mirror of
https://github.com/LostRuins/koboldcpp.git
synced 2026-05-07 17:22:04 +00:00
unstable merge
This commit is contained in:
Concedo
commit
82d562ad7b
34 changed files with 1025 additions and 371 deletions
|
|
@ -3520,15 +3520,15 @@ void common_params_add_preset_options(std::vector<common_arg> & args) {
|
|||
[](common_params &, const std::string &) { /* unused */ }
|
||||
).set_env(COMMON_ARG_PRESET_LOAD_ON_STARTUP).set_preset_only());
|
||||
|
||||
args.push_back(common_arg(
|
||||
{"stop-timeout"}, "SECONDS",
|
||||
"in server router mode, force-kill model instance after this many seconds of graceful shutdown",
|
||||
[](common_params &, int) { /* unused */ }
|
||||
).set_env(COMMON_ARG_PRESET_STOP_TIMEOUT).set_preset_only());
|
||||
|
||||
// args.push_back(common_arg(
|
||||
// {"pin"},
|
||||
// "in server router mode, do not unload this model if models_max is exceeded",
|
||||
// [](common_params &) { /* unused */ }
|
||||
// ).set_preset_only());
|
||||
|
||||
// args.push_back(common_arg(
|
||||
// {"unload-idle-seconds"}, "SECONDS",
|
||||
// "in server router mode, unload models idle for more than this many seconds",
|
||||
// [](common_params &, int) { /* unused */ }
|
||||
// ).set_preset_only());
|
||||
}
|
||||
|
|
|
|||
|
|
@ -10,6 +10,7 @@
|
|||
|
||||
// pseudo-env variable to identify preset-only arguments
|
||||
#define COMMON_ARG_PRESET_LOAD_ON_STARTUP "__PRESET_LOAD_ON_STARTUP"
|
||||
#define COMMON_ARG_PRESET_STOP_TIMEOUT "__PRESET_STOP_TIMEOUT"
|
||||
|
||||
//
|
||||
// CLI argument parsing
|
||||
|
|
|
|||
|
|
@ -7362,6 +7362,90 @@ class MiniMaxM2Model(TextModel):
|
|||
return super().modify_tensors(data_torch, name, bid)
|
||||
|
||||
|
||||
@ModelBase.register("MiMoV2FlashForCausalLM")
|
||||
class MimoV2Model(TextModel):
|
||||
model_arch = gguf.MODEL_ARCH.MIMO2
|
||||
|
||||
def set_gguf_parameters(self):
|
||||
super().set_gguf_parameters()
|
||||
|
||||
assert self.hparams["swa_head_dim"] == self.hparams["head_dim"]
|
||||
assert self.hparams["swa_num_attention_heads"] == self.hparams["num_attention_heads"]
|
||||
assert self.hparams["swa_v_head_dim"] == self.hparams["v_head_dim"]
|
||||
assert self.hparams["topk_method"] == "noaux_tc"
|
||||
|
||||
n_head_kv = self.hparams["num_key_value_heads"]
|
||||
n_head_kv_swa = self.hparams["swa_num_key_value_heads"]
|
||||
n_head_kv_arr = [n_head_kv_swa if use_swa == 1 else n_head_kv for use_swa in self.hparams["hybrid_layer_pattern"]]
|
||||
self.gguf_writer.add_head_count_kv(n_head_kv_arr)
|
||||
|
||||
self.gguf_writer.add_sliding_window(self.hparams["sliding_window"])
|
||||
self.gguf_writer.add_sliding_window_pattern(self.hparams["hybrid_layer_pattern"])
|
||||
self.gguf_writer.add_rope_freq_base_swa(self.hparams["swa_rope_theta"])
|
||||
self.gguf_writer.add_value_length(self.hparams["v_head_dim"])
|
||||
self.gguf_writer.add_expert_count(self.hparams["n_routed_experts"])
|
||||
self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
|
||||
|
||||
rope_dim = int(self.hparams["head_dim"] * self.hparams["partial_rotary_factor"])
|
||||
self.gguf_writer.add_rope_dimension_count(rope_dim)
|
||||
|
||||
self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("layernorm_epsilon", 1e-5))
|
||||
|
||||
_experts: list[dict[str, Tensor]] | None = None
|
||||
|
||||
def modify_tensors(self, data_torch, name, bid):
|
||||
if name.endswith("e_score_correction_bias"):
|
||||
name = name.replace("e_score_correction_bias", "e_score_correction.bias")
|
||||
|
||||
if "attention_sink" in name and not name.endswith(".weight"):
|
||||
name += ".weight"
|
||||
|
||||
# TODO: mimo v2 does not indicate the number of next-token-prediction layers, therefore we cannot do the same way as GLM4_MOE
|
||||
if "model.mtp." in name:
|
||||
return []
|
||||
|
||||
# process the experts separately
|
||||
if name.find("mlp.experts") != -1:
|
||||
n_experts = self.hparams["n_routed_experts"]
|
||||
assert bid is not None
|
||||
|
||||
if self._experts is None:
|
||||
self._experts = [{} for _ in range(self.block_count)]
|
||||
|
||||
self._experts[bid][name] = data_torch
|
||||
|
||||
if len(self._experts[bid]) >= n_experts * 3:
|
||||
tensors: list[tuple[str, Tensor]] = []
|
||||
|
||||
# merge the experts into a single 3d tensor
|
||||
for w_name in ["gate_proj", "up_proj", "down_proj"]:
|
||||
datas: list[Tensor] = []
|
||||
|
||||
for xid in range(n_experts):
|
||||
ename_to_retrieve = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
|
||||
datas.append(self._experts[bid][ename_to_retrieve])
|
||||
del self._experts[bid][ename_to_retrieve]
|
||||
|
||||
data_torch = torch.stack(datas, dim=0)
|
||||
merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
|
||||
new_name = self.map_tensor_name(merged_name)
|
||||
tensors.append((new_name, data_torch))
|
||||
|
||||
return tensors
|
||||
else:
|
||||
return []
|
||||
return [(self.map_tensor_name(name), data_torch)]
|
||||
|
||||
def prepare_tensors(self):
|
||||
super().prepare_tensors()
|
||||
|
||||
if self._experts is not None:
|
||||
# flatten `list[dict[str, Tensor]]` into `list[str]`
|
||||
experts = [k for d in self._experts for k in d.keys()]
|
||||
if len(experts) > 0:
|
||||
raise ValueError(f"Unprocessed experts: {experts}")
|
||||
|
||||
|
||||
@ModelBase.register("PanguEmbeddedForCausalLM")
|
||||
class PanguEmbeddedModel(TextModel):
|
||||
model_arch = gguf.MODEL_ARCH.PANGU_EMBED
|
||||
|
|
|
|||
|
|
@ -19343,7 +19343,17 @@ Current version indicated by LITEVER below.
|
|||
|
||||
for(let i=0;i<mhistory.length;++i)
|
||||
{
|
||||
oai_payload.messages.push({ "role": (mhistory[i].myturn?"user":"assistant"), "content": mhistory[i].msg });
|
||||
let cturn = { "role": (mhistory[i].myturn?"user":"assistant"), "content": mhistory[i].msg };
|
||||
if(!mhistory[i].myturn && mhistory[i].tool_calls)
|
||||
{
|
||||
cturn.content = null;
|
||||
cturn.tool_calls = mhistory[i].tool_calls;
|
||||
}
|
||||
else if(!mhistory[i].myturn && mhistory[i].tool_results)
|
||||
{
|
||||
cturn = mhistory[i].tool_results;
|
||||
}
|
||||
oai_payload.messages.push(cturn);
|
||||
}
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -5,7 +5,7 @@
|
|||
#include "ggml.h"
|
||||
|
||||
#ifdef GGML_CUDA_USE_CUB
|
||||
# include <cub/device/device_scan.cuh>
|
||||
# include <cub/block/block_scan.cuh>
|
||||
#endif // GGML_CUDA_USE_CUB
|
||||
|
||||
template<typename T, int BLOCK_SIZE>
|
||||
|
|
@ -16,12 +16,14 @@ static __global__ void cumsum_cub_kernel(
|
|||
const int64_t s01, const int64_t s02, const int64_t s03,
|
||||
const int64_t s1, const int64_t s2, const int64_t s3) {
|
||||
#ifdef GGML_CUDA_USE_CUB
|
||||
using BlockScan = cub::BlockScan<T, BLOCK_SIZE>;
|
||||
using BlockScanT = cub::BlockScan<T, BLOCK_SIZE>;
|
||||
|
||||
__shared__ typename BlockScan::TempStorage temp_storage;
|
||||
__shared__ T block_carry; // carry from previous tile
|
||||
__shared__ typename BlockScanT::TempStorage temp_storage;
|
||||
__shared__ T block_carry;
|
||||
|
||||
const int tid = threadIdx.x;
|
||||
constexpr int UNROLL_FACTOR = 4;
|
||||
constexpr int TILE_SIZE = BLOCK_SIZE * UNROLL_FACTOR;
|
||||
|
||||
const int64_t i1 = blockIdx.x;
|
||||
const int64_t i2 = blockIdx.y;
|
||||
|
|
@ -39,29 +41,38 @@ static __global__ void cumsum_cub_kernel(
|
|||
}
|
||||
__syncthreads();
|
||||
|
||||
for (int64_t start = 0; start < ne00; start += BLOCK_SIZE) {
|
||||
int64_t idx = start + tid;
|
||||
T x = (idx < ne00) ? src_row[idx] : T(0);
|
||||
for (int64_t start = 0; start < ne00; start += TILE_SIZE) {
|
||||
T items[UNROLL_FACTOR];
|
||||
T thread_sum = T(0);
|
||||
|
||||
T inclusive;
|
||||
T block_total;
|
||||
BlockScan(temp_storage).InclusiveSum(x, inclusive, block_total);
|
||||
|
||||
__syncthreads();
|
||||
|
||||
T final_val = inclusive + block_carry;
|
||||
|
||||
// store result
|
||||
if (idx < ne00) {
|
||||
dst_row[idx] = final_val;
|
||||
#pragma unroll
|
||||
for (int i = 0; i < UNROLL_FACTOR; i++) {
|
||||
int64_t idx = start + tid * UNROLL_FACTOR + i;
|
||||
T val = (idx < ne00) ? src_row[idx] : T(0);
|
||||
thread_sum += val;
|
||||
items[i] = thread_sum;
|
||||
}
|
||||
|
||||
// Block-wide scan on thread sums
|
||||
T thread_prefix;
|
||||
T block_total;
|
||||
BlockScanT(temp_storage).InclusiveSum(thread_sum, thread_prefix, block_total);
|
||||
__syncthreads();
|
||||
|
||||
// Add offset to each item and store
|
||||
T thread_offset = thread_prefix - thread_sum + block_carry;
|
||||
#pragma unroll
|
||||
for (int i = 0; i < UNROLL_FACTOR; i++) {
|
||||
int64_t idx = start + tid * UNROLL_FACTOR + i;
|
||||
if (idx < ne00) {
|
||||
dst_row[idx] = items[i] + thread_offset;
|
||||
}
|
||||
}
|
||||
|
||||
// Update carry for next tile
|
||||
if (tid == 0) {
|
||||
block_carry += block_total;
|
||||
}
|
||||
|
||||
__syncthreads();
|
||||
}
|
||||
#else
|
||||
|
|
@ -69,7 +80,7 @@ static __global__ void cumsum_cub_kernel(
|
|||
#endif // GGML_CUDA_USE_CUB
|
||||
}
|
||||
|
||||
// Fallback kernel implementation (original)
|
||||
// Fallback kernel implementation
|
||||
template<typename T>
|
||||
static __global__ void cumsum_kernel(
|
||||
const T * src, T * dst,
|
||||
|
|
@ -86,10 +97,10 @@ static __global__ void cumsum_kernel(
|
|||
const int warps_per_block = blockDim.x / warp_size;
|
||||
|
||||
extern __shared__ float smem[];
|
||||
float * s_vals = smem;
|
||||
float * s_warp_sums = smem + blockDim.x;
|
||||
float * s_carry = smem + blockDim.x + warps_per_block;
|
||||
float * s_chunk_total = s_carry + 1;
|
||||
float * s_vals = smem;
|
||||
float * s_warp_sums = smem + blockDim.x;
|
||||
float * s_carry = smem + blockDim.x + warps_per_block;
|
||||
float * s_chunk_total = s_carry + 1;
|
||||
|
||||
// Initialize carry
|
||||
if (tid == 0) {
|
||||
|
|
@ -107,21 +118,39 @@ static __global__ void cumsum_kernel(
|
|||
const T * src_row = src + i1 * s01 + i2 * s02 + i3 * s03;
|
||||
T * dst_row = dst + i1 * s1 + i2 * s2 + i3 * s3;
|
||||
|
||||
for (int64_t start = 0; start < ne00; start += blockDim.x) {
|
||||
int64_t idx = start + tid;
|
||||
float val = (idx < ne00) ? ggml_cuda_cast<float, T>(src_row[idx]) : 0.0f;
|
||||
// register blocking: process 4 elements per thread to hide latency
|
||||
// and reduce synchronization overhead
|
||||
constexpr int num_unroll = 4;
|
||||
T temp[num_unroll];
|
||||
|
||||
// 1. Warp inclusive scan
|
||||
for (int64_t i = 0; i < ne00; i += num_unroll * blockDim.x) {
|
||||
int64_t idx = i + tid * num_unroll;
|
||||
|
||||
// thread local sequential scan
|
||||
temp[0] = (idx < ne00 ? src_row[idx] : T(0));
|
||||
#pragma unroll
|
||||
for (int64_t j = 1; j < num_unroll; j++) {
|
||||
temp[j] = temp[j - 1];
|
||||
if (idx + j < ne00) {
|
||||
temp[j] += src_row[idx + j];
|
||||
} else {
|
||||
temp[j] += 0;
|
||||
}
|
||||
}
|
||||
|
||||
// last emenent is sum of all values assigned to thread
|
||||
float val = (idx < ne00) ? ggml_cuda_cast<float, T>(temp[num_unroll - 1]) : 0.0f;
|
||||
|
||||
// Warp inclusive scan
|
||||
val = warp_prefix_inclusive_sum<T, warp_size>(val);
|
||||
s_vals[tid] = val;
|
||||
|
||||
// Store warp total
|
||||
if (lane == warp_size - 1) {
|
||||
s_warp_sums[warp] = val;
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
// 2. Exclusive scan of warp sums (warp 0 only)
|
||||
// Exclusive scan of warp sums (warp 0 only)
|
||||
if (warp == 0) {
|
||||
float w = (tid < warps_per_block) ? s_warp_sums[tid] : 0.0f;
|
||||
float inc = warp_prefix_inclusive_sum<T, warp_size>(w);
|
||||
|
|
@ -134,12 +163,17 @@ static __global__ void cumsum_kernel(
|
|||
}
|
||||
__syncthreads();
|
||||
|
||||
// write back results
|
||||
float carry = *s_carry;
|
||||
float final_val = s_vals[tid] + s_warp_sums[warp] + carry;
|
||||
if (idx < ne00) {
|
||||
dst_row[idx] = ggml_cuda_cast<T, float>(final_val);
|
||||
// calculate sum offset for this thread
|
||||
float final_val_offset = s_vals[tid] + s_warp_sums[warp] + carry - temp[num_unroll - 1];
|
||||
|
||||
#pragma unroll
|
||||
for (int32_t j = 0; j < num_unroll; j++) {
|
||||
if (idx + j < ne00) {
|
||||
dst_row[idx + j] = temp[j] + ggml_cuda_cast<T, float>(final_val_offset);
|
||||
}
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
// Update carry for next chunk
|
||||
if (tid == 0) {
|
||||
|
|
@ -177,7 +211,7 @@ static void cumsum_cuda(
|
|||
const int warps_per_block = block_size / warp_size;
|
||||
const size_t shmem_size = (block_size + warps_per_block + 2) * sizeof(float);
|
||||
|
||||
if (use_cub) {
|
||||
if (use_cub && ne00 >= 1024) {
|
||||
cumsum_cub_kernel<T, CUDA_CUMSUM_BLOCK_SIZE><<<grid_dims, CUDA_CUMSUM_BLOCK_SIZE, 0, stream>>>(
|
||||
src, dst,
|
||||
ne00, ne01, ne02, ne03,
|
||||
|
|
|
|||
|
|
@ -667,7 +667,7 @@ struct vk_device_struct {
|
|||
vk_pipeline pipeline_add_id_f32;
