_n_used() const {
+    return n_used;
+}
+
 std::vector<int32_t> & llama_batch_allocr::get_out_ids() {
     return out_ids;
 }
@@ -419,6 +425,8 @@ llama_pos llama_batch_allocr::seq_pos_max(llama_seq_id seq_id) const {
 void llama_batch_allocr::split_reset() {
     out_ids.clear();
 
+    n_used = 0;
+
     used.clear();
     used.resize(get_n_tokens(), false);
 
@@ -443,6 +451,7 @@ llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
         idxs.push_back(cur_idx);
 
         used[cur_idx] = true;
+        ++n_used;
 
         ++cur_idx;
 
@@ -458,9 +467,17 @@ llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
     return ubatch_add(idxs, idxs.size(), false);
 }
 
-llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch) {
+llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential) {
+    if (sequential && has_cpl) {
+        LLAMA_LOG_ERROR("%s: sequential split is not supported when there are coupled sequences in the input batch\n", __func__);
+
+        return {};
+    }
+
     std::vector<seq_set_t> cur_seq_set;
 
+    llama_seq_id last_seq_id = -1;
+
     // determine the non-overlapping sequence sets participating in this ubatch
     for (int32_t i = 0; i < batch.n_tokens; ++i) {
         if (used[i]) {
@@ -477,9 +494,16 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch) {
             }
         }
 
+        // accept only increasing sequence ids
+        if (sequential) {
+            add = add && (cur_seq_set.empty() || batch.seq_id[i][0] == last_seq_id + 1);
+        }
+
         if (add) {
             cur_seq_set.push_back(seq_set[i]);
 
+            last_seq_id = batch.seq_id[i][0];
+
             if (cur_seq_set.size() > n_ubatch) {
                 break;
             }
@@ -528,6 +552,7 @@ llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch) {
             idxs_per_seq[s].push_back(idx);
 
             used[idx] = true;
+            ++n_used;
 
             ++cur_idx[s];
         }
@@ -569,6 +594,7 @@ llama_ubatch llama_batch_allocr::split_seq(uint32_t n_ubatch) {
         idxs.push_back(cur_idx);
 
         used[cur_idx] = true;
+        ++n_used;
 
         if (idxs.size() >= n_ubatch) {
             break;

src/llama-batch.h

@@ -54,6 +54,7 @@ public:
 
     uint32_t get_n_tokens()  const;
     uint32_t get_n_outputs() const;
+    uint32_t get_n_used()    const;
 
     // the array of output indices in the order they were encountered during the ubatch splitting
     std::vector<int32_t> & get_out_ids();
@@ -69,7 +70,8 @@ public:
     llama_ubatch split_simple(uint32_t n_ubatch);
 
     // make ubatches of equal-length sequences sets
-    llama_ubatch split_equal(uint32_t n_ubatch);
+    // if sequential == true, the tokens in the ubatch will have increasing sequential sequence ids
+    llama_ubatch split_equal(uint32_t n_ubatch, bool sequential);
 
     // sequence-set-wise split - each ubatch contains a single sequence-set
     llama_ubatch split_seq(uint32_t n_ubatch);
@@ -112,6 +114,9 @@ private:
     using pos_set_t = std::set<llama_pos>;
     using seq_cpl_t = std::vector<bool>;
 
+    // helper flag to quickly determine if there are any coupled sequences in the batch
+    bool has_cpl;
+
     std::vector<pos_set_t> seq_pos; // seq_pos[s]: the set of positions in sequence s
     std::vector<seq_cpl_t> seq_cpl; // seq_cpl[s0][s1]: if sequence s0 is coupled to sequence s1
 
@@ -125,6 +130,8 @@ private:
     // batch indices of the output
     std::vector<int32_t> out_ids;
 
+    uint32_t n_used;
+
     // used[i] indicates if token i has already been used in a previous ubatch
     std::vector<bool> used;
 

src/llama-graph.cpp

@@ -281,19 +281,22 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
 }
 
 void llm_graph_input_attn_kv_unified::set_input(const llama_ubatch * ubatch) {
-    if (self_kq_mask) {
+    mctx->set_input_k_idxs(self_k_idxs, ubatch);
+    mctx->set_input_v_idxs(self_v_idxs, ubatch);
+
     mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
-    }
 }
 
 void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch) {
-    if (self_kq_mask) {
+    mctx->get_base()->set_input_k_idxs(self_k_idxs, ubatch);
-        mctx->get_base()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
+    mctx->get_base()->set_input_v_idxs(self_v_idxs, ubatch);
-    }
+
+    mctx->get_base()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
+
+    mctx->get_swa()->set_input_k_idxs(self_k_idxs_swa, ubatch);
+    mctx->get_swa()->set_input_v_idxs(self_v_idxs_swa, ubatch);
 
-    if (self_kq_mask_swa) {
     mctx->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn);
-    }
 }
 
 void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
@@ -333,9 +336,10 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
 }
 
 void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) {
-    if (self_kq_mask) {
+    mctx->get_attn()->set_input_k_idxs(self_k_idxs, ubatch);
+    mctx->get_attn()->set_input_v_idxs(self_v_idxs, ubatch);
+
     mctx->get_attn()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
-    }
 
     const int64_t n_rs = mctx->get_recr()->get_n_rs();
 
@@ -350,7 +354,8 @@ void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) {
     }
 }
 
-void llm_graph_input_one::set_input(const llama_ubatch *) {
+void llm_graph_input_one::set_input(const llama_ubatch * ubatch) {
+    GGML_UNUSED(ubatch);
     GGML_ASSERT(one && ggml_nelements(one) == 1);
     float f_one = 1.0f;
     ggml_backend_tensor_set(one, &f_one, 0, sizeof(float));
@@ -997,8 +1002,10 @@ llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
 
         const auto n_kv = inp->mctx->get_attn()->get_n_kv();
 
-        inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+        inp->self_k_idxs = mctx_cur->get_attn()->build_input_k_idxs(ctx0, ubatch);
-        //cb(inp->self_kq_mask, "KQ_mask", -1);
+        inp->self_v_idxs = mctx_cur->get_attn()->build_input_v_idxs(ctx0, ubatch);
+
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -1135,8 +1142,7 @@ llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() con
     auto inp = std::make_unique<llm_graph_input_attn_no_cache>(hparams, cparams);
 
     // note: there is no KV cache, so the number of KV values is equal to the number of tokens in the batch
-    inp->kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+    inp->kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
-    //cb(inp_kq_mask, "KQ_mask", -1);
     ggml_set_input(inp->kq_mask);
 
     inp->kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->kq_mask, GGML_TYPE_F16) : inp->kq_mask;
@@ -1198,8 +1204,10 @@ llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified()
 
         const auto n_kv = mctx_cur->get_n_kv();
 
-        inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+        inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch);
-        //cb(inp->self_kq_mask, "KQ_mask", -1);
+        inp->self_v_idxs = mctx_cur->build_input_v_idxs(ctx0, ubatch);
+
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -1230,8 +1238,11 @@ ggml_tensor * llm_graph_context::build_attn(
 
     // store to KV cache
     {
-        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, il));
+        const auto & k_idxs = inp->get_k_idxs();
-        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, il));
+        const auto & v_idxs = inp->get_v_idxs();
+
+        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
+        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, v_idxs, il));
     }
 
     const auto & kq_mask = inp->get_kq_mask();
@@ -1290,11 +1301,15 @@ ggml_tensor * llm_graph_context::build_attn(
 
     // optionally store to KV cache
     if (k_cur) {
-        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, il));
+        const auto & k_idxs = is_swa ? inp->get_k_idxs_swa() : inp->get_k_idxs();
+
+        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
     }
 
     if (v_cur) {
-        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, il));
+        const auto & v_idxs = is_swa ? inp->get_v_idxs_swa() : inp->get_v_idxs();
+
+        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, v_idxs, il));
     }
 
     const auto & kq_mask = is_swa ? inp->get_kq_mask_swa() : inp->get_kq_mask();
@@ -1326,7 +1341,7 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
 
     const int32_t n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;
 
-    inp->cross_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+    inp->cross_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
     ggml_set_input(inp->cross_kq_mask);
 
     inp->cross_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->cross_kq_mask, GGML_TYPE_F16) : inp->cross_kq_mask;
@@ -1398,8 +1413,11 @@ ggml_tensor * llm_graph_context::build_attn(
 
     // store to KV cache
     {
-        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, il));
+        const auto & k_idxs = inp->get_k_idxs();
-        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, il));
+        const auto & v_idxs = inp->get_v_idxs();
+
+        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
+        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, v_idxs, il));
     }
 
     const auto & kq_mask = inp->get_kq_mask();
@@ -1434,8 +1452,10 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
     {
         const auto n_kv = mctx_cur->get_base()->get_n_kv();
 
-        inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+        inp->self_k_idxs = mctx_cur->get_base()->build_input_k_idxs(ctx0, ubatch);
-        //cb(inp->self_kq_mask, "KQ_mask", -1);
+        inp->self_v_idxs = mctx_cur->get_base()->build_input_v_idxs(ctx0, ubatch);
+
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -1446,8 +1466,10 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
 
         const auto n_kv = mctx_cur->get_swa()->get_n_kv();
 
-        inp->self_kq_mask_swa = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+        inp->self_k_idxs_swa = mctx_cur->get_swa()->build_input_k_idxs(ctx0, ubatch);
-        //cb(inp->self_kq_mask_swa, "KQ_mask_swa", -1);
+        inp->self_v_idxs_swa = mctx_cur->get_swa()->build_input_v_idxs(ctx0, ubatch);
+
+        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
         ggml_set_input(inp->self_kq_mask_swa);
 
         inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
@@ -1466,7 +1488,7 @@ ggml_tensor * llm_graph_context::build_rs(
            uint32_t   kv_head,
            uint32_t   kv_size,
             int32_t   rs_zero,
-               bool   avoid_copies) const {
+        const llm_graph_get_rows_fn & get_state_rows) const {
 
     ggml_tensor * states = ggml_reshape_2d(ctx0, s, state_size, kv_size);
 