|
||||
|
||||
vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32;
|
||||
vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bicubic_f32;
|
||||
vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bicubic_f32, pipeline_upscale_bilinear_antialias_f32;
|
||||
vk_pipeline pipeline_scale_f32;
|
||||
vk_pipeline pipeline_sqr_f32;
|
||||
vk_pipeline pipeline_sqrt_f32;
|
||||
|
|
@ -779,6 +779,7 @@ struct vk_device_struct {
|
|||
std::map<vk_fa_pipeline_state, vk_pipeline> pipeline_flash_attn_f32_f16[GGML_TYPE_COUNT];
|
||||
|
||||
vk_pipeline pipeline_flash_attn_split_k_reduce;
|
||||
vk_pipeline pipeline_count_experts;
|
||||
|
||||
// [2] is for whether to take n_experts from spec constant (0) or push constant (1)
|
||||
vk_pipeline pipeline_topk_moe[num_topk_moe_pipelines][TOPK_MOE_COUNT][2];
|
||||
|
|
@ -1020,6 +1021,14 @@ struct vk_op_push_constants {
|
|||
float param4;
|
||||
};
|
||||
|
||||
struct vk_op_count_experts_push_constants {
|
||||
uint32_t ne00;
|
||||
uint32_t ne01;
|
||||
uint32_t nb00;
|
||||
uint32_t nb01;
|
||||
uint32_t a_offset;
|
||||
};
|
||||
|
||||
struct vk_op_glu_push_constants {
|
||||
uint32_t N;
|
||||
uint32_t ne00;
|
||||
|
|
@ -1208,6 +1217,7 @@ struct vk_op_diag_mask_push_constants {
|
|||
struct vk_op_rope_push_constants {
|
||||
uint32_t rope_mode;
|
||||
uint32_t ncols;
|
||||
uint32_t nrows;
|
||||
uint32_t n_dims;
|
||||
float freq_scale;
|
||||
uint32_t p_delta_rows;
|
||||
|
|
@ -1580,7 +1590,7 @@ class vk_perf_logger {
|
|||
total_op_times += time;
|
||||
}
|
||||
std::cerr << t.first << ": " << t.second.size() << " x " << (total_op_times / t.second.size() / 1000.0)
|
||||
<< " us";
|
||||
<< " us = " << (total_op_times / 1000.0) << " us";
|
||||
|
||||
// If we have as many flops entries as timing entries for the op, then compute and log the flops/S.
|
||||
auto it = flops.find(t.first);
|
||||
|
|
@ -2845,9 +2855,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
|||
s_mmq_wg_denoms_k = { 32, 64, 1 };
|
||||
|
||||
// spec constants and tile sizes for quant matmul_id
|
||||
l_warptile_mmqid = { 256, 128, 128, 16, 1, device->subgroup_size };
|
||||
m_warptile_mmqid = { 256, 128, 64, 16, 0, device->subgroup_size };
|
||||
s_warptile_mmqid = { 256, 128, 64, 16, 0, device->subgroup_size };
|
||||
l_warptile_mmqid = { 256, 128, 128, 32, 1, device->subgroup_size };
|
||||
m_warptile_mmqid = { 256, 128, 64, 32, 0, device->subgroup_size };
|
||||
s_warptile_mmqid = { 256, 128, 64, 32, 0, device->subgroup_size };
|
||||
l_mmqid_wg_denoms = { 128, 128, 1 };
|
||||
m_mmqid_wg_denoms = { 128, 64, 1 };
|
||||
s_mmqid_wg_denoms = { 128, 64, 1 };
|
||||
|
|
@ -3083,17 +3093,19 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
|||
#endif
|
||||
#undef CREATE_FA
|
||||
|
||||
const int mul_mat_id_param_count = 5;
|
||||
|
||||
#if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
|
||||
if (device->coopmat2) {
|
||||
|
||||
// Create 6 variants, {s,m,l}x{unaligned,aligned}
|
||||
#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
|
||||
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \
|
||||
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \
|
||||
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \
|
||||
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align); \
|
||||
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align); \
|
||||
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \
|
||||
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, true); \
|
||||
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, true); \
|
||||
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, true); \
|
||||
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, true); \
|
||||
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, true); \
|
||||
ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, true); \
|
||||
|
||||
// Create 2 variants, {f16,f32} accumulator
|
||||
#define CREATE_MM2(PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
|
||||
|
|
@ -3129,32 +3141,32 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
|||
|
||||
GGML_ASSERT(device->subgroup_ballot);
|
||||
|
||||
CREATE_MM2(pipeline_matmul_id_f16, matmul_id_subgroup_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_matmul_id_f16, matmul_id_subgroup_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, 5)
|
||||
#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
|
||||
if (device->coopmat_bf16_support) {
|
||||
CREATE_MM(pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM(pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 5)
|
||||
}
|
||||
#endif
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_subgroup_q4_1_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_subgroup_q5_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_subgroup_q6_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S], matmul_id_subgroup_iq1_s_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M], matmul_id_subgroup_iq1_m_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_subgroup_iq2_xxs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS], matmul_id_subgroup_iq2_xs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S], matmul_id_subgroup_iq2_s_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_subgroup_iq3_xxs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S], matmul_id_subgroup_iq3_s_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS], matmul_id_subgroup_iq4_xs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_subgroup_iq4_nl_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4], matmul_id_subgroup_mxfp4_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_subgroup_q4_1_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_subgroup_q5_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_subgroup_q6_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S], matmul_id_subgroup_iq1_s_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M], matmul_id_subgroup_iq1_m_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_subgroup_iq2_xxs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS], matmul_id_subgroup_iq2_xs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S], matmul_id_subgroup_iq2_s_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_subgroup_iq3_xxs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S], matmul_id_subgroup_iq3_s_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS], matmul_id_subgroup_iq4_xs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_subgroup_iq4_nl_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4], matmul_id_subgroup_mxfp4_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 5)
|
||||
#undef CREATE_MM
|
||||
#undef CREATE_MM2
|
||||
} else
|
||||
|
|
@ -3243,35 +3255,35 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
|||
|
||||
GGML_ASSERT(device->subgroup_ballot);
|
||||
|
||||
CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_subgroup_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_subgroup_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
|
||||
CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_subgroup_f16, wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_subgroup_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id);
|
||||
#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
|
||||
if (device->coopmat_bf16_support) {
|
||||
CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
|
||||
CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id);
|
||||
}
|
||||
#endif
|
||||
|
||||
CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_subgroup_q4_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_subgroup_q5_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_subgroup_q6_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S], matmul_id_subgroup_iq1_s_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M], matmul_id_subgroup_iq1_m_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_subgroup_iq2_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS], matmul_id_subgroup_iq2_xs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S], matmul_id_subgroup_iq2_s_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_subgroup_iq3_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S], matmul_id_subgroup_iq3_s_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS], matmul_id_subgroup_iq4_xs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_subgroup_iq4_nl_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4], matmul_id_subgroup_mxfp4_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_subgroup_q4_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_subgroup_q5_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_subgroup_q6_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S], matmul_id_subgroup_iq1_s_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M], matmul_id_subgroup_iq1_m_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_subgroup_iq2_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS], matmul_id_subgroup_iq2_xs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S], matmul_id_subgroup_iq2_s_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_subgroup_iq3_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S], matmul_id_subgroup_iq3_s_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS], matmul_id_subgroup_iq4_xs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_subgroup_iq4_nl_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
CREATE_MM2(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4], matmul_id_subgroup_mxfp4_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
|
||||
#undef CREATE_MM2
|
||||
#undef CREATE_MM
|
||||
} else
|
||||
|
|
@ -3356,91 +3368,91 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
|||
#endif
|
||||
|
||||
if (device->subgroup_ballot && device->subgroup_require_full_support && subgroup_min_size_16) {
|
||||
CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile_id, vk_mat_mat_push_constants, 4, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_subgroup_f16, wg_denoms, warptile_id, vk_mat_mat_push_constants, 4, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_subgroup_f16_f32, wg_denoms, warptile_id, vk_mat_mat_push_constants, 4, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile_id, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile_id, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_subgroup_f16, wg_denoms, warptile_id, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_subgroup_f16_f32, wg_denoms, warptile_id, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile_id, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
|
||||
|
||||
CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_subgroup_q4_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_subgroup_q5_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_subgroup_q6_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S], matmul_id_subgroup_iq1_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M], matmul_id_subgroup_iq1_m_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_subgroup_iq2_xxs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS], matmul_id_subgroup_iq2_xs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S], matmul_id_subgroup_iq2_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_subgroup_iq3_xxs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S], matmul_id_subgroup_iq3_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS], matmul_id_subgroup_iq4_xs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_subgroup_iq4_nl_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4], matmul_id_subgroup_mxfp4_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_subgroup_q4_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_subgroup_q5_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_subgroup_q6_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S], matmul_id_subgroup_iq1_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M], matmul_id_subgroup_iq1_m_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_subgroup_iq2_xxs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS], matmul_id_subgroup_iq2_xs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S], matmul_id_subgroup_iq2_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_subgroup_iq3_xxs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S], matmul_id_subgroup_iq3_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS], matmul_id_subgroup_iq4_xs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_subgroup_iq4_nl_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM2(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4], matmul_id_subgroup_mxfp4_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
|
||||
#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
|