@@ -1475,19 +1497,11 @@ ggml_tensor * llm_graph_context::build_rs(
     ggml_tensor * state_zero = ggml_view_1d(ctx0, states, state_size*(rs_zero >= 0), rs_zero*states->nb[1]*(rs_zero >= 0));
     ggml_build_forward_expand(gf, ggml_scale_inplace(ctx0, state_zero, 0));
 
-    ggml_tensor * output_states;
-
-    if (!avoid_copies) {
     // copy states
     // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv
     // {state_size, kv_size} -> {state_size, n_seqs}
-        output_states = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_seqs, 0));
+    ggml_tensor * output_states = get_state_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_seqs, 0));
     ggml_build_forward_expand(gf, output_states);
-    } else {
-        // FIXME: make the gathering operation happen before the copy below
-        //        (maybe with an optional lambda function passed as a parameter instead of `avoid_copies`?)
-        output_states = states;
-    }
 
     // copy extra states which won't be changed further (between n_seqs and n_kv)
     ggml_tensor * states_extra = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_kv - n_seqs, n_seqs*state_copy->nb[0]));
@@ -1518,10 +1532,10 @@ ggml_tensor * llm_graph_context::build_rs(
         ggml_tensor * s,
             int32_t   state_size,
             int32_t   n_seqs,
-               bool   avoid_copies) const {
+        const llm_graph_get_rows_fn & get_state_rows) const {
-    const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
+    const auto * kv_state = static_cast<const llama_memory_recurrent_context *>(mctx);
 
-    return build_rs(gf, s, inp->s_copy, state_size, n_seqs, mctx_cur->get_n_rs(), mctx_cur->get_head(), mctx_cur->get_size(), mctx_cur->get_rs_z(), avoid_copies);
+    return build_rs(gf, s, inp->s_copy, state_size, n_seqs, kv_state->get_n_rs(), kv_state->get_head(), kv_state->get_size(), kv_state->get_rs_z(), get_state_rows);
 }
 
 ggml_tensor * llm_graph_context::build_rs(
@@ -1530,10 +1544,10 @@ ggml_tensor * llm_graph_context::build_rs(
         ggml_tensor * s,
             int32_t   state_size,
             int32_t   n_seqs,
-               bool   avoid_copies) const {
+        const llm_graph_get_rows_fn & get_state_rows) const {
-    const auto * mctx_cur = static_cast<const llama_memory_hybrid_context *>(mctx)->get_recr();
+    const auto * kv_state = static_cast<const llama_memory_hybrid_context *>(mctx)->get_recr();
 
-    return build_rs(gf, s, inp->s_copy, state_size, n_seqs, mctx_cur->get_n_rs(), mctx_cur->get_head(), mctx_cur->get_size(), mctx_cur->get_rs_z(), avoid_copies);
+    return build_rs(gf, s, inp->s_copy, state_size, n_seqs, kv_state->get_n_rs(), kv_state->get_head(), kv_state->get_size(), kv_state->get_rs_z(), get_state_rows);
 }
 
 ggml_tensor * llm_graph_context::build_rwkv_token_shift_load(

src/llama-graph.h

@@ -228,8 +228,8 @@ public:
 
     ggml_tensor * get_kq_mask() const { return kq_mask_cnv; }
 
-    ggml_tensor * kq_mask     = nullptr; // F32 [n_tokens, n_batch]
+    ggml_tensor * kq_mask     = nullptr; // F32 [n_tokens, n_batch, 1, 1]
-    ggml_tensor * kq_mask_cnv = nullptr; //     [n_tokens, n_batch]
+    ggml_tensor * kq_mask_cnv = nullptr; //     [n_tokens, n_batch, 1, 1]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;
@@ -249,10 +249,16 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    ggml_tensor * get_k_idxs() const { return self_k_idxs; }
+    ggml_tensor * get_v_idxs() const { return self_v_idxs; }
+
     ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
 
-    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch]
+    ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
-    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch]
+    ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch]
+
+    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch, 1, 1]
+    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch, 1, 1]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;
@@ -274,13 +280,23 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    ggml_tensor * get_k_idxs()     const { return self_k_idxs; }
+    ggml_tensor * get_v_idxs()     const { return self_v_idxs; }
+    ggml_tensor * get_k_idxs_swa() const { return self_k_idxs_swa; }
+    ggml_tensor * get_v_idxs_swa() const { return self_v_idxs_swa; }
+
     ggml_tensor * get_kq_mask()     const { return self_kq_mask_cnv; }
     ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }
 
-    ggml_tensor * self_kq_mask         = nullptr; // F32 [n_kv, n_batch]
+    ggml_tensor * self_k_idxs     = nullptr; // I64 [n_batch]
-    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_kv, n_batch]
+    ggml_tensor * self_v_idxs     = nullptr; // I64 [n_batch]
-    ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_kv, n_batch]
+    ggml_tensor * self_k_idxs_swa = nullptr; // I64 [n_batch]
-    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_kv, n_batch]
+    ggml_tensor * self_v_idxs_swa = nullptr; // I64 [n_batch]
+
+    ggml_tensor * self_kq_mask         = nullptr; // F32 [n_kv, n_batch, 1, 1]
+    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_kv, n_batch, 1, 1]
+    ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_kv, n_batch, 1, 1]
+    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_kv, n_batch, 1, 1]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;
@@ -297,8 +313,8 @@ public:
 
     ggml_tensor * get_kq_mask_cross() const { return cross_kq_mask_cnv; }
 
-    ggml_tensor * cross_kq_mask     = nullptr; // F32 [n_outputs_enc, n_batch]
+    ggml_tensor * cross_kq_mask     = nullptr; // F32 [n_outputs_enc, n_batch, 1, 1]
-    ggml_tensor * cross_kq_mask_cnv = nullptr; // F32 [n_outputs_enc, n_batch]
+    ggml_tensor * cross_kq_mask_cnv = nullptr; // F32 [n_outputs_enc, n_batch, 1, 1]
 
     const llama_cross * cross = nullptr;
 };
@@ -319,10 +335,16 @@ public:
 
     ggml_tensor * s_copy; // I32 [kv_size]
 
+    ggml_tensor * get_k_idxs() const { return self_k_idxs; }
+    ggml_tensor * get_v_idxs() const { return self_v_idxs; }
+
     ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
 
-    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch]
+    ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
-    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch]
+    ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch]
+
+    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch, 1, 1]
+    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch, 1, 1]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;
@@ -336,7 +358,7 @@ public:
     llm_graph_input_one() {}
     virtual ~llm_graph_input_one() = default;
 
-    void set_input(const llama_ubatch *) override;
+    void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * one = nullptr; // F32
 };
@@ -424,6 +446,9 @@ struct llm_graph_params {
     const llm_graph_cb & cb;
 };
 
+// used in build_rs to properly order writes and avoid unnecessary copies
+using llm_graph_get_rows_fn = std::function<ggml_tensor * (ggml_context *, ggml_tensor * states, ggml_tensor * ids)>;
+
 struct llm_graph_context {
     const llm_arch arch;
 
@@ -663,7 +688,7 @@ struct llm_graph_context {
                uint32_t   kv_head,
                uint32_t   kv_size,
                 int32_t   rs_zero,
-                   bool   avoid_copies = false) const;
+            const llm_graph_get_rows_fn & get_state_rows = ggml_get_rows) const;
 
     llm_graph_input_rs * build_rs_inp() const;
 
@@ -673,7 +698,7 @@ struct llm_graph_context {
             ggml_tensor * s,
                 int32_t   state_size,
                 int32_t   n_seqs,
-                   bool   avoid_copies = false) const;
+            const llm_graph_get_rows_fn & get_state_rows = ggml_get_rows) const;
 
     ggml_tensor * build_rs(
             llm_graph_input_mem_hybrid * inp,
@@ -681,7 +706,7 @@ struct llm_graph_context {
             ggml_tensor * s,
                 int32_t   state_size,
                 int32_t   n_seqs,
-                   bool   avoid_copies = false) const;
+            const llm_graph_get_rows_fn & get_state_rows = ggml_get_rows) const;
 
     ggml_tensor * build_rwkv_token_shift_load(
         llm_graph_input_rs * inp,

src/llama-hparams.cpp

@@ -73,7 +73,8 @@ uint32_t llama_hparams::n_embd_r() const {
 
     // TODO: maybe support other convolution strides than 1
     // NOTE: since the first column of the conv_state is shifted out each time, it's not actually needed
-    return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * ssm_d_inner;
+    // Corresponds to Mamba's conv_states size
+    return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * (ssm_d_inner + 2*ssm_n_group*ssm_d_state);
 }
 
 uint32_t llama_hparams::n_embd_s() const {

src/llama-hparams.h

@@ -114,6 +114,7 @@ struct llama_hparams {
     uint32_t ssm_d_inner = 0;
     uint32_t ssm_d_state = 0;
     uint32_t ssm_dt_rank = 0;
+    uint32_t ssm_n_group = 0;
 
     // for hybrid state space models
     std::array<bool, LLAMA_MAX_LAYERS> recurrent_layer_arr;

src/llama-kv-cache-unified-iswa.cpp

@@ -117,20 +117,25 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
             ubatches.push_back(std::move(ubatch)); // NOLINT
         }
 
-        auto heads_base = kv_base->prepare(ubatches);
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
-        if (heads_base.empty()) {
+            // failed to find a suitable split
             break;
         }
 
-        auto heads_swa = kv_swa->prepare(ubatches);
+        auto sinfos_base = kv_base->prepare(ubatches);
-        if (heads_swa.empty()) {
+        if (sinfos_base.empty()) {
             break;
         }
 