||||
if (device->integer_dot_product) {
|
||||
CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_1], matmul_id_subgroup_q4_1_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_0], matmul_id_subgroup_q5_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_1], matmul_id_subgroup_q4_1_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_0], matmul_id_subgroup_q5_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_1], matmul_id_subgroup_q5_1_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q8_0], matmul_id_subgroup_q8_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
|
||||
CREATE_MMQ(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_MXFP4], matmul_id_subgroup_mxfp4_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MMQ(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_MXFP4], matmul_id_subgroup_mxfp4_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
|
||||
CREATE_MMQ(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MMQ(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MMQ(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MMQ(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MMQ(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q6_K], matmul_id_subgroup_q6_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MMQ(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q2_K], matmul_id_subgroup_q2_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MMQ(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q3_K], matmul_id_subgroup_q3_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MMQ(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_K], matmul_id_subgroup_q4_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MMQ(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_K], matmul_id_subgroup_q5_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MMQ(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q6_K], matmul_id_subgroup_q6_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
|
||||
}
|
||||
#endif
|
||||
} else {
|
||||
CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
|
||||
CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S], matmul_id_iq1_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M], matmul_id_iq1_m_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_iq2_xxs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS], matmul_id_iq2_xs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S], matmul_id_iq2_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_iq3_xxs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S], matmul_id_iq3_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS], matmul_id_iq4_xs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4], matmul_id_mxfp4_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S], matmul_id_iq1_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M], matmul_id_iq1_m_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_iq2_xxs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS], matmul_id_iq2_xs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S], matmul_id_iq2_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_iq3_xxs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S], matmul_id_iq3_s_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS], matmul_id_iq4_xs_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM2(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4], matmul_id_mxfp4_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
|
||||
#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
|
||||
if (device->integer_dot_product) {
|
||||
CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_0], matmul_id_q4_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_1], matmul_id_q4_1_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_0], matmul_id_q5_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_1], matmul_id_q5_1_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q8_0], matmul_id_q8_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_0], matmul_id_q4_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_1], matmul_id_q4_1_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_0], matmul_id_q5_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_1], matmul_id_q5_1_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q8_0], matmul_id_q8_0_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
|
||||
CREATE_MMQ(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_MXFP4], matmul_id_mxfp4_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_MXFP4], matmul_id_mxfp4_q8_1, mmq_wg_denoms, warptile_mmqid_int, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
|
||||
CREATE_MMQ(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q2_K], matmul_id_q2_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q3_K], matmul_id_q3_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_K], matmul_id_q4_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_K], matmul_id_q5_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q6_K], matmul_id_q6_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q2_K], matmul_id_q2_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q3_K], matmul_id_q3_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q4_K], matmul_id_q4_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q5_K], matmul_id_q5_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MMQ(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id_q8_1[GGML_TYPE_Q6_K], matmul_id_q6_k_q8_1, mmq_wg_denoms, warptile_mmqid_int_k, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
}
|
||||
#endif
|
||||
}
|
||||
|
|
@ -3517,57 +3529,57 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
|||
#endif
|
||||
|
||||
if (device->subgroup_ballot && device->subgroup_require_full_support && subgroup_min_size_16) {
|
||||
CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile_id, vk_mat_mat_push_constants, 4, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_subgroup_f16, , wg_denoms, warptile_id, vk_mat_mat_push_constants, 4, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16_f32.f32acc, matmul_id_subgroup_f16_f32, , wg_denoms, warptile_id, vk_mat_mat_push_constants, 4, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile_id, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile_id, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_subgroup_f16, , wg_denoms, warptile_id, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16_f32.f32acc, matmul_id_subgroup_f16_f32, , wg_denoms, warptile_id, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
|
||||
CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile_id, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
|
||||
|
||||
CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f32acc, matmul_id_subgroup_q4_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f32acc, matmul_id_subgroup_q4_1_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f32acc, matmul_id_subgroup_q5_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f32acc, matmul_id_subgroup_q5_1_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f32acc, matmul_id_subgroup_q8_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f32acc, matmul_id_subgroup_q2_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f32acc, matmul_id_subgroup_q3_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f32acc, matmul_id_subgroup_q4_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f32acc, matmul_id_subgroup_q5_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f32acc, matmul_id_subgroup_q6_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f32acc, matmul_id_subgroup_iq1_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f32acc, matmul_id_subgroup_iq1_m_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f32acc, matmul_id_subgroup_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f32acc, matmul_id_subgroup_iq2_xs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f32acc, matmul_id_subgroup_iq2_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f32acc, matmul_id_subgroup_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f32acc, matmul_id_subgroup_iq3_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f32acc, matmul_id_subgroup_iq4_xs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc, matmul_id_subgroup_iq4_nl_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f32acc, matmul_id_subgroup_mxfp4_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f32acc, matmul_id_subgroup_q4_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f32acc, matmul_id_subgroup_q4_1_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f32acc, matmul_id_subgroup_q5_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f32acc, matmul_id_subgroup_q5_1_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f32acc, matmul_id_subgroup_q8_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f32acc, matmul_id_subgroup_q2_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f32acc, matmul_id_subgroup_q3_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f32acc, matmul_id_subgroup_q4_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f32acc, matmul_id_subgroup_q5_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f32acc, matmul_id_subgroup_q6_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f32acc, matmul_id_subgroup_iq1_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f32acc, matmul_id_subgroup_iq1_m_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f32acc, matmul_id_subgroup_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f32acc, matmul_id_subgroup_iq2_xs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f32acc, matmul_id_subgroup_iq2_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f32acc, matmul_id_subgroup_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f32acc, matmul_id_subgroup_iq3_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f32acc, matmul_id_subgroup_iq4_xs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc, matmul_id_subgroup_iq4_nl_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
CREATE_MM(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f32acc, matmul_id_subgroup_mxfp4_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
|
||||
} else {
|
||||
CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16_f32.f32acc, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16_f32.f32acc, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
|
||||
CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f32acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f32acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f32acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f32acc, matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f32acc, matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f32acc, matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f32acc, matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f32acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f32acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f32acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f32acc, matmul_id_iq1_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f32acc, matmul_id_iq1_m_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f32acc, matmul_id_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f32acc, matmul_id_iq2_xs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f32acc, matmul_id_iq2_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f32acc, matmul_id_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f32acc, matmul_id_iq3_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f32acc, matmul_id_iq4_xs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f32acc, matmul_id_mxfp4_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f32acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f32acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f32acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f32acc, matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f32acc, matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f32acc, matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f32acc, matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f32acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f32acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f32acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f32acc, matmul_id_iq1_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f32acc, matmul_id_iq1_m_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f32acc, matmul_id_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f32acc, matmul_id_iq2_xs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f32acc, matmul_id_iq2_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f32acc, matmul_id_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f32acc, matmul_id_iq3_s_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f32acc, matmul_id_iq4_xs_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_MXFP4, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f32acc, matmul_id_mxfp4_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
}
|
||||
}
|
||||
// reusing CREATE_MM from the fp32 path
|
||||
|
|
@ -3586,7 +3598,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
|||
s_wg_denoms = { 32, 32, 1 };
|
||||
|
||||
CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, , 0);
|
||||
CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id, 0);
|
||||
CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
|
||||
}
|
||||
#undef CREATE_MM
|
||||
|
||||
|
|
@ -3971,6 +3983,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
|||
ggml_vk_create_pipeline(device, device->pipeline_upscale_nearest_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_NEAREST}, 1);
|
||||
ggml_vk_create_pipeline(device, device->pipeline_upscale_bilinear_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_BILINEAR}, 1);
|
||||
ggml_vk_create_pipeline(device, device->pipeline_upscale_bicubic_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_BICUBIC}, 1);
|
||||
ggml_vk_create_pipeline(device, device->pipeline_upscale_bilinear_antialias_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ANTIALIAS}, 1);
|
||||
|
||||
ggml_vk_create_pipeline(device, device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
|
||||
|
||||
|
|
@ -4142,6 +4155,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
|||
|
||||
ggml_vk_create_pipeline(device, device->pipeline_count_equal_i32, "count_equal_i32", count_equal_i32_len, count_equal_i32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, { device->subgroup_size }, 1);
|
||||
|
||||
ggml_vk_create_pipeline(device, device->pipeline_count_experts, "count_experts", count_experts_len, count_experts_data, "main", 2, sizeof(vk_op_count_experts_push_constants), {1, 1, 1}, {}, 1, true);
|
||||
|
||||
for (auto &s : device->pipeline_solve_tri_f32) {
|
||||
const vk_solve_tri_pipeline_state &state = s.first;
|
||||
|
||||
|
|
@ -6553,18 +6568,18 @@ static uint32_t ggml_vk_guess_matmul_id_pipeline_align(ggml_backend_vk_context *
|
|||
|
||||
static void ggml_vk_matmul_id(
|
||||
ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline& pipeline,
|
||||
vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& ids,
|
||||
vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& ids, const vk_subbuffer & expert_count_buf,
|
||||
uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d,
|
||||
uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d,
|
||||
uint32_t n_as, uint32_t nei0, uint32_t nei1, uint32_t nbi1, uint32_t ne11,
|
||||
uint32_t padded_n) {
|
||||
VK_LOG_DEBUG("ggml_vk_matmul_id(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), ids: (" << ids.buffer->buffer << ", " << ids.offset << ", " << ids.size << "), " <<
|
||||
VK_LOG_DEBUG("ggml_vk_matmul_id(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), ids: (" << ids.buffer->buffer << ", " << ids.offset << ", " << ids.size << "), expert_count: (" << expert_count_buf.buffer->buffer << ", " << expert_count_buf.offset << ", " << expert_count_buf.size << "), " <<
|
||||
"m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", " <<
|
||||
"batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", " <<
|
||||
"n_as: " << n_as << ", nei0: " << nei0 << ", nei1: " << nei1 << ", nbi1: " << nbi1 << ", ne11: " << ne11 << ")");
|
||||
const vk_mat_mat_id_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d,
|
||||
nei0, nei1, nbi1, ne11, padded_n };
|
||||
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d, ids }, pc, { m, nei1, n_as });
|
||||
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d, ids, expert_count_buf }, pc, { m, nei1, n_as });
|
||||
}
|
||||
|
||||
static bool ggml_vk_dim01_contiguous(const ggml_tensor * tensor) {
|
||||
|
|
@ -7547,6 +7562,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
|
|||
const uint64_t nei0 = ids->ne[0];
|
||||
const uint64_t nei1 = ids->ne[1];
|
||||
|
||||
const uint32_t nbi0 = ids->nb[0];
|
||||
const uint32_t nbi1 = ids->nb[1];
|
||||
const uint32_t nbi2 = ids->nb[2];
|
||||
|
||||
|
|
@ -7654,6 +7670,9 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
|
|||
if (quantize_y) {
|
||||
to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
|
||||
}
|
||||
vk_pipeline count_experts = ctx->device->pipeline_count_experts;
|
||||
|
||||
uint32_t expert_count_size = sizeof(uint32_t) * n_as;
|
||||
|
||||
{
|
||||
if (
|
||||
|
|
@ -7669,6 +7688,10 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
|
|||
ctx->prealloc_size_y = y_sz;
|
||||
ggml_vk_preallocate_buffers(ctx, subctx);
|
||||
}
|
||||
if (ctx->prealloc_size_split_k < expert_count_size) {
|
||||
ctx->prealloc_size_split_k = expert_count_size;
|
||||
ggml_vk_preallocate_buffers(ctx, subctx);
|
||||
}
|
||||
|
||||
// Request descriptor sets
|
||||
ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
|
||||
|
|
@ -7681,6 +7704,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
|
|||
if (quantize_y) {
|
||||
ggml_pipeline_request_descriptor_sets(ctx, to_q8_1, 1);
|
||||
}
|
||||
ggml_pipeline_request_descriptor_sets(ctx, count_experts, 1);
|
||||
}
|
||||
|
||||
vk_buffer d_D = dst_buf_ctx->dev_buffer;
|
||||
|
|
@ -7730,6 +7754,20 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
|
|||
ggml_vk_sync_buffers(ctx, subctx);
|
||||
}
|
||||
}
|
||||
// Count how many times each expert is used
|
||||
vk_subbuffer expert_count_buf = ggml_vk_subbuffer(ctx, ctx->prealloc_split_k, 0);
|
||||
if (ctx->prealloc_split_k_need_sync) {
|
||||
ggml_vk_sync_buffers(ctx, subctx);
|
||||
}
|
||||
{
|
||||
const std::vector<uint32_t> pc = { (uint32_t)nei0,
|
||||
(uint32_t)nei1,
|
||||
(uint32_t)(nbi0 / ggml_type_size(ids->type)),
|
||||
(uint32_t)(nbi1 / ggml_type_size(ids->type)),
|
||||
(uint32_t)(get_misalign_bytes(ctx, ids) / ggml_type_size(ids->type)) };
|
||||
ggml_vk_dispatch_pipeline(ctx, subctx, count_experts,
|
||||
{ vk_subbuffer{ d_ids, ids_buf_offset, ids_sz }, expert_count_buf }, pc, { (uint32_t)n_as, 1, 1});
|
||||
}
|
||||
|
||||
if (x_non_contig) {
|
||||
ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, ggml_vk_subbuffer(ctx, d_Qx, qx_buf_offset), ggml_vk_subbuffer(ctx, d_X, 0));
|
||||
|
|
@ -7737,7 +7775,6 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
|
|||
const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
|
||||
ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0,
|
||||
{ vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_X, 0, x_sz } }, pc, { (uint32_t)x_ne, 1, 1});
|
||||
ggml_vk_sync_buffers(ctx, subctx);
|
||||
}
|
||||
if (y_non_contig) {
|
||||
if (ctx->prealloc_y_last_pipeline_used != to_fp16_vk_1.get() ||
|
||||
|
|
@ -7761,6 +7798,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
|
|||
ctx->prealloc_y_last_tensor_used = src1;
|
||||
}
|
||||
}
|
||||
ggml_vk_sync_buffers(ctx, subctx);
|
||||
|
||||
uint32_t stride_batch_x = ne00*ne01;
|
||||
uint32_t stride_batch_y = ne10*ne11;
|
||||
|
|
@ -7777,7 +7815,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
|
|||
ggml_vk_matmul_id(
|
||||
ctx, subctx, pipeline,
|
||||
{ d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz },
|
||||
{ d_D, d_buf_offset, d_sz }, { d_ids, ids_buf_offset, ids_sz },
|
||||
{ d_D, d_buf_offset, d_sz }, { d_ids, ids_buf_offset, ids_sz }, expert_count_buf,
|
||||
ne01, ne21, ne10, ne10, ne10, ne01,
|
||||
stride_batch_x, stride_batch_y, ne20*ne21,
|
||||
n_as, nei0, nei1, nbi1 / ggml_type_size(ids->type), ne11, padded_n
|
||||
|
|
@ -7789,6 +7827,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
|
|||
if (y_non_contig || quantize_y) {
|
||||
ctx->prealloc_y_need_sync = true;
|
||||
}
|
||||
ctx->prealloc_split_k_need_sync = true;
|
||||
}
|
||||
|
||||
static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx) {
|
||||
|
|
@ -8462,7 +8501,7 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
|
|||
return nullptr;
|
||||
case GGML_OP_UPSCALE:
|
||||
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
|
||||
ggml_scale_mode mode = (ggml_scale_mode)(ggml_get_op_params_i32(dst, 0) & 0xFF);
|
||||
uint32_t mode = (ggml_get_op_params_i32(dst, 0) & (0xFF | GGML_SCALE_FLAG_ANTIALIAS));
|
||||
switch (mode) {
|
||||
case GGML_SCALE_MODE_NEAREST:
|
||||
return ctx->device->pipeline_upscale_nearest_f32;
|
||||
|
|
@ -8470,6 +8509,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
|
|||
return ctx->device->pipeline_upscale_bilinear_f32;
|
||||
case GGML_SCALE_MODE_BICUBIC:
|
||||
return ctx->device->pipeline_upscale_bicubic_f32;
|
||||
case GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ANTIALIAS:
|
||||
return ctx->device->pipeline_upscale_bilinear_antialias_f32;
|
||||
default:
|
||||
return nullptr;
|
||||
}
|
||||
|
|
@ -9120,10 +9161,20 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
|
|||
elements = { num_groups * (uint32_t)src0->ne[3], 1, 1 };
|
||||
} break;
|
||||
case GGML_OP_DIAG_MASK_INF:
|
||||
case GGML_OP_ROPE:
|
||||
case GGML_OP_ROPE_BACK:
|
||||
elements = { (uint32_t)ggml_nrows(src0), (uint32_t)ne00, 1 };
|
||||
break;
|
||||
case GGML_OP_ROPE:
|
||||
case GGML_OP_ROPE_BACK:
|
||||
{
|
||||
uint32_t nrows = (uint32_t)ggml_nrows(src0);
|
||||
uint32_t z = 1;
|
||||
if (nrows > ctx->device->properties.limits.maxComputeWorkGroupCount[0]) {
|
||||
z = CEIL_DIV(nrows, 32768);
|
||||
nrows = 32768;
|
||||
}
|
||||
elements = { nrows, (uint32_t)ne00, z };
|
||||
|
||||
} break;
|
||||
case GGML_OP_GET_ROWS:
|
||||
elements = { (uint32_t)ne00, (uint32_t)ne10, (uint32_t)(ne11 * ne12) };
|
||||
elements[1] = std::min(elements[1], ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
|
||||
|
|
@ -10051,7 +10102,7 @@ static vk_op_rope_push_constants ggml_vk_make_rope_constants(const ggml_tensor *
|
|||
uint32_t nb02 = src0->nb[2] / ggml_type_size(src0->type);
|
||||
|
||||
vk_op_rope_push_constants rope {
|
||||
(uint32_t)mode, (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1],
|
||||
(uint32_t)mode, (uint32_t)src0->ne[0], (uint32_t)ggml_nrows(src0), (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1],
|
||||
freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale,
|
||||
has_ff, (uint32_t)src0->ne[2], nb01, nb02,
|
||||
{ sections[0], sections[1], sections[2], sections[3] }, is_imrope, backprop, set_rows_stride,
|
||||
|
|
@ -14360,7 +14411,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
|
|||
}
|
||||
return true;
|
||||
case GGML_OP_UPSCALE:
|
||||
return op->src[0]->type == GGML_TYPE_F32 && !(op->op_params[0] & GGML_SCALE_FLAG_ANTIALIAS);
|
||||
if (op->op_params[0] & GGML_SCALE_FLAG_ANTIALIAS) {
|
||||
if ((op->op_params[0] & 0xFF) != GGML_SCALE_MODE_BILINEAR) {
|
||||
return false;
|
||||
}
|
||||
}
|
||||
return op->src[0]->type == GGML_TYPE_F32;
|
||||
case GGML_OP_ACC:
|
||||
return op->src[0]->type == GGML_TYPE_F32;
|
||||
case GGML_OP_CONCAT:
|
||||
|
|
|
|||
51
ggml/src/ggml-vulkan/vulkan-shaders/count_experts.comp
Normal file
51
ggml/src/ggml-vulkan/vulkan-shaders/count_experts.comp
Normal file
|
|
@ -0,0 +1,51 @@
|
|||
#version 450
|
||||
|
||||
#extension GL_EXT_control_flow_attributes : enable
|
||||
|
||||
#include "types.glsl"
|
||||
|
||||
layout (push_constant) uniform parameter
|
||||
{
|
||||
uint32_t ne00;
|
||||
uint32_t ne01;
|
||||
uint32_t nb00;
|
||||
uint32_t nb01;
|
||||
uint32_t a_offset;
|
||||
} p;
|
||||
|
||||
#define BLOCK_SIZE 256
|
||||
|
||||
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
|
||||
|
||||
layout (binding = 0) readonly buffer A {uint data_a[];};
|
||||
layout (binding = 1) writeonly buffer D {uint data_d[];};
|
||||
|
||||
shared uint vals[BLOCK_SIZE];
|
||||
|
||||
void main() {
|
||||
const uint expert_id = gl_WorkGroupID.x;
|
||||
const uint num_elements = p.ne00 * p.ne01;
|
||||
const uint tid = gl_LocalInvocationID.x;
|
||||
|
||||
uint count = 0;
|
||||
for (uint idx = tid; idx < num_elements; idx += BLOCK_SIZE) {
|
||||
const uint i01 = idx / p.ne00;
|
||||
const uint i00 = idx % p.ne00;
|
||||
const uint a = data_a[p.a_offset + i01 * p.nb01 + i00 * p.nb00];
|
||||
|
||||
count += uint(a == expert_id);
|
||||
}
|
||||
|
||||
vals[tid] = count;
|
||||
barrier();
|
||||
[[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
|
||||
if (tid < s) {
|
||||
vals[tid] += vals[tid + s];
|
||||
}
|
||||
barrier();
|
||||
}
|
||||
|
||||
if (tid == 0) {
|
||||
data_d[expert_id] = vals[0];
|
||||
}
|
||||
}
|
||||
|
|
@ -401,13 +401,7 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
|
|||
const uint sl = (data_a[a_offset + ib].scales_l[ib32/2] >> (4 * (ib32 & 1))) & 0xF;
|
||||
const uint sh = (data_a[a_offset + ib].scales_h >> (2 * ib32)) & 3;
|
||||