-        assert(heads_base.size() == heads_swa.size());
+        auto sinfos_swa = kv_swa->prepare(ubatches);
+        if (sinfos_swa.empty()) {
+            break;
+        }
+
+        assert(sinfos_base.size() == sinfos_swa.size());
 
         return std::make_unique<llama_kv_cache_unified_iswa_context>(
-                this, std::move(heads_base), std::move(heads_swa), std::move(ubatches));
+                this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches));
     } while (false);
 
     // if it fails, try equal split
@@ -139,7 +144,7 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
 
         std::vector<llama_ubatch> ubatches;
         while (true) {
-            auto ubatch = balloc.split_equal(n_ubatch);
+            auto ubatch = balloc.split_equal(n_ubatch, false);
 
             if (ubatch.n_tokens == 0) {
                 break;
@@ -148,20 +153,25 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
             ubatches.push_back(std::move(ubatch)); // NOLINT
         }
 
-        auto heads_base = kv_base->prepare(ubatches);
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
-        if (heads_base.empty()) {
+            // failed to find a suitable split
             break;
         }
 
-        auto heads_swa = kv_swa->prepare(ubatches);
+        auto sinfos_base = kv_base->prepare(ubatches);
-        if (heads_swa.empty()) {
+        if (sinfos_base.empty()) {
             break;
         }
 
-        assert(heads_base.size() == heads_swa.size());
+        auto sinfos_swa = kv_swa->prepare(ubatches);
+        if (sinfos_swa.empty()) {
+            break;
+        }
+
+        assert(sinfos_base.size() == sinfos_swa.size());
 
         return std::make_unique<llama_kv_cache_unified_iswa_context>(
-                this, std::move(heads_base), std::move(heads_swa), std::move(ubatches));
+                this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches));
     } while (false);
 
     // TODO: if we fail again, we should attempt different splitting strategies
@@ -224,13 +234,13 @@ llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
 
 llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
         llama_kv_cache_unified_iswa * kv,
-        std::vector<uint32_t> heads_base,
+        slot_info_vec_t sinfos_base,
-        std::vector<uint32_t> heads_swa,
+        slot_info_vec_t sinfos_swa,
         std::vector<llama_ubatch> ubatches) :
     ubatches(std::move(ubatches)),
     // note: here we copy the ubatches. not sure if this is ideal
-    ctx_base(new llama_kv_cache_unified_context(kv->get_base(), std::move(heads_base), this->ubatches)),
+    ctx_base(new llama_kv_cache_unified_context(kv->get_base(), std::move(sinfos_base), this->ubatches)),
-    ctx_swa (new llama_kv_cache_unified_context(kv->get_swa (), std::move(heads_swa),  this->ubatches)),
+    ctx_swa (new llama_kv_cache_unified_context(kv->get_swa (), std::move(sinfos_swa),  this->ubatches)),
     status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
 }
 

src/llama-kv-cache-unified-iswa.h

@@ -74,6 +74,8 @@ private:
 
 class llama_kv_cache_unified_iswa_context : public llama_memory_context_i {
 public:
+    using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t;
+
     // used for errors
     llama_kv_cache_unified_iswa_context(llama_memory_status status);
 
@@ -90,8 +92,8 @@ public:
     // used to create a batch processing context from a batch
     llama_kv_cache_unified_iswa_context(
             llama_kv_cache_unified_iswa * kv,
-            std::vector<uint32_t> heads_base,
+            slot_info_vec_t sinfos_base,
-            std::vector<uint32_t> heads_swa,
+            slot_info_vec_t sinfos_swa,
             std::vector<llama_ubatch> ubatches);
 
     virtual ~llama_kv_cache_unified_iswa_context();

src/llama-kv-cache-unified.cpp

@@ -156,6 +156,13 @@ llama_kv_cache_unified::llama_kv_cache_unified(
 
     const char * LLAMA_KV_CACHE_DEBUG = getenv("LLAMA_KV_CACHE_DEBUG");
     debug = LLAMA_KV_CACHE_DEBUG ? atoi(LLAMA_KV_CACHE_DEBUG) : 0;
+
+    const char * LLAMA_SET_ROWS = getenv("LLAMA_SET_ROWS");
+    supports_set_rows = LLAMA_SET_ROWS ? atoi(LLAMA_SET_ROWS) : 0;
+
+    if (!supports_set_rows) {
+        LLAMA_LOG_WARN("%s: LLAMA_SET_ROWS=0, using old ggml_cpy() method for backwards compatibility\n", __func__);
+    }
 }
 
 void llama_kv_cache_unified::clear(bool data) {
@@ -353,13 +360,18 @@ llama_memory_context_ptr llama_kv_cache_unified::init_batch(
             ubatches.push_back(std::move(ubatch)); // NOLINT
         }
 
-        auto heads = prepare(ubatches);
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
-        if (heads.empty()) {
+            // failed to find a suitable split
+            break;
+        }
+
+        auto sinfos = prepare(ubatches);
+        if (sinfos.empty()) {
             break;
         }
 
         return std::make_unique<llama_kv_cache_unified_context>(
-                this, std::move(heads), std::move(ubatches));
+                this, std::move(sinfos), std::move(ubatches));
     } while (false);
 
     return std::make_unique<llama_kv_cache_unified_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
@@ -402,12 +414,13 @@ llama_memory_context_ptr llama_kv_cache_unified::init_update(llama_context * lct
     return std::make_unique<llama_kv_cache_unified_context>(this, lctx, do_shift, std::move(dinfo));
 }
 
-llama_kv_cache_unified::ubatch_heads llama_kv_cache_unified::prepare(const std::vector<llama_ubatch> & ubatches) {
+llama_kv_cache_unified::slot_info_vec_t llama_kv_cache_unified::prepare(const std::vector<llama_ubatch> & ubatches) {
-    llama_kv_cache_unified::ubatch_heads res;
+    llama_kv_cache_unified::slot_info_vec_t res;
 
     struct state {
         uint32_t head_old; // old position of the head, before placing the ubatch
-        uint32_t head_new; // new position of the head, after placing the ubatch
+
+        slot_info sinfo; // slot info for the ubatch
 
         llama_kv_cells_unified cells; // copy of the old cells, before placing the ubatch
     };
@@ -418,26 +431,29 @@ llama_kv_cache_unified::ubatch_heads llama_kv_cache_unified::prepare(const std::
     bool success = true;
 
     for (const auto & ubatch : ubatches) {
+        // non-continuous slots require support for ggml_set_rows()
+        const bool cont = supports_set_rows ? false : true;
+
         // only find a suitable slot for the ubatch. don't modify the cells yet
-        const int32_t head_new = find_slot(ubatch);
+        const auto sinfo_new = find_slot(ubatch, cont);
-        if (head_new < 0) {
+        if (sinfo_new.empty()) {
             success = false;
             break;
         }
 
         // remeber the position that we found
-        res.push_back(head_new);
+        res.push_back(sinfo_new);
 
         // store the old state of the cells in the recovery stack
-        states.push_back({head, (uint32_t) head_new, cells.cp(head_new, ubatch.n_tokens)});
+        states.push_back({head, sinfo_new, cells.cp(sinfo_new.idxs)});
 
         // now emplace the ubatch
-        apply_ubatch(head_new, ubatch);
+        apply_ubatch(sinfo_new, ubatch);
     }
 
     // iterate backwards and restore the cells to their original state
     for (auto it = states.rbegin(); it != states.rend(); ++it) {
-        cells.set(it->head_new, it->cells);
+        cells.set(it->sinfo.idxs, it->cells);
         head = it->head_old;
     }
 
@@ -539,7 +555,7 @@ bool llama_kv_cache_unified::update(llama_context * lctx, bool do_shift, const d
     return updated;
 }
 
-int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
+llama_kv_cache_unified::slot_info llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch, bool cont) const {
     const uint32_t n_tokens = ubatch.n_tokens;
 
     uint32_t head_cur = this->head;
@@ -552,7 +568,7 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
 
     if (n_tokens > cells.size()) {
         LLAMA_LOG_ERROR("%s: n_tokens = %d > size = %u\n", __func__, n_tokens, cells.size());
-        return -1;
+        return { };
     }
 
     if (debug > 0) {
@@ -615,15 +631,26 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
 
     uint32_t n_tested = 0;
 
+    // for continuous slots, we test that all tokens in the ubatch fit, starting from the current head
+    // for non-continuous slots, we test the tokens one by one
+    const uint32_t n_test = cont ? n_tokens : 1;
+
+    slot_info res;
+
+    auto & idxs = res.idxs;
+
+    idxs.reserve(n_tokens);
+
     while (true) {
-        if (head_cur + n_tokens > cells.size()) {
+        if (head_cur + n_test > cells.size()) {
             n_tested += cells.size() - head_cur;
             head_cur = 0;
             continue;
         }
 
-        bool found = true;
+        for (uint32_t i = 0; i < n_test; i++) {
-        for (uint32_t i = 0; i < n_tokens; i++) {
+            const auto idx = head_cur;
+
             //const llama_pos    pos    = ubatch.pos[i];
             //const llama_seq_id seq_id = ubatch.seq_id[i][0];
 
@@ -633,19 +660,19 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
             //    - (disabled) mask causally, if the sequence is the same as the one we are inserting
             //    - mask SWA, using current max pos for that sequence in the cache
             //                always insert in the cell with minimum pos
-            bool can_use = cells.is_empty(head_cur + i);
+            bool can_use = cells.is_empty(idx);
 
-            if (!can_use && cells.seq_count(head_cur + i) == 1) {
+            if (!can_use && cells.seq_count(idx) == 1) {
-                const llama_pos pos_cell = cells.pos_get(head_cur + i);
+                const llama_pos pos_cell = cells.pos_get(idx);
 
                 // (disabled) causal mask
                 // note: it's better to purge any "future" tokens beforehand
-                //if (cells.seq_has(head_cur + i, seq_id)) {
+                //if (cells.seq_has(idx, seq_id)) {
                 //    can_use = pos_cell >= pos;
                 //}
 
                 if (!can_use) {
-                    const llama_seq_id seq_id_cell = cells.seq_get(head_cur + i);
+                    const llama_seq_id seq_id_cell = cells.seq_get(idx);
 
                     // SWA mask
                     if (is_masked_swa(pos_cell, cells.seq_pos_max(seq_id_cell) + 1)) {
@@ -654,28 +681,39 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
                 }
             }
 
-            if (!can_use) {
+            head_cur++;
-                found = false;
+            n_tested++;
-                head_cur += i + 1;
+
-                n_tested += i + 1;
+            if (can_use) {
+                idxs.push_back(idx);
+            } else {
                 break;
             }
         }
 
-        if (found) {
+        if (idxs.size() == n_tokens) {
             break;
         }
 
+        if (cont) {
+            idxs.clear();
+        }
+
         if (n_tested >= cells.size()) {
             //LLAMA_LOG_ERROR("%s: failed to find a slot for %d tokens\n", __func__, n_tokens);
-            return -1;
+            return { };
         }
     }
 
-    return head_cur;
+    // we didn't find a suitable slot - return empty result
+    if (idxs.size() < n_tokens) {
+        res.clear();
+    }
+
+    return res;
 }
 
-void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch & ubatch) {
+void llama_kv_cache_unified::apply_ubatch(const slot_info & sinfo, const llama_ubatch & ubatch) {
     // keep track of the max sequence position that we would overwrite with this ubatch
     // for non-SWA cache, this would be always empty
     llama_seq_id seq_pos_max_rm[LLAMA_MAX_SEQ];
@@ -683,22 +721,26 @@ void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch
         seq_pos_max_rm[s] = -1;
     }
 
-    for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
+    assert(ubatch.n_tokens == sinfo.idxs.size());
-        if (!cells.is_empty(head_cur + i)) {
-            assert(cells.seq_count(head_cur + i) == 1);
 
-            const llama_seq_id seq_id = cells.seq_get(head_cur + i);
+    for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
-            const llama_pos    pos    = cells.pos_get(head_cur + i);
+        const auto idx = sinfo.idxs.at(i);
+
+        if (!cells.is_empty(idx)) {
+            assert(cells.seq_count(idx) == 1);
+
+            const llama_seq_id seq_id = cells.seq_get(idx);
+            const llama_pos    pos    = cells.pos_get(idx);
 
             seq_pos_max_rm[seq_id] = std::max(seq_pos_max_rm[seq_id], pos);
 
-            cells.rm(head_cur + i);
+            cells.rm(idx);
         }
 
-        cells.pos_set(head_cur + i, ubatch.pos[i]);
+        cells.pos_set(idx, ubatch.pos[i]);
 
         for (int32_t s = 0; s < ubatch.n_seq_id[i]; s++) {
-            cells.seq_add(head_cur + i, ubatch.seq_id[i][s]);
+            cells.seq_add(idx, ubatch.seq_id[i][s]);
         }
     }
 
@@ -719,7 +761,7 @@ void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch
     }
 
     // move the head at the end of the slot
-    head = head_cur + ubatch.n_tokens;
+    head = sinfo.idxs.back() + 1;
 }
 
 bool llama_kv_cache_unified::get_can_shift() const {
@@ -772,47 +814,133 @@ ggml_tensor * llama_kv_cache_unified::get_v(ggml_context * ctx, int32_t il, uint
             0);
 }
 
-ggml_tensor * llama_kv_cache_unified::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il, uint32_t head_cur) const {
+ggml_tensor * llama_kv_cache_unified::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const {
     const int32_t ikv = map_layer_ids.at(il);
 
     auto * k = layers[ikv].k;
 
+    const int64_t n_embd_k_gqa = k->ne[0];
     const int64_t n_tokens = k_cur->ne[2];
 
+    k_cur = ggml_reshape_2d(ctx, k_cur, k->ne[0], n_tokens);
+
+    if (k_idxs && supports_set_rows) {
+        return ggml_set_rows(ctx, k, k_cur, k_idxs);
+    }
+
+    // TODO: fallback to old ggml_cpy() method for backwards compatibility
+    //       will be removed when ggml_set_rows() is adopted by all backends
+
     ggml_tensor * k_view = ggml_view_1d(ctx, k,
-            n_tokens*hparams.n_embd_k_gqa(il),
+            n_tokens*n_embd_k_gqa,
-            ggml_row_size(k->type, hparams.n_embd_k_gqa(il))*head_cur);
+            ggml_row_size(k->type, n_embd_k_gqa)*sinfo.head());
 
     return ggml_cpy(ctx, k_cur, k_view);
 }
 
-ggml_tensor * llama_kv_cache_unified::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il, uint32_t head_cur) const {
+ggml_tensor * llama_kv_cache_unified::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il, const slot_info & sinfo) const {
     const int32_t ikv = map_layer_ids.at(il);
 
     auto * v = layers[ikv].v;
 
+    const int64_t n_embd_v_gqa = v->ne[0];
     const int64_t n_tokens = v_cur->ne[2];
 
-    v_cur = ggml_reshape_2d(ctx, v_cur, hparams.n_embd_v_gqa(il), n_tokens);
+    v_cur = ggml_reshape_2d(ctx, v_cur, n_embd_v_gqa, n_tokens);
+
+    if (v_idxs && supports_set_rows) {
+        if (!v_trans) {
+            return ggml_set_rows(ctx, v, v_cur, v_idxs);
+        }
+
+        // the row becomes a single element
+        ggml_tensor * v_view = ggml_reshape_3d(ctx, v, 1, v->ne[1], v->ne[0]);
+
+        // note: the V cache is transposed when not using flash attention
+        v_cur = ggml_permute(ctx, ggml_reshape_3d(ctx, v_cur, v_cur->ne[0], 1, v_cur->ne[1]), 2, 0, 1, 3);
+
+        // note: we can be more explicit here at the cost of extra cont
+        //       however, above we take advantage that a row of single element is always continuous regardless of the row stride
+        //v_cur = ggml_transpose(ctx, v_cur);
+        //v_cur = ggml_cont_3d(ctx, v_cur, 1, v_cur->ne[0], v_cur->ne[1]);
+
+        // we broadcast the KV indices n_embd_v_gqa times
+        // v      [1,        n_kv,     n_embd_v_gqa]
+        // v_cur  [1,        n_tokens, n_embd_v_gqa]
+        // v_idxs [n_tokens, 1,        1]
+        return ggml_set_rows(ctx, v_view, v_cur, v_idxs);
+    }
+
+    // TODO: fallback to old ggml_cpy() method for backwards compatibility
+    //       will be removed when ggml_set_rows() is adopted by all backends
 
     ggml_tensor * v_view = nullptr;
 
     if (!v_trans) {
         v_view = ggml_view_1d(ctx, v,
-                n_tokens*hparams.n_embd_v_gqa(il),
+                n_tokens*n_embd_v_gqa,
-                ggml_row_size(v->type, hparams.n_embd_v_gqa(il))*head_cur);
+                ggml_row_size(v->type, n_embd_v_gqa)*sinfo.head());
     } else {
-        // note: the V cache is transposed when not using flash attention
-        v_view = ggml_view_2d(ctx, v, n_tokens, hparams.n_embd_v_gqa(il),
-                (v->ne[1])*ggml_element_size(v),
-                (head_cur)*ggml_element_size(v));
-
         v_cur = ggml_transpose(ctx, v_cur);
+
+        v_view = ggml_view_2d(ctx, v, n_tokens, n_embd_v_gqa,
+                (v->ne[1]    )*ggml_element_size(v),
+                (sinfo.head())*ggml_element_size(v));
     }
 
     return ggml_cpy(ctx, v_cur, v_view);
 }
 
+ggml_tensor * llama_kv_cache_unified::build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
+    const uint32_t n_tokens = ubatch.n_tokens;
+
+    ggml_tensor * k_idxs = ggml_new_tensor_1d(ctx, GGML_TYPE_I64, n_tokens);
+
+    ggml_set_input(k_idxs);
+
+    return k_idxs;
+}
+
+ggml_tensor * llama_kv_cache_unified::build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
+    const uint32_t n_tokens = ubatch.n_tokens;
+
+    ggml_tensor * v_idxs = ggml_new_tensor_1d(ctx, GGML_TYPE_I64, n_tokens);
+
+    ggml_set_input(v_idxs);
+
+    return v_idxs;
+}
+
+void llama_kv_cache_unified::set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const {
+    if (!supports_set_rows) {
+        return;
+    }
+
+    const uint32_t n_tokens = ubatch->n_tokens;
+
+    GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
+    int64_t * data = (int64_t *) dst->data;
+
+    for (int64_t i = 0; i < n_tokens; ++i) {
+        data[i] = sinfo.idxs.at(i);
+    }
+}
+
+void llama_kv_cache_unified::set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const {
+    if (!supports_set_rows) {
+        return;
+    }
+
+    const uint32_t n_tokens = ubatch->n_tokens;
+
+    GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
+    int64_t * data = (int64_t *) dst->data;
+
+    for (int64_t i = 0; i < n_tokens; ++i) {
+        data[i] = sinfo.idxs.at(i);
+    }
+}
+
 void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const {
     const uint32_t n_tokens = ubatch->n_tokens;
 
@@ -1552,13 +1680,15 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell
             ubatch.seq_id[i]   = &dest_seq_id;
         }
 
-        const auto head_cur = find_slot(ubatch);
+        const auto sinfo = find_slot(ubatch, true);
-        if (head_cur < 0) {
+        if (sinfo.empty()) {
             LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__);
             return false;
         }
 
-        apply_ubatch(head_cur, ubatch);
+        apply_ubatch(sinfo, ubatch);
+
+        const auto head_cur = sinfo.head();
 