const uint qshift = (iqs & 16) >> 2;
|
||||
u8vec4 qs = u8vec4(
|
||||
data_a[a_offset + ib].qs[iq + 0],
|
||||
data_a[a_offset + ib].qs[iq + 1],
|
||||
data_a[a_offset + ib].qs[iq + 2],
|
||||
data_a[a_offset + ib].qs[iq + 3]
|
||||
);
|
||||
qs = (qs >> qshift) & uint8_t(0xF);
|
||||
const u8vec4 qs = unpack8((data_a_packed32[a_offset + ib].qs[iq/4] >> qshift) & 0x0F0F0F0F);
|
||||
|
||||
const float dl = float(int(sl | (sh << 4)) - 32);
|
||||
return dl * vec4(
|
||||
|
|
|
|||
|
|
@ -68,6 +68,7 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
|
|||
|
||||
#ifdef MUL_MAT_ID
|
||||
layout (binding = 3) readonly buffer IDS {int data_ids[];};
|
||||
layout (binding = 4) readonly buffer Counts {int data_expert_count[];};
|
||||
#endif
|
||||
|
||||
layout (push_constant) uniform parameter
|
||||
|
|
@ -135,13 +136,19 @@ shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
|
|||
#include "mul_mm_funcs.glsl"
|
||||
|
||||
void main() {
|
||||
const uint ic = gl_WorkGroupID.y;
|
||||
|
||||
#ifdef MUL_MAT_ID
|
||||
const uint expert_idx = gl_GlobalInvocationID.z;
|
||||
if (ic * BN >= data_expert_count[expert_idx]) {
|
||||
return;
|
||||
}
|
||||
#endif
|
||||
#ifdef NEEDS_INIT_IQ_SHMEM
|
||||
init_iq_shmem(gl_WorkGroupSize);
|
||||
#endif
|
||||
|
||||
#ifdef MUL_MAT_ID
|
||||
const uint expert_idx = gl_GlobalInvocationID.z;
|
||||
#else
|
||||
#ifndef MUL_MAT_ID
|
||||
const uint batch_idx = gl_GlobalInvocationID.z;
|
||||
|
||||
const uint i13 = batch_idx / p.ne12;
|
||||
|
|
@ -156,7 +163,6 @@ void main() {
|
|||
const uint blocks_m = (p.M + BM - 1) / BM;
|
||||
const uint ir = gl_WorkGroupID.x % blocks_m;
|
||||
const uint ik = gl_WorkGroupID.x / blocks_m;
|
||||
const uint ic = gl_WorkGroupID.y;
|
||||
|
||||
const uint WNITER = (WM * WN) / (WARP * TM * TN * WMITER);
|
||||
const uint WSUBM = WM / WMITER;
|
||||
|
|
|
|||
|
|
@ -92,6 +92,7 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
|
|||
|
||||
#ifdef MUL_MAT_ID
|
||||
layout (binding = 3) readonly buffer IDS {int data_ids[];};
|
||||
layout (binding = 4) readonly buffer Counts {int data_expert_count[];};
|
||||
|
||||
shared u16vec4 row_ids[BN];
|
||||
|
||||
|
|
@ -107,11 +108,7 @@ B_TYPE decodeFuncB(const in decodeBufB bl, const in uint blockCoords[2], const i
|
|||
{
|
||||
const uint row_i = blockCoords[0];
|
||||
|
||||
if (row_i >= _ne1) {
|
||||
return B_TYPE(0.0);
|
||||
}
|
||||
|
||||
const u16vec4 row_idx = row_ids[row_i & (BN - 1)];
|
||||
const u16vec4 row_idx = row_ids[row_i];
|
||||
B_TYPE ret = data_b[row_idx.y * p.batch_stride_b + row_idx.x * p.stride_b + blockCoords[1]];
|
||||
|
||||
return ret;
|
||||
|
|
@ -138,6 +135,8 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
|
|||
uint ids[16];
|
||||
uint iter = 0;
|
||||
|
||||
uint expert_count = data_expert_count[expert_idx];
|
||||
|
||||
for (uint j = 0; j < num_elements; j += BLOCK_SIZE) {
|
||||
// prefetch up to 16 elements
|
||||
if (iter == 0) {
|
||||
|
|
@ -185,7 +184,7 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
|
|||
}
|
||||
_ne1 += total;
|
||||
iter &= 15;
|
||||
if (_ne1 >= (ic + 1) * BN) {
|
||||
if (_ne1 >= (ic + 1) * BN || _ne1 == expert_count) {
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
|
@ -194,15 +193,28 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
|
|||
#endif
|
||||
|
||||
void main() {
|
||||
const uint tid = gl_LocalInvocationIndex;
|
||||
const uint ic = gl_WorkGroupID.y;
|
||||
|
||||
#ifdef MUL_MAT_ID
|
||||
const uint expert_idx = gl_GlobalInvocationID.z;
|
||||
if (ic * BN >= data_expert_count[expert_idx]) {
|
||||
return;
|
||||
}
|
||||
// initialize to row 0 so we don't need to bounds check
|
||||
if (tid < BN) {
|
||||
row_ids[tid] = u16vec4(0);
|
||||
}
|
||||
#if !defined(NEEDS_INIT_IQ_SHMEM)
|
||||
barrier();
|
||||
#endif
|
||||
#endif
|
||||
|
||||
#ifdef NEEDS_INIT_IQ_SHMEM
|
||||
init_iq_shmem(gl_WorkGroupSize);
|
||||
#endif
|
||||
|
||||
const uint tid = gl_LocalInvocationIndex;
|
||||
|
||||
#ifdef MUL_MAT_ID
|
||||
const uint expert_idx = gl_GlobalInvocationID.z;
|
||||
#else
|
||||
#ifndef MUL_MAT_ID
|
||||
const uint batch_idx = gl_GlobalInvocationID.z;
|
||||
|
||||
const uint i13 = batch_idx / p.ne12;
|
||||
|
|
@ -217,7 +229,6 @@ void main() {
|
|||
const uint blocks_m = (p.M + BM - 1) / BM;
|
||||
const uint ir = gl_WorkGroupID.x % blocks_m;
|
||||
const uint ik = gl_WorkGroupID.x / blocks_m;
|
||||
const uint ic = gl_WorkGroupID.y;
|
||||
|
||||
#ifdef MUL_MAT_ID
|
||||
if (bitCount(p.nei0) == 1) {
|
||||
|
|
@ -482,7 +493,7 @@ void main() {
|
|||
coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
|
||||
|
||||
coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
|
||||
coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB);
|
||||
coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB);
|
||||
|
||||
sum = coopMatMulAdd(mat_a, mat_b, sum);
|
||||
} else {
|
||||
|
|
@ -490,7 +501,7 @@ void main() {
|
|||
coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
|
||||
|
||||
coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
|
||||
coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB);
|
||||
coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB);
|
||||
|
||||
sum = coopMatMulAdd(mat_a, mat_b, sum);
|
||||
}
|
||||
|
|
@ -526,7 +537,7 @@ void main() {
|
|||
coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
|
||||
|
||||
coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
|
||||
coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB);
|
||||
coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB);
|
||||
|
||||
sum = coopMatMulAdd(mat_a, mat_b, sum);
|
||||
} else {
|
||||
|
|
@ -534,7 +545,7 @@ void main() {
|
|||
coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
|
||||
|
||||
coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
|
||||
coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB);
|
||||
coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB);
|
||||
|
||||
sum = coopMatMulAdd(mat_a, mat_b, sum);
|
||||
}
|
||||
|
|
@ -571,7 +582,7 @@ void main() {
|
|||
|
||||
coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
|
||||
#ifdef MUL_MAT_ID
|
||||
coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
|
||||
coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
|
||||
#else
|
||||
coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose);
|
||||
#endif
|
||||
|
|
@ -583,7 +594,7 @@ void main() {
|
|||
|
||||
coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
|
||||
#ifdef MUL_MAT_ID
|
||||
coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
|
||||
coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
|
||||
#else
|
||||
coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose);
|
||||
#endif
|
||||
|
|
|
|||
|
|
@ -159,14 +159,16 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
|
|||
const uint is = iqs / 8; // 0..15
|
||||
const uint halfsplit = ((iqs % 64) / 16); // 0,1,2,3
|
||||
const uint qsshift = halfsplit * 2; // 0,2,4,6
|
||||
const uint m = 1 << (4 * n + halfsplit); // 1,2,4,8,16,32,64,128
|
||||
|
||||
const int8_t us = int8_t(((data_a[ib].scales[is % 8] >> (4 * int(is / 8))) & 0xF)
|
||||
| (((data_a[ib].scales[8 + (is % 4)] >> (2 * int(is / 4))) & 3) << 4));
|
||||
const float dl = float(data_a[ib].d) * float(us - 32);
|
||||
|
||||
buf_a[buf_idx] = FLOAT_TYPE_VEC2(dl * float(int8_t((data_a[ib].qs[qsi ] >> qsshift) & 3) - (((data_a[ib].hmask[hmi ] & m) != 0) ? 0 : 4)),
|
||||
dl * float(int8_t((data_a[ib].qs[qsi + 1] >> qsshift) & 3) - (((data_a[ib].hmask[hmi + 1] & m) != 0) ? 0 : 4)));
|
||||
const vec2 qs = vec2(unpack8((uint(data_a_packed16[ib].qs[qsi / 2]) >> qsshift) & 0x0303).xy);
|
||||
const vec2 hm = vec2(unpack8(((uint(data_a_packed16[ib].hmask[hmi / 2]) >> (4 * n + halfsplit)) & 0x0101 ^ 0x0101) << 2).xy);
|
||||
|
||||
buf_a[buf_idx] = FLOAT_TYPE_VEC2(dl * (qs.x - hm.x),
|
||||
dl * (qs.y - hm.y));
|
||||
#elif defined(DATA_A_Q4_K)
|
||||
const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
|
||||
const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
|
||||
|
|
@ -198,8 +200,10 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
|
|||
const float d = loadd.x * sc;
|
||||
const float m = -loadd.y * mbyte;
|
||||
|
||||
buf_a[buf_idx] = FLOAT_TYPE_VEC2(fma(d, float((data_a[ib].qs[qsi ] >> (b * 4)) & 0xF), m),
|
||||
fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF), m));
|
||||
const vec2 q = vec2(unpack8((uint(data_a_packed16[ib].qs[qsi / 2]) >> (b * 4)) & 0x0F0F).xy);
|
||||
|
||||
buf_a[buf_idx] = FLOAT_TYPE_VEC2(fma(d, q.x, m),
|
||||
fma(d, q.y, m));
|
||||
#elif defined(DATA_A_Q5_K)
|
||||
const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
|
||||
const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
|
||||
|
|
@ -213,8 +217,6 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
|
|||
const uint qsi = n * 32 + (iqs % 16) * 2; // 0,2,4..126
|
||||
const uint qhi = (iqs % 16) * 2; // 0,2,4..30
|
||||
|
||||
const uint8_t hm = uint8_t(1 << (iqs / 16));
|
||||
|
||||
const vec2 loadd = vec2(data_a[ib].dm);
|
||||
|
||||
const uint scidx0 = (is < 4) ? is : (is + 4);
|
||||
|
|
@ -234,8 +236,12 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
|
|||
const float d = loadd.x * sc;
|
||||
const float m = -loadd.y * mbyte;
|
||||
|
||||
buf_a[buf_idx] = FLOAT_TYPE_VEC2(fma(d, float((data_a[ib].qs[qsi ] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi ] & hm) != 0 ? 16 : 0), m),
|
||||
fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi + 1] & hm) != 0 ? 16 : 0), m));
|
||||
const uint qs = (uint(data_a_packed16[ib].qs[qsi / 2]) >> (b * 4)) & 0x0F0F;
|
||||
const uint qh = ((uint(data_a_packed16[ib].qh[qhi / 2]) >> (iqs / 16)) & 0x0101) << 4;
|
||||
const vec2 q = vec2(unpack8(qs | qh).xy);
|
||||
|
||||
buf_a[buf_idx] = FLOAT_TYPE_VEC2(fma(d, q.x, m),
|
||||
fma(d, q.y, m));
|
||||
#elif defined(DATA_A_Q6_K)
|
||||
const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
|
||||
const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
|
||||
|
|
@ -394,11 +400,9 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
|
|||
|
||||
const float d = float(data_a[ib].d);
|
||||
const uint qs = data_a[ib].qs[iqs];
|
||||
const uint signs = pack32(u8vec4(
|
||||
data_a[ib].qs[is+0],
|
||||
data_a[ib].qs[is+1],
|
||||
data_a[ib].qs[is+2],
|
||||
data_a[ib].qs[is+3]
|
||||
const uint signs = pack32(u16vec2(
|
||||
data_a_packed16[ib].qs[is/2],
|
||||
data_a_packed16[ib].qs[is/2+1]
|
||||
));
|
||||
const float db = d * 0.5 * (0.5 + (signs >> 28));
|
||||
const uint32_t sign7 = bitfieldExtract(signs, 7 * (int(iqs / 2) % 4), 7);
|
||||
|
|
@ -443,8 +447,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
|
|||
const uint sl = (data_a[ib].scales_l[ib32/2] >> (4 * (ib32 & 1))) & 0xF;
|
||||
const uint sh = ((data_a[ib].scales_h) >> (2 * ib32)) & 3;
|
||||
const uint qshift = (idx & 8) >> 1;
|
||||
u8vec2 qs = u8vec2(data_a[ib].qs[iq], data_a[ib].qs[iq + 1]);
|
||||
qs = (qs >> qshift) & uint8_t(0xF);
|
||||
u8vec2 qs = unpack8((uint(data_a_packed16[ib].qs[iq/2]) >> qshift) & 0x0F0F).xy;
|
||||
|
||||
const float d = float(data_a[ib].d);
|
||||
const vec2 v = d * float(int(sl | (sh << 4)) - 32) * vec2(kvalues_iq4nl[qs.x], kvalues_iq4nl[qs.y]);
|
||||
|
|
|
|||
|
|
@ -13,6 +13,8 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
|
|||
uint ids[16];
|
||||
uint iter = 0;
|
||||
|
||||
uint expert_count = data_expert_count[expert_idx];
|
||||
|
||||
for (uint j = 0; j < num_elements; j += BLOCK_SIZE) {
|
||||
// prefetch up to 16 elements
|
||||
if (iter == 0) {
|
||||
|
|
@ -60,7 +62,7 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
|
|||
}
|
||||
_ne1 += total;
|
||||
iter &= 15;
|
||||
if (_ne1 >= (ic + 1) * BN) {
|
||||
if (_ne1 >= (ic + 1) * BN || _ne1 == expert_count) {
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
|
|
|||
|
|
@ -35,6 +35,7 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
|
|||
|
||||
#ifdef MUL_MAT_ID
|
||||
layout (binding = 3) readonly buffer IDS {int data_ids[];};
|
||||
layout (binding = 4) readonly buffer Counts {int data_expert_count[];};
|
||||
#endif
|
||||
|
||||
layout (push_constant) uniform parameter
|
||||
|
|
@ -104,13 +105,19 @@ block_b_cache cache_b;
|
|||
#include "mul_mmq_funcs.glsl"
|
||||
|
||||
void main() {
|
||||
const uint ic = gl_WorkGroupID.y;
|
||||
|
||||
#ifdef MUL_MAT_ID
|
||||
const uint expert_idx = gl_GlobalInvocationID.z;
|
||||
if (ic * BN >= data_expert_count[expert_idx]) {
|
||||
return;
|
||||
}
|
||||
#endif
|
||||
#ifdef NEEDS_INIT_IQ_SHMEM
|
||||
init_iq_shmem(gl_WorkGroupSize);
|
||||
#endif
|
||||
|
||||
#ifdef MUL_MAT_ID
|
||||
const uint expert_idx = gl_GlobalInvocationID.z;
|
||||
#else
|
||||
#ifndef MUL_MAT_ID
|
||||
const uint batch_idx = gl_GlobalInvocationID.z;