         // keep the head at the old position because we will read the KV data into it in state_read_data()
         head = head_cur;
@@ -1744,7 +1874,11 @@ llama_kv_cache_unified_context::llama_kv_cache_unified_context(llama_memory_stat
 llama_kv_cache_unified_context::llama_kv_cache_unified_context(
         llama_kv_cache_unified * kv) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv) {
     n_kv = kv->get_size();
-    head = 0;
+
+    // create a dummy slot info - the actual data is irrelevant. we just need to build the graph
+    sinfos.resize(1);
+    sinfos[0].idxs.resize(1);
+    sinfos[0].idxs[0] = 0;
 }
 
 llama_kv_cache_unified_context::llama_kv_cache_unified_context(
@@ -1759,8 +1893,8 @@ llama_kv_cache_unified_context::llama_kv_cache_unified_context(
 
 llama_kv_cache_unified_context::llama_kv_cache_unified_context(
         llama_kv_cache_unified * kv,
-        llama_kv_cache_unified::ubatch_heads heads,
+        llama_kv_cache_unified::slot_info_vec_t sinfos,
-        std::vector<llama_ubatch> ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), heads(std::move(heads)), ubatches(std::move(ubatches)) {
+        std::vector<llama_ubatch> ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), sinfos(std::move(sinfos)), ubatches(std::move(ubatches)) {
 }
 
 llama_kv_cache_unified_context::~llama_kv_cache_unified_context() = default;
@@ -1768,7 +1902,7 @@ llama_kv_cache_unified_context::~llama_kv_cache_unified_context() = default;
 bool llama_kv_cache_unified_context::next() {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    if (++i_next >= ubatches.size()) {
+    if (++i_cur >= ubatches.size()) {
         return false;
     }
 
@@ -1785,10 +1919,9 @@ bool llama_kv_cache_unified_context::apply() {
         return true;
     }
 
-    kv->apply_ubatch(heads[i_next], ubatches[i_next]);
+    kv->apply_ubatch(sinfos[i_cur], ubatches[i_cur]);
 
     n_kv = kv->get_n_kv();
-    head = heads[i_next];
 
     return true;
 }
@@ -1800,7 +1933,7 @@ llama_memory_status llama_kv_cache_unified_context::get_status() const {
 const llama_ubatch & llama_kv_cache_unified_context::get_ubatch() const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    return ubatches[i_next];
+    return ubatches[i_cur];
 }
 
 uint32_t llama_kv_cache_unified_context::get_n_kv() const {
@@ -1815,18 +1948,34 @@ ggml_tensor * llama_kv_cache_unified_context::get_v(ggml_context * ctx, int32_t
     return kv->get_v(ctx, il, n_kv);
 }
 
-ggml_tensor * llama_kv_cache_unified_context::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il) const {
+ggml_tensor * llama_kv_cache_unified_context::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il) const {
-    return kv->cpy_k(ctx, k_cur, il, head);
+    return kv->cpy_k(ctx, k_cur, k_idxs, il, sinfos[i_cur]);
 }
 
-ggml_tensor * llama_kv_cache_unified_context::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il) const {
+ggml_tensor * llama_kv_cache_unified_context::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il) const {
-    return kv->cpy_v(ctx, v_cur, il, head);
+    return kv->cpy_v(ctx, v_cur, v_idxs, il, sinfos[i_cur]);
+}
+
+ggml_tensor * llama_kv_cache_unified_context::build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
+    return kv->build_input_k_idxs(ctx, ubatch);
+}
+
+ggml_tensor * llama_kv_cache_unified_context::build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
+    return kv->build_input_v_idxs(ctx, ubatch);
 }
 
 void llama_kv_cache_unified_context::set_input_k_shift(ggml_tensor * dst) const {
     kv->set_input_k_shift(dst);
 }
 
+void llama_kv_cache_unified_context::set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const {
+    kv->set_input_k_idxs(dst, ubatch, sinfos[i_cur]);
+}
+
+void llama_kv_cache_unified_context::set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const {
+    kv->set_input_v_idxs(dst, ubatch, sinfos[i_cur]);
+}
+
 void llama_kv_cache_unified_context::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const {
     kv->set_input_kq_mask(dst, ubatch, causal_attn);
 }

src/llama-kv-cache-unified.h

@@ -24,8 +24,6 @@ public:
     // this callback is used to filter out layers that should not be included in the cache
     using layer_filter_cb = std::function<bool(int32_t il)>;
 
-    using ubatch_heads = std::vector<uint32_t>;
-
     struct defrag_info {
         bool empty() const {
             return ids.empty();
@@ -37,6 +35,32 @@ public:
         std::vector<uint32_t> ids;
     };
 
+    // for each ubatch, create a slot_info that contains information about where the ubatch should be inserted in the
+    //   KV cells. for example, cell indices for each token, such that: token[i] -> goes to cells[idxs[i]]
+    struct slot_info {
+        // data for ggml_set_rows
+        using idx_vec_t = std::vector<uint32_t>;
+
+        idx_vec_t idxs;
+
+        uint32_t head() const {
+            return idxs.at(0);
+        }
+
+        bool empty() const {
+            return idxs.empty();
+        }
+
+        void clear() {
+            idxs.clear();
+        }
+
+        // TODO: implement
+        //std::vector<idx_vec_t> seq_idxs;
+    };
+
+    using slot_info_vec_t = std::vector<slot_info>;
+
     llama_kv_cache_unified(
             const llama_model &  model,
               layer_filter_cb && filter,
@@ -102,30 +126,37 @@ public:
     ggml_tensor * get_v(ggml_context * ctx, int32_t il, uint32_t n_kv) const;
 
     // store k_cur and v_cur in the cache based on the provided head location
-    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il, uint32_t head_cur) const;
+    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const;
-    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il, uint32_t head_cur) const;
+    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il, const slot_info & sinfo) const;
 
     //
     // preparation API
     //
 
-    // find places for the provided ubatches in the cache, returns the head locations
+    // find places for the provided ubatches in the cache, returns the slot infos
     // return empty vector on failure
-    ubatch_heads prepare(const std::vector<llama_ubatch> & ubatches);
+    slot_info_vec_t prepare(const std::vector<llama_ubatch> & ubatches);
 
     bool update(llama_context * lctx, bool do_shift, const defrag_info & dinfo);
 
-    // return the cell position where we can insert the ubatch
+    // find a slot of kv cells that can hold the ubatch
-    // return -1 on failure to find a contiguous slot of kv cells
+    // if cont == true, then the slot must be continuous
-    int32_t find_slot(const llama_ubatch & ubatch) const;
+    // return empty slot_info on failure
+    slot_info find_slot(const llama_ubatch & ubatch, bool cont) const;
 
-    // emplace the ubatch context into slot: [head_cur, head_cur + ubatch.n_tokens)
+    // emplace the ubatch context into slot: [sinfo.idxs[0...ubatch.n_tokens - 1]]
-    void apply_ubatch(uint32_t head_cur, const llama_ubatch & ubatch);
+    void apply_ubatch(const slot_info & sinfo, const llama_ubatch & ubatch);
 
     //
-    // set_input API
+    // input API
     //
 
+    ggml_tensor * build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
+    ggml_tensor * build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
+
+    void set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const;
+    void set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const;
+
     void set_input_kq_mask   (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const;
     void set_input_k_shift   (ggml_tensor * dst) const;
     void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const;
@@ -157,8 +188,13 @@ private:
     // SWA
     const uint32_t n_swa = 0;
 
+    // env: LLAMA_KV_CACHE_DEBUG
     int debug = 0;
 
+    // env: LLAMA_SET_ROWS (temporary)
+    // ref: https://github.com/ggml-org/llama.cpp/pull/14285
+    int supports_set_rows = false;
+
     const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
 
     std::vector<ggml_context_ptr>        ctxs;
@@ -211,7 +247,7 @@ private:
 class llama_kv_cache_unified_context : public llama_memory_context_i {
 public:
     // some shorthands
-    using ubatch_heads = llama_kv_cache_unified::ubatch_heads;
+    using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t;
     using defrag_info     = llama_kv_cache_unified::defrag_info;
 
     // used for errors
@@ -231,7 +267,7 @@ public:
     // used to create a batch procesing context from a batch
     llama_kv_cache_unified_context(
             llama_kv_cache_unified * kv,
-            ubatch_heads heads,
+            slot_info_vec_t sinfos,
             std::vector<llama_ubatch> ubatches);
 
     virtual ~llama_kv_cache_unified_context();
@@ -257,11 +293,16 @@ public:
     ggml_tensor * get_v(ggml_context * ctx, int32_t il) const;
 
     // store k_cur and v_cur in the cache based on the provided head location
-    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il) const;
+    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il) const;
-    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il) const;
+    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il) const;
 
-    void set_input_k_shift(ggml_tensor * dst) const;
+    ggml_tensor * build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
+    ggml_tensor * build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
 
+    void set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const;
+    void set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const;
+
+    void set_input_k_shift   (ggml_tensor * dst) const;
     void set_input_kq_mask   (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const;
     void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const;
 
@@ -283,10 +324,10 @@ private:
     // batch processing context
     //
 
-    // the index of the next ubatch to process
+    // the index of the cur ubatch to process
-    size_t i_next = 0;
+    size_t i_cur = 0;
 
-    ubatch_heads heads;
+    slot_info_vec_t sinfos;
 
     std::vector<llama_ubatch> ubatches;
 
@@ -297,7 +338,4 @@ private:
     // a heuristic, to avoid attending the full cache if it is not yet utilized
     // as the cache gets filled, the benefit from this heuristic disappears
     int32_t n_kv;
-
-    // the beginning of the current slot in which the ubatch will be inserted
-    int32_t head;
 };

src/llama-kv-cells.h

@@ -105,10 +105,30 @@ public:
         res.resize(n);
 
         for (uint32_t j = 0; j < n; ++j) {
-            res.pos[j] = pos[i + j];
+            const auto idx = i + j;
-            res.seq[j] = seq[i + j];
 
-            assert(shift[i + j] == 0);
+            res.pos[j] = pos[idx];
+            res.seq[j] = seq[idx];
+
+            assert(shift[idx] == 0);
+        }
+
+        return res;
+    }
+
+    // copy the state of cells [idxs[0], idxs[1], ..., idxs[idxs.size() - 1])
+    llama_kv_cells_unified cp(const std::vector<uint32_t> & idxs) const {
+        llama_kv_cells_unified res;
+
+        res.resize(idxs.size());
+
+        for (uint32_t j = 0; j < idxs.size(); ++j) {
+            const auto idx = idxs[j];
+
+            res.pos[j] = pos[idx];
+            res.seq[j] = seq[idx];
+
+            assert(shift[idx] == 0);
         }
 
         return res;
@@ -119,26 +139,58 @@ public:
         assert(i + other.pos.size() <= pos.size());
 
         for (uint32_t j = 0; j < other.pos.size(); ++j) {
-            if (pos[i + j] == -1 && other.pos[j] != -1) {
+            const auto idx = i + j;
+
+            if (pos[idx] == -1 && other.pos[j] != -1) {
                 used.insert(i + j);
             }
 
-            if (pos[i + j] != -1 && other.pos[j] == -1) {
+            if (pos[idx] != -1 && other.pos[j] == -1) {
                 used.erase(i + j);
             }
 
-            if (pos[i + j] != -1) {
+            if (pos[idx] != -1) {
                 seq_pos_rm(i + j);
             }
 
-            pos[i + j] = other.pos[j];
+            pos[idx] = other.pos[j];
-            seq[i + j] = other.seq[j];
+            seq[idx] = other.seq[j];
 
-            if (pos[i + j] != -1) {
+            if (pos[idx] != -1) {
                 seq_pos_add(i + j);
             }
 
-            assert(shift[i + j] == 0);
+            assert(shift[idx] == 0);
+        }
+    }
+
+    // set the state of cells [idxs[0], idxs[1], ..., idxs[idxs.size() - 1])
+    void set(const std::vector<uint32_t> & idxs, const llama_kv_cells_unified & other) {
+        assert(idxs.size() == other.pos.size());
+
+        for (uint32_t j = 0; j < other.pos.size(); ++j) {
+            const auto idx = idxs[j];
+
+            if (pos[idx] == -1 && other.pos[j] != -1) {
+                used.insert(idx);
+            }
+
+            if (pos[idx] != -1 && other.pos[j] == -1) {
+                used.erase(idx);
+            }
+
+            if (pos[idx] != -1) {
+                seq_pos_rm(idx);
+            }
+
+            pos[idx] = other.pos[j];
+            seq[idx] = other.seq[j];
+
+            if (pos[idx] != -1) {
+                seq_pos_add(idx);
+            }
+
+            assert(shift[idx] == 0);
         }
     }
 

src/llama-memory-hybrid.cpp

@@ -70,7 +70,7 @@ llama_memory_context_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & ba
                 // if all tokens are output, split by sequence
                 ubatch = balloc.split_seq(n_ubatch);
             } else {
-                ubatch = balloc.split_equal(n_ubatch);
+                ubatch = balloc.split_equal(n_ubatch, false);
             }
 
             if (ubatch.n_tokens == 0) {
@@ -80,6 +80,11 @@ llama_memory_context_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & ba
             ubatches.push_back(std::move(ubatch)); // NOLINT
         }
 
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
+            break;
+        }
+
         // prepare the recurrent batches first
         if (!mem_recr->prepare(ubatches)) {
             // TODO: will the recurrent cache be in an undefined context at this point?
@@ -195,11 +200,11 @@ llama_memory_hybrid_context::llama_memory_hybrid_context(
 
 llama_memory_hybrid_context::llama_memory_hybrid_context(
               llama_memory_hybrid * mem,
-            std::vector<uint32_t>   heads_attn,
+                  slot_info_vec_t   sinfos_attn,
         std::vector<llama_ubatch>   ubatches) :
     ubatches(std::move(ubatches)),
     // note: here we copy the ubatches. not sure if this is ideal
-    ctx_attn(new llama_kv_cache_unified_context(mem->get_mem_attn(), std::move(heads_attn), this->ubatches)),
+    ctx_attn(new llama_kv_cache_unified_context(mem->get_mem_attn(), std::move(sinfos_attn), this->ubatches)),
     ctx_recr(new llama_memory_recurrent_context(mem->get_mem_recr(),                        this->ubatches)),
     status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
 }

src/llama-memory-hybrid.h

@@ -92,6 +92,8 @@ private:
 
 class llama_memory_hybrid_context : public llama_memory_context_i {
 public:
+    using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t;
+
     // init failure
     explicit llama_memory_hybrid_context(llama_memory_status status);
 
@@ -107,7 +109,7 @@ public:
     // init success
     llama_memory_hybrid_context(
               llama_memory_hybrid * mem,
-            std::vector<uint32_t>   heads_attn,
+                  slot_info_vec_t   sinfos_attn,
         std::vector<llama_ubatch>   ubatches);
 
     ~llama_memory_hybrid_context() = default;

src/llama-memory-recurrent.cpp

@@ -374,10 +374,11 @@ llama_memory_context_ptr llama_memory_recurrent::init_batch(llama_batch_allocr &
                 // if all tokens are output, split by sequence
                 ubatch = balloc.split_seq(n_ubatch);
             } else {
-                ubatch = balloc.split_equal(n_ubatch);
+                ubatch = balloc.split_equal(n_ubatch, false);
             }
 
-            if (ubatch.n_tokens == 0) {
+            if (balloc.get_n_used() < balloc.get_n_tokens()) {
+                // failed to find a suitable split
                 break;
             }
 

src/llama-model.cpp

@@ -213,23 +213,27 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
             } break;
         case GGML_OP_SSM_CONV:
             {
-                // FIXME
+                const int64_t n_seq_tokens = 512;
-                ggml_tensor * conv_x = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, 12345, w->ne[1], 6789);
+                const int64_t n_seqs       = 3;
+                ggml_tensor * conv_x = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, w->ne[0] - 1 + n_seq_tokens, w->ne[1], n_seqs);
                 op_tensor = ggml_ssm_conv(ctx, conv_x, w);
             } break;
         case GGML_OP_SSM_SCAN:
             {
-                // FIXME
+                // w is ssm_a, which is used to distinguish Mamba-1 and Mamba-2
-                const int64_t d_state      = w->ne[0];
+                const int64_t d_state      = w->ne[0] == 1 ? hparams.ssm_d_state : w->ne[0];
-                const int64_t d_inner      = w->ne[1];
+                const int64_t n_head       = w->ne[1];
+                const int64_t head_dim     = hparams.ssm_d_inner / n_head;
+                const int64_t n_group      = hparams.ssm_n_group ? hparams.ssm_n_group : 1;
                 const int64_t n_seq_tokens = 512;
-                const int64_t n_seqs       = 1;
+                const int64_t n_seqs       = 3;
-                ggml_tensor * s  = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_state, d_inner, n_seqs);
+                ggml_tensor * s   = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, d_state, head_dim, n_head, n_seqs);
-                ggml_tensor * x = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_inner, n_seq_tokens, n_seqs);
+                ggml_tensor * x   = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, head_dim, n_head, n_seq_tokens, n_seqs);
-                ggml_tensor * dt = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_inner, n_seq_tokens, n_seqs);
+                ggml_tensor * dt  = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_head, n_seq_tokens, n_seqs);
-                ggml_tensor * B = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_state, n_seq_tokens, n_seqs);
+                ggml_tensor * B   = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, d_state, n_group, n_seq_tokens, n_seqs);
-                ggml_tensor * C = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_state, n_seq_tokens, n_seqs);
+                ggml_tensor * C   = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, d_state, n_group, n_seq_tokens, n_seqs);
-                op_tensor = ggml_ssm_scan(ctx, s, x, dt, w, B, C);
+                ggml_tensor * ids = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n_seqs);
+                op_tensor = ggml_ssm_scan(ctx, s, x, dt, w, B, C, ids);
             } break;
         case GGML_OP_RWKV_WKV6:
             {
@@ -1086,6 +1090,38 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_MAMBA2:
+            {
+                ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
+                ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
+                ml.get_key(LLM_KV_SSM_STATE_SIZE,     hparams.ssm_d_state);
+                ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
+                ml.get_key(LLM_KV_SSM_GROUP_COUNT,    hparams.ssm_n_group);
+
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                switch (hparams.n_layer) {
+                    case 24:
+                        switch (hparams.n_embd) {
+                            case 768: type = LLM_TYPE_SMALL; break;
+                            default: type = LLM_TYPE_UNKNOWN;
+                        } break;
+                    case 48:
+                        switch (hparams.n_embd) {
+                            case 1024: type = LLM_TYPE_MEDIUM; break;
+                            case 1536: type = LLM_TYPE_LARGE; break;
+                            case 2048: type = LLM_TYPE_XL; break;
+                            default: type = LLM_TYPE_UNKNOWN;
+                        } break;
+                    case 64:
+                        switch (hparams.n_embd) {
+                            case 2560: type = LLM_TYPE_3B; break;
+                            case 4096: type = LLM_TYPE_7B; break;
+                            default: type = LLM_TYPE_UNKNOWN;
+                        } break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_XVERSE:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -3216,6 +3252,54 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {d_state, d_inner}, 0);
                         layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {d_inner}, 0);
 