|
||||
|
||||
const uint i13 = batch_idx / p.ne12;
|
||||
|
|
@ -125,7 +132,6 @@ void main() {
|
|||
const uint blocks_m = (p.M + BM - 1) / BM;
|
||||
const uint ir = gl_WorkGroupID.x % blocks_m;
|
||||
const uint ik = gl_WorkGroupID.x / blocks_m;
|
||||
const uint ic = gl_WorkGroupID.y;
|
||||
|
||||
const uint WNITER = (WM * WN) / (WARP * TM * TN * WMITER);
|
||||
const uint WSUBM = WM / WMITER;
|
||||
|
|
|
|||
|
|
@ -6,6 +6,9 @@
|
|||
void main() {
|
||||
const uint i0 = 2*gl_GlobalInvocationID.y;
|
||||
// i1 is actually i2*nb2+i1, but the rows are contiguous
|
||||
const uint i1 = gl_GlobalInvocationID.x;
|
||||
const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
|
||||
if (i1 >= pc.nrows) {
|
||||
return;
|
||||
}
|
||||
rope_multi(i0, i1, pc);
|
||||
}
|
||||
|
|
|
|||
|
|
@ -6,6 +6,9 @@
|
|||
void main() {
|
||||
const uint i0 = 2*gl_GlobalInvocationID.y;
|
||||
// i1 is actually i2*nb2+i1, but the rows are contiguous
|
||||
const uint i1 = gl_GlobalInvocationID.x;
|
||||
const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
|
||||
if (i1 >= pc.nrows) {
|
||||
return;
|
||||
}
|
||||
rope_neox(i0, i1, pc);
|
||||
}
|
||||
|
|
|
|||
|
|
@ -6,6 +6,9 @@
|
|||
void main() {
|
||||
const uint i0 = 2*gl_GlobalInvocationID.y;
|
||||
// i1 is actually i2*nb2+i1, but the rows are contiguous
|
||||
const uint i1 = gl_GlobalInvocationID.x;
|
||||
const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
|
||||
if (i1 >= pc.nrows) {
|
||||
return;
|
||||
}
|
||||
rope_norm(i0, i1, pc);
|
||||
}
|
||||
|
|
|
|||
|
|
@ -6,6 +6,7 @@
|
|||
struct rope_params {
|
||||
uint rope_mode;
|
||||
uint ncols;
|
||||
uint nrows;
|
||||
uint n_dims;
|
||||
float freq_scale;
|
||||
uint p_delta_rows;
|
||||
|
|
|
|||
|
|
@ -6,6 +6,9 @@
|
|||
void main() {
|
||||
const uint i0 = 2*gl_GlobalInvocationID.y;
|
||||
// i1 is actually i2*nb2+i1, but the rows are contiguous
|
||||
const uint i1 = gl_GlobalInvocationID.x;
|
||||
const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
|
||||
if (i1 >= pc.nrows) {
|
||||
return;
|
||||
}
|
||||
rope_vision(i0, i1, pc);
|
||||
}
|
||||
|
|
|
|||
|
|
@ -172,16 +172,12 @@ struct block_q8_0
|
|||
float16_t d;
|
||||
int8_t qs[32];
|
||||
};
|
||||
|
||||
struct block_q8_0_packed16
|
||||
{
|
||||
float16_t d;
|
||||
int16_t qs[32/2];
|
||||
};
|
||||
struct block_q8_0_packed32
|
||||
{
|
||||
float16_t d;
|
||||
int32_t qs[32/4];
|
||||
};
|
||||
|
||||
#if defined(DATA_A_Q8_0)
|
||||
#define QUANT_K QUANT_K_Q8_0
|
||||
|
|
@ -189,7 +185,6 @@ struct block_q8_0_packed32
|
|||
#define QUANT_AUXF 1
|
||||
#define A_TYPE block_q8_0
|
||||
#define A_TYPE_PACKED16 block_q8_0_packed16
|
||||
#define A_TYPE_PACKED32 block_q8_0_packed32
|
||||
#define DATA_A_QUANT_LEGACY
|
||||
#endif
|
||||
|
||||
|
|
@ -201,11 +196,13 @@ struct block_q8_1
|
|||
f16vec2 ds;
|
||||
int8_t qs[32];
|
||||
};
|
||||
|
||||
struct block_q8_1_packed16
|
||||
{
|
||||
f16vec2 ds;
|
||||
int16_t qs[16];
|
||||
};
|
||||
|
||||
struct block_q8_1_packed32
|
||||
{
|
||||
f16vec2 ds;
|
||||
|
|
@ -218,6 +215,7 @@ struct block_q8_1_x4
|
|||
f16vec2 ds[4];
|
||||
int32_t qs[32];
|
||||
};
|
||||
|
||||
struct block_q8_1_x4_packed128
|
||||
{
|
||||
f16vec2 ds[4];
|
||||
|
|
@ -1346,10 +1344,28 @@ struct block_iq4_xs
|
|||
uint8_t qs[QUANT_K_IQ4_XS/2];
|
||||
};
|
||||
|
||||
struct block_iq4_xs_packed16
|
||||
{
|
||||
float16_t d;
|
||||
uint16_t scales_h;
|
||||
uint16_t scales_l[QUANT_K_IQ4_XS/128];
|
||||
uint16_t qs[QUANT_K_IQ4_XS/4];
|
||||
};
|
||||
|
||||
struct block_iq4_xs_packed32
|
||||
{
|
||||
float16_t d;
|
||||
uint16_t scales_h;
|
||||
uint32_t scales_l;
|
||||
uint32_t qs[QUANT_K_IQ4_XS/8];
|
||||
};
|
||||
|
||||
#if defined(DATA_A_IQ4_XS)
|
||||
#define QUANT_K QUANT_K_IQ4_XS
|
||||
#define QUANT_R QUANT_R_IQ4_XS
|
||||
#define A_TYPE block_iq4_xs
|
||||
#define A_TYPE_PACKED16 block_iq4_xs_packed16
|
||||
#define A_TYPE_PACKED32 block_iq4_xs_packed32
|
||||
#endif
|
||||
|
||||
#define QUANT_K_IQ4_NL 32
|
||||
|
|
|
|||
|
|
@ -21,6 +21,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
|
|||
#define NEAREST 0
|
||||
#define BILINEAR 1
|
||||
#define BICUBIC 2
|
||||
#define BILINEAR_ANTIALIAS 513
|
||||
|
||||
layout (constant_id = 0) const uint scale_mode = 0;
|
||||
|
||||
|
|
@ -62,6 +63,56 @@ float interpolate_bilinear(uint i10, uint i11, uint i12, uint i13) {
|
|||
return fetch_bilinear(c0, c1, d, i12, i13);
|
||||
}
|
||||
|
||||
float triangle_filter(float x) {
|
||||
return max(1.0f - abs(x), 0.0f);
|
||||
}
|
||||
|
||||
float interpolate_bilinear_antialias(uint i10, uint i11, uint i12, uint i13) {
|
||||
const float support1 = max(1.0f, 1.0f / p.sf1);
|
||||
const float invscale1 = 1.0f / support1;
|
||||
const float support0 = max(1.0f, 1.0f / p.sf0);
|
||||
const float invscale0 = 1.0f / support0;
|
||||
|
||||
const uint i02 = uint(i12 / p.sf2);
|
||||
const uint i03 = uint(i13 / p.sf3);
|
||||
|
||||
const float y = (float(i11) + p.pixel_offset) / p.sf1;
|
||||
const float x = (float(i10) + p.pixel_offset) / p.sf0;
|
||||
|
||||
// the range of source pixels that contribute
|
||||
const int x_min = max(int(x - support0 + p.pixel_offset), 0);
|
||||
const int x_max = min(int(x + support0 + p.pixel_offset), int(p.ne00));
|
||||
const int y_min = max(int(y - support1 + p.pixel_offset), 0);
|
||||
const int y_max = min(int(y + support1 + p.pixel_offset), int(p.ne01));
|
||||
|
||||
// bilinear filter with antialiasing
|
||||
float val = 0.0f;
|
||||
float total_weight = 0.0f;
|
||||
|
||||
for (int sy = y_min; sy < y_max; sy++) {
|
||||
const float weight_y = triangle_filter((sy - y + p.pixel_offset) * invscale1);
|
||||
|
||||
for (int sx = x_min; sx < x_max; sx++) {
|
||||
const float weight_x = triangle_filter((sx - x + p.pixel_offset) * invscale0);
|
||||
const float weight = weight_x * weight_y;
|
||||
|
||||
if (weight <= 0.0f) {
|
||||
continue;
|
||||
}
|
||||
|
||||
const float pixel = data_a[p.a_offset + i03 * p.nb03 + i02 * p.nb02 + sy * p.nb01 + sx * p.nb00];
|
||||
val += pixel * weight;
|
||||
total_weight += weight;
|
||||
}
|
||||
}
|
||||
|
||||
if (total_weight > 0.0f) {
|
||||
val /= total_weight;
|
||||
}
|
||||
|
||||
return val;
|
||||
}
|
||||
|
||||
// Bicubic interpolation with alpha = -0.75
|
||||
// https://en.wikipedia.org/wiki/Bicubic_interpolation#Bicubic_convolution_algorithm
|
||||
const vec4 bcoeffs1 = vec4( 1.25, -2.25, 0.0, 1.0);
|
||||
|
|
@ -118,6 +169,9 @@ void main() {
|
|||
case BICUBIC:
|
||||
result = interpolate_bicubic(i10, i11, i12, i13);
|
||||
break;
|
||||
case BILINEAR_ANTIALIAS:
|
||||
result = interpolate_bilinear_antialias(i10, i11, i12, i13);
|
||||
break;
|
||||
}
|
||||
|
||||
data_d[p.d_offset + idx] = D_TYPE(result);
|
||||
|
|
|
|||
|
|
@ -962,6 +962,8 @@ void process_shaders() {
|
|||
string_to_spv("count_equal_i32", "count_equal.comp", merge_maps(base_dict, {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}}));
|
||||
string_to_spv("cumsum_f32", "cumsum.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
|
||||
|
||||
string_to_spv("count_experts", "count_experts.comp", merge_maps(base_dict, {{"A_TYPE", "uint"}, {"D_TYPE", "uint"}}));
|
||||
|
||||
for (std::string dim_str : {"", "_3d"}) {
|
||||
for (bool bda : {false, true}) {
|
||||
std::string bda_str = bda ? "_bda" : "";
|
||||
|
|
|
|||
|
|
@ -449,6 +449,7 @@ class MODEL_ARCH(IntEnum):
|
|||
RND1 = auto()
|
||||
PANGU_EMBED = auto()
|
||||
MISTRAL3 = auto()
|
||||
MIMO2 = auto()
|
||||
LLAMA_EMBED = auto()
|
||||
|
||||
|
||||
|
|
@ -845,6 +846,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
|
|||
MODEL_ARCH.RND1: "rnd1",
|
||||
MODEL_ARCH.PANGU_EMBED: "pangu-embedded",
|
||||
MODEL_ARCH.MISTRAL3: "mistral3",
|
||||
MODEL_ARCH.MIMO2: "mimo2",
|
||||
MODEL_ARCH.LLAMA_EMBED: "llama-embed",
|
||||
}
|
||||
|
||||
|
|
@ -3198,6 +3200,26 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
|
|||
MODEL_TENSOR.FFN_DOWN_EXP,
|
||||
MODEL_TENSOR.FFN_UP_EXP,
|
||||
],
|
||||
MODEL_ARCH.MIMO2: [
|
||||
MODEL_TENSOR.TOKEN_EMBD,
|
||||
MODEL_TENSOR.OUTPUT_NORM,
|
||||
MODEL_TENSOR.OUTPUT,
|
||||
MODEL_TENSOR.ATTN_NORM,
|
||||
MODEL_TENSOR.ATTN_Q,
|
||||
MODEL_TENSOR.ATTN_K,
|
||||
MODEL_TENSOR.ATTN_V,
|
||||
MODEL_TENSOR.ATTN_SINKS,
|
||||
MODEL_TENSOR.ATTN_OUT,
|
||||
MODEL_TENSOR.FFN_NORM,
|
||||
MODEL_TENSOR.FFN_GATE,
|
||||
MODEL_TENSOR.FFN_DOWN,
|
||||
MODEL_TENSOR.FFN_UP,
|
||||
MODEL_TENSOR.FFN_GATE_INP,
|
||||
MODEL_TENSOR.FFN_GATE_EXP,
|
||||
MODEL_TENSOR.FFN_DOWN_EXP,
|
||||
MODEL_TENSOR.FFN_UP_EXP,
|
||||
MODEL_TENSOR.FFN_EXP_PROBS_B,
|
||||
],
|
||||
MODEL_ARCH.LLAMA_EMBED: [
|
||||
MODEL_TENSOR.TOKEN_EMBD,
|
||||
MODEL_TENSOR.OUTPUT_NORM,
|
||||
|
|
@ -3217,7 +3239,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
|
|||
MODEL_TENSOR.FFN_GATE_EXP,
|
||||
MODEL_TENSOR.FFN_DOWN_EXP,
|
||||
MODEL_TENSOR.FFN_UP_EXP,
|
||||
]
|
||||
],
|
||||
# TODO
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -320,6 +320,7 @@ class TensorNameMap:
|
|||
|
||||
MODEL_TENSOR.ATTN_SINKS: (
|
||||
"model.layers.{bid}.self_attn.sinks", # openai-moe
|
||||
"model.layers.{bid}.self_attn.attention_sink_bias", # mimov2
|
||||
),
|
||||
|
||||
MODEL_TENSOR.ATTN_GATE: (
|
||||
|
|
|
|||
|
|
@ -115,6 +115,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
|
|||
{ LLM_ARCH_RND1, "rnd1" },
|
||||
{ LLM_ARCH_PANGU_EMBED, "pangu-embedded" },
|
||||
{ LLM_ARCH_MISTRAL3, "mistral3" },
|
||||
{ LLM_ARCH_MIMO2, "mimo2" },
|
||||
{ LLM_ARCH_LLAMA_EMBED, "llama-embed" },
|
||||
{ LLM_ARCH_UNKNOWN, "(unknown)" },
|
||||
};
|
||||
|
|
@ -2190,6 +2191,27 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
|
|||
LLM_TENSOR_VISEXP_FFN_DOWN,
|
||||
LLM_TENSOR_VISEXP_FFN_UP,
|
||||
};
|
||||
case LLM_ARCH_MIMO2:
|
||||
return {
|
||||
LLM_TENSOR_TOKEN_EMBD,
|
||||
LLM_TENSOR_OUTPUT_NORM,
|
||||
LLM_TENSOR_OUTPUT,
|
||||
LLM_TENSOR_ATTN_NORM,
|
||||
LLM_TENSOR_ATTN_Q,
|
||||
LLM_TENSOR_ATTN_K,
|
||||
LLM_TENSOR_ATTN_V,
|
||||
LLM_TENSOR_ATTN_SINKS,
|
||||
LLM_TENSOR_ATTN_OUT,
|
||||
LLM_TENSOR_FFN_NORM,
|
||||
LLM_TENSOR_FFN_GATE,
|
||||
LLM_TENSOR_FFN_DOWN,
|
||||
LLM_TENSOR_FFN_UP,
|
||||
LLM_TENSOR_FFN_GATE_INP,
|
||||
LLM_TENSOR_FFN_GATE_EXPS,
|
||||
LLM_TENSOR_FFN_DOWN_EXPS,
|
||||
LLM_TENSOR_FFN_UP_EXPS,
|
||||
LLM_TENSOR_FFN_EXP_PROBS_B,
|
||||
};
|
||||
case LLM_ARCH_GPTJ:
|
||||
case LLM_ARCH_UNKNOWN:
|
||||
return {
|
||||
|
|
|
|||
|
|
@ -119,6 +119,7 @@ enum llm_arch {
|
|||
LLM_ARCH_RND1,
|
||||
LLM_ARCH_PANGU_EMBED,
|
||||
LLM_ARCH_MISTRAL3,
|
||||
LLM_ARCH_MIMO2,
|
||||
LLM_ARCH_LLAMA_EMBED,
|
||||
LLM_ARCH_UNKNOWN,
|
||||
};
|
||||
|
|
|
|||
|
|
@ -123,10 +123,11 @@ struct llama_hparams {
|
|||
llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
|
||||
// the size of the sliding window (0 - no SWA)
|
||||
uint32_t n_swa = 0;
|
||||
// if swa_layers[il] == true, then layer il is SWA
|
||||
// if swa_layers[il] == false, then layer il is dense (i.e. non-SWA)
|
||||
// if swa_layers[il] == 1, then layer il is SWA
|
||||
// if swa_layers[il] == 0, then layer il is dense (i.e. non-SWA)
|
||||
// by default, all layers are dense
|
||||
std::array<bool, LLAMA_MAX_LAYERS> swa_layers;
|
||||
// note: using uint32_t type for compatibility reason
|
||||
std::array<uint32_t, LLAMA_MAX_LAYERS> swa_layers;
|
||||
|
||||
// for State Space Models
|
||||
uint32_t ssm_d_conv = 0;
|
||||
|
|
|
|||
|
|
@ -79,6 +79,7 @@
|
|||
#include "models/llama-iswa.cpp"
|
||||
#include "models/llama.cpp"
|
||||
#include "models/mamba.cpp"
|
||||
#include "models/mimo2-iswa.cpp"
|
||||
#include "models/minicpm3.cpp"
|
||||
#include "models/minimax-m2.cpp"
|
||||
#include "models/modern-bert.cpp"
|
||||
|
|
@ -236,6 +237,7 @@ const char * llm_type_name(llm_type type) {
|
|||
case LLM_TYPE_230B_A10B: return "230B.A10B";
|
||||
case LLM_TYPE_235B_A22B: return "235B.A22B";
|
||||
case LLM_TYPE_300B_A47B: return "300B.A47B";
|
||||
case LLM_TYPE_310B_A15B: return "310B.A15B";
|
||||
case LLM_TYPE_355B_A32B: return "355B.A32B";
|
||||
case LLM_TYPE_E2B: return "E2B";
|
||||
case LLM_TYPE_E4B: return "E4B";
|
||||
|
|
@ -2445,6 +2447,22 @@ void llama_model::load_hparams(llama_model_loader & ml) {
|
|||
default: type = LLM_TYPE_UNKNOWN;
|
||||
}
|
||||
} break;
|
||||
case LLM_ARCH_MIMO2:
|
||||
{
|
||||
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
|
||||
|
||||
hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
|
||||
|
||||
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
|
||||
ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
|
||||
ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa);
|
||||
ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.swa_layers, hparams.n_layer);
|
||||
|
||||
switch (hparams.n_layer) {
|
||||
case 48: type = LLM_TYPE_310B_A15B; break;