+                        // out_proj
+                        layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);
+                    }
+                } break;
+            case LLM_ARCH_MAMBA2:
+                {
+                    const int64_t d_conv  = hparams.ssm_d_conv;
+                    const int64_t d_inner = hparams.ssm_d_inner;
+                    const int64_t d_state = hparams.ssm_d_state;
+                    const int64_t n_head  = hparams.ssm_dt_rank;
+                    const int64_t n_group = hparams.ssm_n_group;
+                    const int64_t d_in_proj = 2*d_inner + 2*n_group*d_state + n_head;
+
+                    // only an expansion factor of 2 is supported for now
+                    GGML_ASSERT(2 * n_embd == d_inner);
+
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    {
+                        output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+
+                        output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        // if output is NULL, init from the input tok embed, duplicated to allow offloading
+                        if (output == NULL) {
+                            output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        // norm
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, 0);
+
+                        layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner + 2*n_group*d_state}, 0);
+                        layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner + 2*n_group*d_state}, 0);
+
+                        layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {n_head}, 0);
+
+                        // no "weight" suffix for these
+                        layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_head}, 0);
+                        layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, n_head}, 0);
+
+                        layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {d_inner / n_group, n_group}, 0);
+
                         // out_proj
                         layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);
                     }
@@ -4727,10 +4811,14 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: freq_scale_train = %g\n",     __func__, hparams.rope_freq_scale_train);
         LLAMA_LOG_INFO("%s: n_ctx_orig_yarn  = %u\n",     __func__, hparams.n_ctx_orig_yarn);
         LLAMA_LOG_INFO("%s: rope_finetuned   = %s\n",     __func__, hparams.rope_finetuned ? "yes" : "unknown");
+    }
+
+    if (arch == LLM_ARCH_MAMBA || arch == LLM_ARCH_MAMBA2) {
         LLAMA_LOG_INFO("%s: ssm_d_conv       = %u\n",     __func__, hparams.ssm_d_conv);
         LLAMA_LOG_INFO("%s: ssm_d_inner      = %u\n",     __func__, hparams.ssm_d_inner);
         LLAMA_LOG_INFO("%s: ssm_d_state      = %u\n",     __func__, hparams.ssm_d_state);
         LLAMA_LOG_INFO("%s: ssm_dt_rank      = %u\n",     __func__, hparams.ssm_dt_rank);
+        LLAMA_LOG_INFO("%s: ssm_n_group      = %u\n",     __func__, hparams.ssm_n_group);
         LLAMA_LOG_INFO("%s: ssm_dt_b_c_rms   = %d\n",     __func__, hparams.ssm_dt_b_c_rms);
 
         if (!classifier_labels.empty()) {
@@ -9765,9 +9853,7 @@ struct llm_build_starcoder2 : public llm_graph_context {
 };
 
 struct llm_build_mamba : public llm_graph_context {
-    const llama_model & model;
+    llm_build_mamba(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
-
-    llm_build_mamba(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params), model(model) {
         ggml_tensor * cur;
         ggml_tensor * inpL;
 
@@ -9785,7 +9871,11 @@ struct llm_build_mamba : public llm_graph_context {
                     LLM_NORM_RMS, il);
             cb(cur, "attn_norm", il);
 
-            cur = build_mamba_layer(rs_inp, gf, cur, ubatch, il);
+            if (model.arch == LLM_ARCH_MAMBA2) {
+                cur = build_mamba2_layer(rs_inp, gf, cur, model, ubatch, il);
+            } else {
+                cur = build_mamba_layer(rs_inp, gf, cur, model, ubatch, il);
+            }
 
             if (il == n_layer - 1 && inp_out_ids) {
                 cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
@@ -9819,11 +9909,11 @@ struct llm_build_mamba : public llm_graph_context {
         ggml_build_forward_expand(gf, cur);
     }
 
-    // TODO: split
     ggml_tensor * build_mamba_layer(
         llm_graph_input_rs * inp,
                ggml_cgraph * gf,
                ggml_tensor * cur,
+         const llama_model & model,
         const llama_ubatch & ubatch,
                        int   il) const {
         const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
@@ -9834,6 +9924,8 @@ struct llm_build_mamba : public llm_graph_context {
         const int64_t d_inner = hparams.ssm_d_inner;
         const int64_t d_state = hparams.ssm_d_state;
         const int64_t dt_rank = hparams.ssm_dt_rank;
+        const int64_t n_head  = d_inner;
+        const int64_t head_dim = 1;
         const int64_t n_seqs  = ubatch.n_seqs;
         // Some variants of Mamba arch (e.g. FalconMamba do apply layer norm on B and Dt layers)
         const bool ssm_dt_b_c_rms = hparams.ssm_dt_b_c_rms;
@@ -9849,15 +9941,8 @@ struct llm_build_mamba : public llm_graph_context {
         ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
         ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);
 
-        // (ab)using the KV cache to store the states
+        ggml_tensor * conv = build_rs(inp, gf, conv_states_all, hparams.n_embd_r(), n_seqs);
-        ggml_tensor * conv = build_rs(
-                inp, gf, conv_states_all,
-                hparams.n_embd_r(), n_seqs);
         conv = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner, n_seqs);
-        ggml_tensor * ssm = build_rs(
-                inp, gf, ssm_states_all,
-                hparams.n_embd_s(), n_seqs);
-        ssm = ggml_reshape_3d(ctx0, ssm, d_state, d_inner, n_seqs);
 
         // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
         cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
@@ -9906,8 +9991,8 @@ struct llm_build_mamba : public llm_graph_context {
             ggml_tensor * x_db = build_lora_mm(model.layers[il].ssm_x, x);
             // split
             ggml_tensor * dt = ggml_view_3d(ctx0, x_db, dt_rank, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], 0);
-            ggml_tensor * B  = ggml_view_3d(ctx0, x_db, d_state, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*dt_rank);
+            ggml_tensor * B  = ggml_view_4d(ctx0, x_db, d_state, /* n_group */ 1, n_seq_tokens, n_seqs, d_state*x_db->nb[0], x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*dt_rank);
-            ggml_tensor * C  = ggml_view_3d(ctx0, x_db, d_state, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*(dt_rank+d_state));
+            ggml_tensor * C  = ggml_view_4d(ctx0, x_db, d_state, /* n_group */ 1, n_seq_tokens, n_seqs, d_state*x_db->nb[0], x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*(dt_rank+d_state));
 
             // Some Mamba variants (e.g. FalconMamba) apply RMS norm in B, C & Dt layers
             if (ssm_dt_b_c_rms) {
@@ -9920,23 +10005,36 @@ struct llm_build_mamba : public llm_graph_context {
             dt = build_lora_mm(model.layers[il].ssm_dt, dt);
             dt = ggml_add(ctx0, dt, model.layers[il].ssm_dt_b);
 
+            cur = x;
+            x = ggml_reshape_4d(ctx0, x, head_dim, n_head, n_seq_tokens, n_seqs);
+
+            ggml_tensor * A = model.layers[il].ssm_a;
+
+            // use the states and the indices provided by build_recurrent_state
+            // (this is necessary in order to properly use the states before they are overwritten,
+            //  while avoiding to make unnecessary copies of the states)
+            auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) {
+                ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_head, mctx_cur->get_size());
+
                 // Custom operator to optimize the parallel associative scan
                 // as described in the Annex D of the Mamba paper.
                 // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs}
-            ggml_tensor * y_ssm = ggml_ssm_scan(ctx0, ssm, x, dt, model.layers[il].ssm_a, B, C);
+                return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids);
+            };
+
+            ggml_tensor * y_ssm = build_rs(inp, gf, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows);
 
             // store last states
             ggml_build_forward_expand(gf,
                 ggml_cpy(ctx0,
-                    ggml_view_1d(ctx0, y_ssm, d_state*d_inner*n_seqs, x->nb[3]),
+                    ggml_view_1d(ctx0, y_ssm, d_state*d_inner*n_seqs, x->nb[3]*x->ne[3]),
                     ggml_view_1d(ctx0, ssm_states_all, d_state*d_inner*n_seqs, kv_head*d_state*d_inner*ggml_element_size(ssm_states_all))));
 
-            ggml_tensor * y = ggml_view_3d(ctx0, y_ssm, d_inner, n_seq_tokens, n_seqs, x->nb[1], x->nb[2], 0);
+            ggml_tensor * y = ggml_view_3d(ctx0, y_ssm, d_inner, n_seq_tokens, n_seqs, x->nb[2], x->nb[3], 0);
 
             // TODO: skip computing output earlier for unused tokens
 
-            // {d_inner, n_seq_tokens, n_seqs} * {d_inner} => {d_inner, n_seq_tokens, n_seqs}
+            y = ggml_add(ctx0, y, ggml_mul(ctx0, cur, model.layers[il].ssm_d));
-            y = ggml_add(ctx0, y, ggml_mul(ctx0, x, model.layers[il].ssm_d));
             y = ggml_mul(ctx0, y, ggml_silu(ctx0, ggml_cont(ctx0, z)));
 
             // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}
@@ -9945,7 +10043,136 @@ struct llm_build_mamba : public llm_graph_context {
 
         // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
         cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs);
-        //cb(cur, "mamba_out", il);
+        // cb(cur, "mamba_out", il);
+
+        return cur;
+    }
+
+    ggml_tensor * build_mamba2_layer(
+        llm_graph_input_rs * inp,
+             ggml_cgraph * gf,
+             ggml_tensor * cur,
+       const llama_model & model,
+      const llama_ubatch & ubatch,
+                     int   il) const {
+        const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
+
+        const auto kv_head = mctx_cur->get_head();
+
+        const int64_t d_conv  = hparams.ssm_d_conv;
+        const int64_t d_inner = hparams.ssm_d_inner;
+        const int64_t d_state = hparams.ssm_d_state;
+        const int64_t n_head  = hparams.ssm_dt_rank;
+        const int64_t head_dim = d_inner / n_head;
+        const int64_t n_group = hparams.ssm_n_group;
+        const int64_t n_seqs  = ubatch.n_seqs;
+
+        const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+
+        GGML_ASSERT(n_seqs != 0);
+        GGML_ASSERT(ubatch.equal_seqs);
+        GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
+
+        ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
+        ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);
+
+        ggml_tensor * conv = build_rs(inp, gf, conv_states_all, hparams.n_embd_r(), n_seqs);
+        conv = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner + 2*n_group*d_state, n_seqs);
+
+        // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
+        cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
+
+        // d_in_proj = 2 * self.d_inner + 2 * self.ngroups * self.d_state + self.nheads
+
+        // {n_embd, d_in_proj} @ {n_embd, n_seq_tokens, n_seqs} => {d_in_proj, n_seq_tokens, n_seqs}
+        ggml_tensor * zxBCdt = build_lora_mm(model.layers[il].ssm_in, cur);
+
+        // split the above in three
+        ggml_tensor * z = ggml_view_4d(ctx0, zxBCdt, head_dim, n_head, n_seq_tokens, n_seqs, head_dim*zxBCdt->nb[0], zxBCdt->nb[1], zxBCdt->nb[2], 0);
+        ggml_tensor * xBC = ggml_view_3d(ctx0, zxBCdt, d_inner + 2*n_group*d_state, n_seq_tokens, n_seqs, zxBCdt->nb[1], zxBCdt->nb[2], d_inner*ggml_element_size(zxBCdt));
+        ggml_tensor * dt = ggml_view_3d(ctx0, zxBCdt, n_head, n_seq_tokens, n_seqs, zxBCdt->nb[1], zxBCdt->nb[2], (2*d_inner + 2*n_group*d_state)*ggml_element_size(zxBCdt));
+
+        // conv
+        {
+            // => {d_conv - 1 + n_seq_tokens, d_inner + 2*n_group*d_state, n_seqs}
+            ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, xBC), 0);
+
+            // copy last (d_conv - 1) columns back into the state cache
+            ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner + 2*n_group*d_state, n_seqs, conv_x->nb[1], conv_x->nb[2], n_seq_tokens*(conv_x->nb[0]));
+
+            ggml_build_forward_expand(gf,
+                ggml_cpy(ctx0, last_conv,
+                    ggml_view_1d(ctx0, conv_states_all,
+                        (d_conv - 1)*(d_inner + 2*n_group*d_state)*(n_seqs),
+                        kv_head*(d_conv - 1)*(d_inner + 2*n_group*d_state)*ggml_element_size(conv_states_all))));
+
+            // 1D convolution
+            // The equivalent is to make a self-overlapping view of conv_x
+            // over d_conv columns at each stride in the 3rd dimension,
+            // then element-wise multiply that with the conv1d weight,
+            // then sum the elements of each row,
+            // (the last two steps are a dot product over rows (also doable with mul_mat))
+            // then permute away the ne[0] dimension,
+            // and then you're left with the resulting x tensor.
+            // For simultaneous sequences, all sequences need to have the same length.
+            xBC = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d);
+
+            // bias
+            xBC = ggml_add(ctx0, xBC, model.layers[il].ssm_conv1d_b);
+
+            xBC = ggml_silu(ctx0, xBC);
+        }
+
+        // ssm
+        {
+            // These correspond to V K Q in SSM/attention duality
+            ggml_tensor * x = ggml_view_4d(ctx0, xBC, head_dim, n_head, n_seq_tokens, n_seqs, head_dim*xBC->nb[0], xBC->nb[1], xBC->nb[2], 0);
+            ggml_tensor * B = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state*xBC->nb[0], xBC->nb[1], xBC->nb[2], d_inner*ggml_element_size(xBC));
+            ggml_tensor * C = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state*xBC->nb[0], xBC->nb[1], xBC->nb[2], (d_inner + n_group*d_state)*ggml_element_size(xBC));
+
+            // {n_head, n_seq_tokens, n_seqs}
+            dt = ggml_add(ctx0, ggml_cont(ctx0, dt), model.layers[il].ssm_dt_b);
+
+            ggml_tensor * A = model.layers[il].ssm_a;
+
+            // use the states and the indices provided by build_recurrent_state
+            // (this is necessary in order to properly use the states before they are overwritten,
+            //  while avoiding to make unnecessary copies of the states)
+            auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) {
+                ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_head, mctx_cur->get_size());
+
+                // TODO: use semistructured matrices to implement state-space duality
+                // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs}
+                return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids);
+            };
+
+            ggml_tensor * y_ssm = build_rs(inp, gf, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows);
+
+            // store last states
+            ggml_build_forward_expand(gf,
+                ggml_cpy(ctx0,
+                    ggml_view_1d(ctx0, y_ssm, d_state*d_inner*n_seqs, ggml_nelements(x)*x->nb[0]),
+                    ggml_view_1d(ctx0, ssm_states_all, d_state*d_inner*n_seqs, kv_head*d_state*d_inner*ggml_element_size(ssm_states_all))));
+
+            ggml_tensor * y = ggml_view_4d(ctx0, y_ssm, head_dim, n_head, n_seq_tokens, n_seqs, x->nb[1], n_head*x->nb[1], n_seq_tokens*n_head*x->nb[1], 0);
+
+            // TODO: skip computing output earlier for unused tokens
+
+            y = ggml_add(ctx0, y, ggml_mul(ctx0, x, model.layers[il].ssm_d));
+            y = ggml_mul(ctx0, y, ggml_silu(ctx0, ggml_cont(ctx0, z)));
+
+            // grouped RMS norm
+            y = ggml_reshape_4d(ctx0, y, d_inner / n_group, n_group, n_seq_tokens, n_seqs);
+            y = build_norm(y, model.layers[il].ssm_norm, NULL, LLM_NORM_RMS, il);
+            y = ggml_reshape_3d(ctx0, y, d_inner, n_seq_tokens, n_seqs);
+
+            // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}
+            cur = build_lora_mm(model.layers[il].ssm_out, y);
+        }
+
+        // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
+        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs);
+        // cb(cur, "mamba_out", il);
 
         return cur;
     }
@@ -14768,6 +14995,7 @@ llm_graph_result_ptr llama_model::build_graph(
                 llm = std::make_unique<llm_build_starcoder2>(*this, params, gf);
             } break;
         case LLM_ARCH_MAMBA:
+        case LLM_ARCH_MAMBA2:
             {
                 llm = std::make_unique<llm_build_mamba>(*this, params, gf);
             } break;
@@ -15028,6 +15256,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_REFACT:
         case LLM_ARCH_BLOOM:
         case LLM_ARCH_MAMBA:
+        case LLM_ARCH_MAMBA2:
         case LLM_ARCH_JINA_BERT_V2:
         case LLM_ARCH_T5:
         case LLM_ARCH_T5ENCODER:

src/llama-model.h

@@ -172,6 +172,7 @@ struct llama_layer {
     struct ggml_tensor * ffn_sub_norm    = nullptr;
     struct ggml_tensor * attn_norm_cross = nullptr;
     struct ggml_tensor * attn_norm_enc   = nullptr;
+    struct ggml_tensor * ssm_norm        = nullptr;
 
     // attention
     struct ggml_tensor * wq        = nullptr;

tools/mtmd/clip.cpp

@@ -1430,8 +1430,7 @@ struct clip_graph {
                 ggml_tensor * x = embeddings;
                 embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, embeddings);
                 x = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w,x);
-                embeddings = ggml_silu_inplace(ctx0, embeddings);
+                embeddings = ggml_swiglu_split(ctx0, embeddings, x);
-                embeddings = ggml_mul(ctx0, embeddings,x);
                 embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, embeddings);
             }
             // arrangement of BOI/EOI token embeddings
@@ -1527,15 +1526,8 @@ struct clip_graph {
                 cur = ggml_mul_mat(ctx0, model.mm_1_w, cur);
 
                 // swiglu
-                {
-                    int64_t split_point = cur->ne[0] / 2;
-                    ggml_tensor * x0 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], 0));
-                    ggml_tensor * x1 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], split_point * ggml_element_size(cur)));
-
                 // see SwiGLU in ultravox_model.py, the second half passed through is silu, not the first half
-                    x1 = ggml_silu(ctx0, x1);
+                cur = ggml_swiglu_swapped(ctx0, cur);
-                    cur = ggml_mul(ctx0, x0, x1);
-                }
 
                 // mid-norm
                 cur = ggml_rms_norm(ctx0, cur, 1e-6);
@@ -1794,33 +1786,40 @@ private:
             cur = tmp;
         }
 
+        // we only support parallel ffn for now
         switch (type_op) {
             case FFN_SILU:
-                {
+                if (gate) {
+                    cur = ggml_swiglu_split(ctx0, cur, tmp);
+                    cb(cur, "ffn_swiglu", il);
+                } else {
                     cur = ggml_silu(ctx0, cur);
                     cb(cur, "ffn_silu", il);
                 } break;
             case FFN_GELU:
-                {
+                if (gate) {
+                    cur = ggml_geglu_split(ctx0, cur, tmp);
+                    cb(cur, "ffn_geglu", il);
+                } else {
                     cur = ggml_gelu(ctx0, cur);
                     cb(cur, "ffn_gelu", il);
                 } break;
             case FFN_GELU_ERF:
-                {
+                if (gate) {
+                    cur = ggml_geglu_erf_split(ctx0, cur, tmp);
+                    cb(cur, "ffn_geglu_erf", il);
+                } else {
                     cur = ggml_gelu_erf(ctx0, cur);
-                    cb(cur, "ggml_gelu_erf", il);
+                    cb(cur, "ffn_gelu_erf", il);
                 } break;
             case FFN_GELU_QUICK:
-                {
-                    cur = ggml_gelu_quick(ctx0, cur);
-                    cb(cur, "ffn_relu", il);
-                } break;
-        }
-
-        // we only support parallel ffn for now
                 if (gate) {
-            cur = ggml_mul(ctx0, cur, tmp);
+                    cur = ggml_geglu_quick_split(ctx0, cur, tmp);
-            cb(cur, "ffn_gate_par", il);
+                    cb(cur, "ffn_geglu_quick", il);
+                } else {
+                    cur = ggml_gelu_quick(ctx0, cur);
+                    cb(cur, "ffn_gelu_quick", il);
+                } break;
         }
 
         if (down) {