|
||||
default: type = LLM_TYPE_UNKNOWN;
|
||||
}
|
||||
} break;
|
||||
default: throw std::runtime_error("unsupported model architecture");
|
||||
}
|
||||
|
||||
|
|
@ -6807,6 +6825,44 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
|
|||
layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { hparams.n_ff_shexp, n_embd }, 0);
|
||||
}
|
||||
} break;
|
||||
case LLM_ARCH_MIMO2:
|
||||
{
|
||||
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}, 0);
|
||||
|
||||
for (int i = 0; i < n_layer; ++i) {
|
||||
auto & layer = layers[i];
|
||||
uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i);
|
||||
uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i);
|
||||
uint32_t n_head = hparams.n_head(i);
|
||||
|
||||
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head }, 0);
|
||||
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, 0);
|
||||
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_v_gqa }, 0);
|
||||
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_v * n_head, n_embd }, 0);
|
||||
|
||||
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
|
||||
layer.attn_sinks = create_tensor(tn(LLM_TENSOR_ATTN_SINKS, "weight", i), {n_head}, TENSOR_NOT_REQUIRED);
|
||||
|
||||
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
|
||||
|
||||
// non-MoE branch
|
||||
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);
|
||||
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, TENSOR_NOT_REQUIRED);
|
||||
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);
|
||||
|
||||
// MoE branch
|
||||
int64_t n_ff_exp = hparams.n_ff_exp;
|
||||
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED);
|
||||
layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED);
|
||||
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, TENSOR_NOT_REQUIRED);
|
||||
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED);
|
||||
layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
|
||||
}
|
||||
} break;
|
||||
default:
|
||||
throw std::runtime_error("unknown architecture");
|
||||
}
|
||||
|
|
@ -7873,6 +7929,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
|
|||
{
|
||||
llm = std::make_unique<llm_build_mistral3>(*this, params);
|
||||
} break;
|
||||
case LLM_ARCH_MIMO2:
|
||||
{
|
||||
llm = std::make_unique<llm_build_mimo2_iswa>(*this, params);
|
||||
} break;
|
||||
default:
|
||||
GGML_ABORT("fatal error");
|
||||
}
|
||||
|
|
@ -8103,6 +8163,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
|
|||
case LLM_ARCH_PANGU_EMBED:
|
||||
case LLM_ARCH_AFMOE:
|
||||
case LLM_ARCH_QWEN3NEXT:
|
||||
case LLM_ARCH_MIMO2:
|
||||
return LLAMA_ROPE_TYPE_NEOX;
|
||||
|
||||
case LLM_ARCH_QWEN2VL:
|
||||
|
|
|
|||
|
|
@ -123,6 +123,7 @@ enum llm_type {
|
|||
LLM_TYPE_230B_A10B, // Minimax M2
|
||||
LLM_TYPE_235B_A22B,
|
||||
LLM_TYPE_300B_A47B, // Ernie MoE big
|
||||
LLM_TYPE_310B_A15B, // /MiMo-V2-Flash
|
||||
LLM_TYPE_355B_A32B, // GLM-4.5
|
||||
LLM_TYPE_E2B,
|
||||
LLM_TYPE_E4B,
|
||||
|
|
|
|||
|
|
@ -205,12 +205,11 @@ static void llama_params_fit_impl(
|
|||
}
|
||||
}
|
||||
|
||||
int64_t sum_total = 0;
|
||||
int64_t sum_free = 0;
|
||||
int64_t sum_projected_free = 0;
|
||||
int64_t min_projected_free = INT64_MAX;
|
||||
int64_t sum_projected_used = 0;
|
||||
int64_t sum_projected_model = 0;
|
||||
int64_t sum_projected_ctx = 0;
|
||||
|
||||
if (nd > 1) {
|
||||
LLAMA_LOG_INFO("%s: projected memory use with initial parameters [MiB]:\n", __func__);
|
||||
|
|
@ -221,12 +220,11 @@ static void llama_params_fit_impl(
|
|||
const int64_t projected_used = dmd.mb.total();
|
||||
const int64_t projected_free = dmd.free - projected_used;
|
||||
|
||||
sum_total += dmd.total;
|
||||
sum_free += dmd.free;
|
||||
sum_projected_used += projected_used;
|
||||
sum_projected_free += projected_free;
|
||||
min_projected_free = std::min(min_projected_free, projected_free);
|
||||
sum_projected_model += dmd.mb.model;
|
||||
sum_projected_ctx += dmd.mb.context;
|
||||
|
||||
if (nd > 1) {
|
||||
LLAMA_LOG_INFO("%s: - %s: %6" PRId64 " total, %6" PRId64 " used, %6" PRId64 " %s\n",
|
||||
|
|
@ -234,10 +232,9 @@ static void llama_params_fit_impl(
|
|||
projected_free >= 0 ? "surplus" : "deficit");
|
||||
}
|
||||
}
|
||||
assert(sum_total >= 0 && sum_projected_used >= 0 && sum_projected_ctx >= 0);
|
||||
assert(sum_projected_used >= sum_projected_ctx);
|
||||
assert(sum_free >= 0 && sum_projected_used >= 0);
|
||||
LLAMA_LOG_INFO("%s: projected to use %" PRId64 " MiB of device memory vs. %" PRId64 " MiB of free device memory\n",
|
||||
__func__, sum_projected_used/MiB, sum_total/MiB);
|
||||
__func__, sum_projected_used/MiB, sum_free/MiB);
|
||||
if (min_projected_free >= margin) {
|
||||
if (nd == 1) {
|
||||
LLAMA_LOG_INFO("%s: will leave %" PRId64 " >= %" PRId64 " MiB of free device memory, no changes needed\n",
|
||||
|
|
@ -260,9 +257,7 @@ static void llama_params_fit_impl(
|
|||
__func__, margin/MiB, -global_surplus/MiB);
|
||||
if (cparams->n_ctx == 0) {
|
||||
if (hp_nct > n_ctx_min) {
|
||||
const int64_t bytes_per_ctx = sum_projected_ctx / hp_nct;
|
||||
|
||||
int64_t memory_reduction = -global_surplus;
|
||||
int64_t sum_used_target = sum_free - nd*margin_s;
|
||||
if (nd > 1) {
|
||||
// for multiple devices we need to be more conservative in terms of how much context we think can fit:
|
||||
// - for dense models only whole layers can be assigned to devices
|
||||
|
|
@ -270,24 +265,34 @@ static void llama_params_fit_impl(
|
|||
// - on average we expect a waste of 0.5 layers/tensors per device
|
||||
// - use slightly more than the expected average for nd devices to be safe
|
||||
const int64_t model_per_layer = sum_projected_model / std::min(uint32_t(mparams->n_gpu_layers), hp_ngl);
|
||||
memory_reduction += (nd + 1) * model_per_layer / (hp_nex == 0 ? 2 : 6);
|
||||
sum_used_target -= (nd + 1) * model_per_layer / (hp_nex == 0 ? 2 : 6);
|
||||
}
|
||||
|
||||
uint32_t ctx_reduction = std::min(uint32_t((memory_reduction + bytes_per_ctx - 1) / bytes_per_ctx), hp_nct - n_ctx_min);
|
||||
cparams->n_ctx = hp_nct - ctx_reduction;
|
||||
cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % 256, n_ctx_min); // round down context for CUDA backend
|
||||
int64_t sum_projected_used_min_ctx = 0;
|
||||
cparams->n_ctx = n_ctx_min;
|
||||
const dmds_t dmds_min_ctx = llama_get_device_memory_data(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
|
||||
for (const auto & dmd : dmds_min_ctx) {
|
||||
sum_projected_used_min_ctx += dmd.mb.total();
|
||||
}
|
||||
if (sum_used_target > sum_projected_used_min_ctx) {
|
||||
// linear interpolation between minimum and maximum context size:
|
||||
cparams->n_ctx += (hp_nct - n_ctx_min) * (sum_used_target - sum_projected_used_min_ctx)
|
||||
/ (sum_projected_used - sum_projected_used_min_ctx);
|
||||
cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % 256, n_ctx_min); // round down context for CUDA backend
|
||||
|
||||
ctx_reduction = hp_nct - cparams->n_ctx;
|
||||
memory_reduction = ctx_reduction * bytes_per_ctx;
|
||||
global_surplus += memory_reduction;
|
||||
LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
|
||||
__func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
|
||||
if (global_surplus >= 0) {
|
||||
const int64_t bytes_per_ctx = (sum_projected_used - sum_projected_used_min_ctx) / (hp_nct - n_ctx_min);
|
||||
const int64_t memory_reduction = (hp_nct - cparams->n_ctx) * bytes_per_ctx;
|
||||
LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
|
||||
__func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
|
||||
if (nd == 1) {
|
||||
LLAMA_LOG_INFO("%s: entire model can be fit by reducing context\n", __func__);
|
||||
return;
|
||||
}
|
||||
LLAMA_LOG_INFO("%s: entire model should be fit across devices by reducing context\n", __func__);
|
||||
} else {
|
||||
const int64_t memory_reduction = sum_projected_used - sum_projected_used_min_ctx;
|
||||
LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
|
||||
__func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
|
||||
}
|
||||
} else {
|
||||
LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
|
||||
|
|
|
|||
123
src/models/mimo2-iswa.cpp
Normal file
123
src/models/mimo2-iswa.cpp
Normal file
|
|
@ -0,0 +1,123 @@
|
|||
|
||||
#include "models.h"
|
||||
|
||||
llm_build_mimo2_iswa::llm_build_mimo2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
|
||||
ggml_tensor * cur;
|
||||
ggml_tensor * inpL;
|
||||
|
||||
inpL = build_inp_embd(model.tok_embd);
|
||||
|
||||
ggml_tensor * inp_pos = build_inp_pos();
|
||||
auto * inp_attn = build_attn_inp_kv_iswa();
|
||||
ggml_tensor * inp_out_ids = build_inp_out_ids();
|
||||
|
||||
for (int il = 0; il < n_layer; ++il) {
|
||||
ggml_tensor * inpSA = inpL;
|
||||
|
||||
uint32_t n_head_l = hparams.n_head(il);
|
||||
uint32_t n_head_kv_l = hparams.n_head_kv(il);
|
||||
const float freq_base_l = model.get_rope_freq_base(cparams, il);
|
||||
const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
|
||||
|
||||
cur = inpL;
|
||||
|
||||
// self_attention
|
||||
{
|
||||
cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
|
||||
cb(cur, "attn_norm", il);
|
||||
|
||||
// compute Q and K and RoPE them
|
||||
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
|
||||
cb(Qcur, "Qcur", il);
|
||||
|
||||
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
|
||||
cb(Kcur, "Kcur", il);
|
||||
|
||||
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
|
||||
cb(Vcur, "Vcur", il);
|
||||
|
||||
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head_l, n_tokens);
|
||||
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv_l, n_tokens);
|
||||
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv_l, n_tokens);
|
||||
|
||||
Qcur = ggml_rope_ext(
|
||||
ctx0, Qcur, inp_pos, nullptr,
|
||||
n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
|
||||
ext_factor, attn_factor, beta_fast, beta_slow
|
||||
);
|
||||
|
||||
Kcur = ggml_rope_ext(
|
||||
ctx0, Kcur, inp_pos, nullptr,
|
||||
n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
|
||||
ext_factor, attn_factor, beta_fast, beta_slow
|
||||
);
|
||||
|
||||
cb(Qcur, "Qcur", il);
|
||||
cb(Kcur, "Kcur", il);
|
||||
cb(Vcur, "Vcur", il);
|
||||
|
||||
ggml_tensor * sinks = model.layers[il].attn_sinks;
|
||||
|
||||
cur = build_attn(inp_attn,
|
||||
model.layers[il].wo, NULL,
|
||||
Qcur, Kcur, Vcur, nullptr, sinks, nullptr, 1.0f/sqrtf(float(n_embd_head_k)), il);
|
||||
}
|
||||
|
||||
if (il == n_layer - 1 && inp_out_ids) {
|
||||
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
|
||||
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
|
||||
}
|
||||
|
||||
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
|
||||
cb(ffn_inp, "ffn_inp", il);
|
||||
|
||||
cur = build_norm(ffn_inp,
|
||||
model.layers[il].ffn_norm, NULL,
|
||||
LLM_NORM_RMS, il);
|
||||
cb(cur, "ffn_norm", il);
|
||||
|
||||
// feed-forward network
|
||||
if (model.layers[il].ffn_gate_inp == nullptr) {
|
||||
// dense branch
|
||||
cur = build_ffn(cur,
|
||||
model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
|
||||
model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
|
||||
model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
|
||||
NULL,
|
||||
LLM_FFN_SILU, LLM_FFN_PAR, il);
|
||||
cb(cur, "ffn_out", il);
|
||||
} else {
|
||||
// MoE branch
|
||||
cur = build_moe_ffn(cur, model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps,
|
||||
model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
|
||||
model.layers[il].ffn_exp_probs_b, n_expert, n_expert_used, LLM_FFN_SILU, true, false,
|
||||
0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID, il);
|
||||
cb(cur, "ffn_moe_out", il);
|
||||
}
|
||||
|
||||
cur = ggml_add(ctx0, cur, ffn_inp);
|
||||
|
||||
cur = build_cvec(cur, il);
|
||||
cb(cur, "l_out", il);
|
||||
|
||||
// input for next layer
|
||||
inpL = cur;
|
||||
}
|
||||
|
||||
cur = inpL;
|
||||
|
||||
cur = build_norm(cur,
|
||||
model.output_norm, NULL,
|
||||
LLM_NORM_RMS, -1);
|
||||
|
||||
cb(cur, "result_norm", -1);
|
||||
res->t_embd = cur;
|
||||
|
||||
// lm_head
|
||||
cur = build_lora_mm(model.output, cur);
|
||||
|
||||
cb(cur, "result_output", -1);
|
||||
res->t_logits = cur;
|
||||
|
||||
ggml_build_forward_expand(gf, cur);
|
||||
}
|
||||
|
|
@ -316,6 +316,10 @@ struct llm_build_mamba : public llm_graph_context_mamba {
|
|||
llm_build_mamba(const llama_model & model, const llm_graph_params & params);
|
||||
};
|
||||
|
||||
struct llm_build_mimo2_iswa : public llm_graph_context {
|
||||
llm_build_mimo2_iswa(const llama_model & model, const llm_graph_params & params);
|
||||
};
|
||||
|
||||
struct llm_build_minicpm3 : public llm_graph_context {
|
||||
llm_build_minicpm3(const llama_model & model, const llm_graph_params & params);
|
||||
};
|
||||
|
|
|
|||
|
|
@ -1007,8 +1007,10 @@ private:
|
|||
return ret;
|
||||
}
|
||||
|
||||
void clear_slot(server_slot & slot) const {
|
||||
GGML_ASSERT(!slot.is_processing());
|
||||
void clear_slot(server_slot & slot, bool allow_processing = false) const {
|
||||
if (!allow_processing) {
|
||||
GGML_ASSERT(!slot.is_processing());
|
||||
}
|
||||
|
||||
SLT_WRN(slot, "clearing slot with %zu tokens\n", slot.prompt.tokens.size());
|
||||
|
||||
|
|
@ -2336,7 +2338,7 @@ private:
|
|||
if (!llama_memory_seq_rm(llama_get_memory(ctx), slot.id, p0, -1)) {
|
||||
SLT_WRN(slot, "failed to truncate tokens with position >= %d - clearing the memory\n", p0);
|
||||
|
||||
clear_slot(slot);
|
||||
clear_slot(slot, /*allow_processing=*/true);
|
||||
|
||||
// there is no common part left
|
||||
slot.n_prompt_tokens_cache = 0;
|
||||
|
|
|
|||
|
|
@ -34,6 +34,8 @@
|
|||
#include <limits.h>
|
||||
#endif
|
||||
|
||||
#define DEFAULT_STOP_TIMEOUT 10 // seconds
|
||||
|
||||
#define CMD_ROUTER_TO_CHILD_EXIT "cmd_router_to_child:exit"
|
||||
#define CMD_CHILD_TO_ROUTER_READY "cmd_child_to_router:ready"
|
||||
|
||||
|
|
@ -203,13 +205,14 @@ void server_models::load_models() {
|
|||
// convert presets to server_model_meta and add to mapping
|
||||
for (const auto & preset : final_presets) {
|
||||
server_model_meta meta{
|
||||
/* preset */ preset.second,
|
||||
/* name */ preset.first,
|
||||
/* port */ 0,
|
||||
/* status */ SERVER_MODEL_STATUS_UNLOADED,
|
||||
/* last_used */ 0,
|
||||
/* args */ std::vector<std::string>(),
|
||||
/* exit_code */ 0
|
||||
/* preset */ preset.second,
|
||||
/* name */ preset.first,
|
||||
/* port */ 0,
|
||||
/* status */ SERVER_MODEL_STATUS_UNLOADED,
|
||||
/* last_used */ 0,
|
||||
/* args */ std::vector<std::string>(),
|
||||
/* exit_code */ 0,
|
||||
/* stop_timeout */ DEFAULT_STOP_TIMEOUT,
|
||||
};
|
||||
add_model(std::move(meta));
|
||||
}
|
||||
|
|
@ -227,6 +230,20 @@ void server_models::load_models() {
|
|||
}
|
||||
}
|
||||
|
||||
// handle custom stop-timeout option
|
||||
for (auto & [name, inst] : mapping) {
|
||||
std::string val;
|
||||
if (inst.meta.preset.get_option(COMMON_ARG_PRESET_STOP_TIMEOUT, val)) {
|
||||
try {
|
||||
inst.meta.stop_timeout = std::stoi(val);
|
||||
} catch (...) {
|
||||
SRV_WRN("invalid stop-timeout value '%s' for model '%s', using default %d seconds\n",
|
||||
val.c_str(), name.c_str(), DEFAULT_STOP_TIMEOUT);
|
||||
inst.meta.stop_timeout = DEFAULT_STOP_TIMEOUT;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// load any autoload models
|
||||
std::vector<std::string> models_to_load;
|
||||
for (const auto & [name, inst] : mapping) {
|
||||
|
|
@ -362,7 +379,7 @@ void server_models::unload_lru() {
|
|||
int64_t lru_last_used = ggml_time_ms();
|
||||
size_t count_active = 0;
|
||||
{
|
||||
std::lock_guard<std::mutex> lk(mutex);
|
||||
std::unique_lock<std::mutex> lk(mutex);
|
||||
for (const auto & m : mapping) {
|
||||
if (m.second.meta.is_active()) {
|
||||
count_active++;
|
||||
|
|
@ -376,6 +393,13 @@ void server_models::unload_lru() {
|
|||
if (!lru_model_name.empty() && count_active >= (size_t)base_params.models_max) {
|
||||
SRV_INF("models_max limit reached, removing LRU name=%s\n", lru_model_name.c_str());
|
||||
unload(lru_model_name);
|
||||
// wait for unload to complete
|
||||
{
|
||||
std::unique_lock<std::mutex> lk(mutex);
|
||||
cv.wait(lk, [this, &lru_model_name]() {
|
||||
return mapping[lru_model_name].meta.status == SERVER_MODEL_STATUS_UNLOADED;
|
||||
});
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -436,38 +460,83 @@ void server_models::load(const std::string & name) {
|
|||
|
||||
// start a thread to manage the child process
|
||||
// captured variables are guaranteed to be destroyed only after the thread is joined
|
||||
inst.th = std::thread([this, name, child_proc = inst.subproc, port = inst.meta.port]() {
|
||||
// read stdout/stderr and forward to main server log
|
||||
bool state_received = false; // true if child state received
|
||||
FILE * p_stdout_stderr = subprocess_stdout(child_proc.get());
|
||||
if (p_stdout_stderr) {
|
||||
char buffer[4096];
|
||||
while (fgets(buffer, sizeof(buffer), p_stdout_stderr) != nullptr) {
|
||||
LOG("[%5d] %s", port, buffer);
|
||||
if (!state_received && std::strstr(buffer, CMD_CHILD_TO_ROUTER_READY) != nullptr) {
|
||||
// child process is ready
|
||||
this->update_status(name, SERVER_MODEL_STATUS_LOADED);
|
||||
state_received = true;
|
||||
inst.th = std::thread([this, name, child_proc = inst.subproc, port = inst.meta.port, stop_timeout = inst.meta.stop_timeout]() {
|
||||
FILE * stdin_file = subprocess_stdin(child_proc.get());
|
||||
FILE * stdout_file = subprocess_stdout(child_proc.get()); // combined stdout/stderr
|
||||
|
||||
std::thread log_thread([&]() {
|
||||
// read stdout/stderr and forward to main server log
|
||||
// also handle status report from child process
|
||||
bool state_received = false; // true if child state received
|
||||
if (stdout_file) {
|
||||
char buffer[4096];
|
||||
while (fgets(buffer, sizeof(buffer), stdout_file) != nullptr) {
|
||||
LOG("[%5d] %s", port, buffer);
|
||||
if (!state_received && std::strstr(buffer, CMD_CHILD_TO_ROUTER_READY) != nullptr) {
|
||||
// child process is ready
|
||||
this->update_status(name, SERVER_MODEL_STATUS_LOADED, 0);
|
||||
state_received = true;
|
||||
}
|
||||
}
|
||||
} else {
|
||||
SRV_ERR("failed to get stdout/stderr of child process for name=%s\n", name.c_str());
|
||||
}
|
||||
} else {
|
||||
SRV_ERR("failed to get stdout/stderr of child process for name=%s\n", name.c_str());
|
||||
}
|
||||
});
|
||||
|
||||
std::thread stopping_thread([&]() {
|
||||
// thread to monitor stopping signal
|
||||
auto is_stopping = [this, &name]() {
|
||||
return this->stopping_models.find(name) != this->stopping_models.end();
|
||||
};
|
||||
{
|
||||
std::unique_lock<std::mutex> lk(this->mutex);
|
||||
this->cv_stop.wait(lk, is_stopping);
|
||||
}
|
||||
SRV_INF("stopping model instance name=%s\n", name.c_str());
|
||||
// send interrupt to child process
|
||||
fprintf(stdin_file, "%s\n", CMD_ROUTER_TO_CHILD_EXIT);
|
||||
fflush(stdin_file);
|
||||
// wait to stop gracefully or timeout
|
||||
int64_t start_time = ggml_time_ms();
|
||||
while (true) {
|
||||
std::unique_lock<std::mutex> lk(this->mutex);
|
||||
if (!is_stopping()) {
|
||||
return; // already stopped
|
||||
}
|
||||
int64_t elapsed = ggml_time_ms() - start_time;
|
||||
if (elapsed >= stop_timeout * 1000) {
|
||||
// timeout, force kill
|
||||
SRV_WRN("force-killing model instance name=%s after %d seconds timeout\n", name.c_str(), stop_timeout);
|
||||
subprocess_terminate(child_proc.get());
|
||||
return;
|
||||
}
|
||||
this->cv_stop.wait_for(lk, std::chrono::seconds(1));
|
||||
}
|
||||
});
|
||||
|
||||
// we reach here when the child process exits
|
||||
// note: we cannot join() prior to this point because it will close stdin_file
|
||||
if (log_thread.joinable()) {
|
||||
log_thread.join();
|
||||
}
|
||||
|
||||
// stop the timeout monitoring thread
|
||||
{
|
||||
std::lock_guard<std::mutex> lk(this->mutex);
|
||||
stopping_models.erase(name);
|
||||
cv_stop.notify_all();
|
||||
}
|
||||
if (stopping_thread.joinable()) {
|
||||
stopping_thread.join();
|
||||
}
|
||||
|
||||
// get the exit code
|
||||
int exit_code = 0;
|
||||
subprocess_join(child_proc.get(), &exit_code);
|
||||
subprocess_destroy(child_proc.get());
|
||||
// update PID and status
|
||||
{
|
||||
std::lock_guard<std::mutex> lk(mutex);
|
||||
auto it = mapping.find(name);
|
||||
if (it != mapping.end()) {
|
||||
auto & meta = it->second.meta;
|
||||
meta.exit_code = exit_code;
|
||||
meta.status = SERVER_MODEL_STATUS_UNLOADED;
|
||||
}
|
||||
cv.notify_all();
|
||||
}
|
||||
|
||||
// update status and exit code
|
||||
this->update_status(name, SERVER_MODEL_STATUS_UNLOADED, exit_code);
|
||||
SRV_INF("instance name=%s exited with status %d\n", name.c_str(), exit_code);
|
||||
});
|
||||
|
||||
|
|
@ -488,22 +557,14 @@ void server_models::load(const std::string & name) {
|
|||
cv.notify_all();
|
||||
}
|
||||
|
||||
static void interrupt_subprocess(FILE * stdin_file) {
|
||||
// because subprocess.h does not provide a way to send SIGINT,
|
||||
// we will send a command to the child process to exit gracefully
|
||||
if (stdin_file) {
|
||||
fprintf(stdin_file, "%s\n", CMD_ROUTER_TO_CHILD_EXIT);
|
||||
fflush(stdin_file);
|
||||
}
|
||||
}
|
||||
|
||||
void server_models::unload(const std::string & name) {
|
||||
std::lock_guard<std::mutex> lk(mutex);
|
||||
auto it = mapping.find(name);
|
||||
if (it != mapping.end()) {
|
||||
if (it->second.meta.is_active()) {
|
||||
SRV_INF("unloading model instance name=%s\n", name.c_str());
|
||||
interrupt_subprocess(it->second.stdin_file);
|
||||
stopping_models.insert(name);
|
||||
cv_stop.notify_all();
|
||||
// status change will be handled by the managing thread
|
||||
} else {
|
||||
SRV_WRN("model instance name=%s is not loaded\n", name.c_str());
|
||||
|
|
@ -518,7 +579,8 @@ void server_models::unload_all() {
|
|||
for (auto & [name, inst] : mapping) {
|
||||
if (inst.meta.is_active()) {
|
||||
SRV_INF("unloading model instance name=%s\n", name.c_str());
|
||||
interrupt_subprocess(inst.stdin_file);
|
||||
stopping_models.insert(name);
|
||||
cv_stop.notify_all();
|
||||
// status change will be handled by the managing thread
|
||||
}
|
||||
// moving the thread to join list to avoid deadlock
|
||||
|
|
@ -532,16 +594,15 @@ void server_models::unload_all() {
|
|||
}
|
||||
}
|
||||
|
||||
void server_models::update_status(const std::string & name, server_model_status status) {
|
||||
// for now, we only allow updating to LOADED status
|
||||
if (status != SERVER_MODEL_STATUS_LOADED) {
|
||||
throw std::runtime_error("invalid status value");
|
||||
}
|
||||
auto meta = get_meta(name);
|
||||
if (meta.has_value()) {
|
||||
meta->status = status;
|
||||
update_meta(name, meta.value());
|
||||
void server_models::update_status(const std::string & name, server_model_status status, int exit_code) {
|
||||
std::unique_lock<std::mutex> lk(mutex);
|
||||
auto it = mapping.find(name);
|
||||
if (it != mapping.end()) {
|
||||
auto & meta = it->second.meta;
|
||||
meta.status = status;
|
||||
meta.exit_code = exit_code;
|
||||
}
|
||||
cv.notify_all();
|
||||
}
|
||||
|
||||
void server_models::wait_until_loaded(const std::string & name) {
|
||||
|
|
@ -568,6 +629,7 @@ bool server_models::ensure_model_loaded(const std::string & name) {
|
|||
load(name);
|
||||
}
|
||||
|
||||
// for loading state
|
||||
SRV_INF("waiting until model name=%s is fully loaded...\n", name.c_str());
|
||||
wait_until_loaded(name);
|
||||
|
||||
|
|
@ -795,7 +857,7 @@ void server_models_routes::init_routes() {
|
|||
res_err(res, format_error_response("model is not found", ERROR_TYPE_INVALID_REQUEST));
|
||||
return res;
|
||||
}
|
||||
if (model->status != SERVER_MODEL_STATUS_LOADED) {
|
||||
if (!model->is_active()) {
|
||||
res_err(res, format_error_response("model is not loaded", ERROR_TYPE_INVALID_REQUEST));
|
||||
return res;
|
||||
}
|
||||
|
|
|
|||
|
|
@ -9,6 +9,7 @@
|
|||
#include <condition_variable>
|
||||
#include <functional>
|
||||
#include <memory>
|
||||
#include <set>
|
||||
|
||||
/**
|
||||
* state diagram:
|
||||
|
|
@ -56,6 +57,7 @@ struct server_model_meta {
|
|||
int64_t last_used = 0; // for LRU unloading
|
||||
std::vector<std::string> args; // args passed to the model instance, will be populated by render_args()
|
||||
int exit_code = 0; // exit code of the model instance process (only valid if status == FAILED)
|
||||
int stop_timeout = 0; // seconds to wait before force-killing the model instance during shutdown
|
||||
|
||||
bool is_active() const {
|
||||
return status == SERVER_MODEL_STATUS_LOADED || status == SERVER_MODEL_STATUS_LOADING;
|
||||
|
|
@ -83,6 +85,10 @@ private:
|
|||
std::condition_variable cv;
|
||||
std::map<std::string, instance_t> mapping;
|
||||
|
||||
// for stopping models
|
||||
std::condition_variable cv_stop;
|
||||
std::set<std::string> stopping_models;
|
||||
|
||||
common_preset_context ctx_preset;
|
||||
|
||||
common_params base_params;
|
||||
|
|
@ -119,7 +125,7 @@ public:
|
|||
void unload_all();
|
||||
|
||||
// update the status of a model instance (thread-safe)
|
||||
void update_status(const std::string & name, server_model_status status);
|
||||
void update_status(const std::string & name, server_model_status status, int exit_code);
|
||||
|
||||
// wait until the model instance is fully loaded (thread-safe)
|
||||
// return when the model is loaded or failed to load
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue