Merge branch 'fix_precision_MLA' of https://github.com/kvcache-ai/ktransformers into server-prefix-cache

ktransformers/local_chat.py

@@ -30,6 +30,7 @@ from ktransformers.models.modeling_llama import LlamaForCausalLM
 from ktransformers.models.modeling_mixtral import MixtralForCausalLM
 from ktransformers.util.utils import prefill_and_generate
 from ktransformers.server.config.config import Config
+from ktransformers.operators.flashinfer_wrapper import flashinfer_enabled
 
 custom_models = {
     "DeepseekV2ForCausalLM": DeepseekV2ForCausalLM,
@@ -170,9 +171,16 @@ def local_chat(
         torch.set_default_dtype(
             torch.bfloat16
         )  # TODO: Remove this, replace dtype using config
-        generated = prefill_and_generate(
+
-            model, tokenizer, input_tensor.cuda(), max_new_tokens, use_cuda_graph, mode, force_think
+        if system != "Windows" and (config.architectures[0] == "DeepseekV2ForCausalLM" or "DeepseekV3ForCausalLM") and flashinfer_enabled:
-        )
+            generated = prefill_and_generate(
+                model, tokenizer, input_tensor.cuda(), max_new_tokens, use_cuda_graph, mode = mode, force_think = force_think,
+                use_flashinfer_mla = True, num_heads = config.num_attention_heads, head_dim_ckv = config.kv_lora_rank, head_dim_kpe = config.qk_rope_head_dim, q_head_dim = config.qk_rope_head_dim + config.qk_nope_head_dim
+            )
+        else:
+            generated = prefill_and_generate(
+                model, tokenizer, input_tensor.cuda(), max_new_tokens, use_cuda_graph, mode = mode, force_think = force_think,
+            )
 
 
 if __name__ == "__main__":

ktransformers/models/custom_cache.py

@@ -138,8 +138,6 @@ class StaticCache(transformers.StaticCache):
             page_idx = cache_position // self.page_size
             page_offset = cache_position % self.page_size
             # key shape (self.max_pages, self.page_size, 1, config.kv_lora_rank + config.qk_rope_head_dim)
-            #print("page_idx", page_idx)
-            #print("page_offset", page_offset)
             k_out[page_idx, page_offset, :, :self.kv_lora_rank] = key_states
             k_out[page_idx, page_offset, :, self.kv_lora_rank:] = value_states
             return k_out, self.page_table_list[layer_idx]

ktransformers/operators/RoPE.py

@@ -42,7 +42,7 @@ class RotaryEmbedding(BaseInjectedModule, DeepseekV2RotaryEmbedding):
         **kwargs,
     ):
         BaseInjectedModule.__init__(
-            self, key, gguf_loader, config, orig_module, generate_device, **kwargs
+            self, key, gguf_loader, config, orig_module, prefill_device, generate_device, **kwargs
         )
         self.orig_module.__init__(
             orig_module.dim, orig_module.max_position_embeddings, orig_module.base
@@ -72,7 +72,7 @@ class RotaryEmbeddingV3(BaseInjectedModule):
         **kwargs,
     ):
         BaseInjectedModule.__init__(
-            self, key, gguf_loader, config, orig_module, generate_device, **kwargs
+            self, key, gguf_loader, config, orig_module, prefill_device, generate_device, **kwargs
         )
         self.generate_device = generate_device
         self.prefill_device = prefill_device
@@ -122,7 +122,7 @@ class RotaryEmbeddingV2(BaseInjectedModule, LlamaRotaryEmbedding):
         **kwargs,
     ):
         BaseInjectedModule.__init__(
-            self, key, gguf_loader, config, orig_module, generate_device, **kwargs
+            self, key, gguf_loader, config, orig_module, prefill_device, generate_device, **kwargs
         )
         self.orig_module.__init__(
             orig_module.dim,
@@ -160,7 +160,7 @@ class YarnRotaryEmbedding(BaseInjectedModule, DeepseekV2YarnRotaryEmbedding):
         **kwargs,
     ):
         BaseInjectedModule.__init__(
-            self, key, gguf_loader, config, orig_module, generate_device, **kwargs
+            self, key, gguf_loader, config, orig_module, prefill_device, generate_device, **kwargs
         )
         self.orig_module.__init__(
             orig_module.dim,
@@ -204,7 +204,7 @@ class YarnRotaryEmbedding(BaseInjectedModule, DeepseekV2YarnRotaryEmbedding):
 #         **kwargs,
 #     ):
 #         BaseInjectedModule.__init__(
-#             self, key, gguf_loader, config, orig_module, generate_device, **kwargs
+#             self, key, gguf_loader, config, orig_module, prefill_device, generate_device, **kwargs
 #         )
 #         self.generate_device = generate_device
 #         self.prefill_device = prefill_device
@@ -230,7 +230,7 @@ class YarnRotaryEmbeddingV3(BaseInjectedModule):
         **kwargs,
     ):
         BaseInjectedModule.__init__(
-            self, key, gguf_loader, config, orig_module, generate_device, **kwargs
+            self, key, gguf_loader, config, orig_module, prefill_device, generate_device, **kwargs
         )
         self.generate_device = generate_device
         self.prefill_device = prefill_device
@@ -332,11 +332,12 @@ class DynamicNTKScalingRotaryEmbedding(
         gguf_loader: GGUFLoader,
         config: PretrainedConfig,
         orig_module: nn.Module,
-        device: str = "cuda",
+        prefill_device: str = "cuda",
+        generate_device: str = "cuda",
         **kwargs,
     ):
         BaseInjectedModule.__init__(
-            self, key, gguf_loader, config, orig_module, device, **kwargs
+            self, key, gguf_loader, config, orig_module, prefill_device, generate_device, **kwargs
         )
         self.orig_module.__init__(
             orig_module.dim,

ktransformers/operators/attention.py

@@ -19,9 +19,13 @@ from ktransformers.util.custom_gguf import GGUFLoader
 import logging
 from transformers.configuration_utils import PretrainedConfig
 from transformers.cache_utils import Cache
-from flash_attn import flash_attn_with_kvcache, flash_attn_func
+from flash_attn import flash_attn_func
 from ktransformers.operators.triton_attention import decode_attention_fwd_grouped
 import os
+from ktransformers.operators.flashinfer_wrapper import flashinfer_enabled
+if flashinfer_enabled:
+    from ktransformers.operators.flashinfer_wrapper import MLAWrapperSingleton, attention_ref
+
 logger = logging.getLogger("attention")
 
 # Copied from transformers.models.llama.modeling_llama.rotate_half
@@ -41,15 +45,15 @@ class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
                  gguf_loader : GGUFLoader,
                  config: PretrainedConfig,
                  orig_module: nn.Module,
-                 device: str = "cuda",
+                 prefill_device: str = "cuda",
+                 generate_device: str = "cuda",
                  chunck_size: int = 1000,
-                 use_triton: bool = False,
                  **kwargs):
-        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, device, **kwargs)
+        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, prefill_device, generate_device, **kwargs)
         self.orig_module.__init__(orig_module.config,
             orig_module.layer_idx)
         self.chunck_size = chunck_size # TODO, generate chunck_size automatically.
-        self.use_triton = use_triton
+        self.mla_wrapper = None
 
     def get_absorbed(self) -> Tuple[torch.Tensor, torch.Tensor]:
         if not (hasattr(self, 'q_absorb') and hasattr(self, 'out_absorb')):
@@ -141,6 +145,7 @@ class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
         #print(compressed_kv.shape)
         
         attn_weights = (torch.matmul(q_pe, k_pe.mT) + torch.matmul(q_nope, compressed_kv.mT)) * self.softmax_scale
+        
         #attn_weights [bsz, self.num_heads, q_len, kv_seq_len]
         compressed_kv = compressed_kv.squeeze(1)
         """
@@ -168,6 +173,7 @@ class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
         attn_weights = nn.functional.dropout(
             attn_weights, p=self.attention_dropout, training=self.training
         )
+        
         attn_output = torch.einsum('bhql,blc->bhqc', attn_weights, compressed_kv)
         
         attn_output = torch.matmul(attn_output, out_absorb.mT) 
@@ -186,7 +192,7 @@ class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
 
         return attn_output, None, past_key_value
 
-    def forward_linux(
+    def forward_linux_triton(
             self,
             hidden_states: torch.Tensor,
             attention_mask: Optional[torch.Tensor] = None,
@@ -232,7 +238,7 @@ class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
         # q_pe [bsz, q_len, self.num_heads, self.qk_rope_head_dim] k_pe [bsz, q_len, 1, self.qk_rope_head_dim]
         
         # decode
-        if self.use_triton and q_len == 1:
+        if q_len == 1:
             if past_key_value is not None:
                 cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
                 compressed_kv_with_k_pe, page_table = past_key_value.update(compressed_kv, k_pe, self.layer_idx, cache_kwargs)
@@ -277,7 +283,7 @@ class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
             # use triton attention kernel adapted from vLLM and SGLang for MQA
             decode_attention_fwd_grouped(query_states, compressed_kv_with_k_pe, compressed_kv, attn_output,
                              page_table,
-                             position_ids.squeeze(0).to(torch.int32), attn_logits,
+                             position_ids.squeeze(0).to(torch.int32)+1, attn_logits,
                              4, #num_kv_splits # follow vLLM, fix it TODO
                              self.softmax_scale,
                              past_key_value.page_size)
@@ -338,6 +344,154 @@ class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
             attn_output = self.o_proj(attn_output)
             return attn_output, None, past_key_value
 
+    def forward_linux_flashinfer(
+            self,
+            hidden_states: torch.Tensor,
+            attention_mask: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.Tensor] = None,
+            past_key_value: Optional[Cache] = None,
+            output_attentions: bool = False,
+            use_cache: bool = False,
+            cache_position: Optional[torch.Tensor] = None,
+            **kwargs,
+        ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+
+        bsz, q_len, _ = hidden_states.size()
+
+        if self.q_lora_rank is None:
+            q = self.q_proj(hidden_states)
+        else:
+            q = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
+        q = q.view(bsz, q_len, self.num_heads, self.q_head_dim)
+        q_nope, q_pe = torch.split(
+            q, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1
+        )
+
+        compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
+        compressed_kv, k_pe = torch.split(
+            compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
+        )
+        compressed_kv = self.kv_a_layernorm(compressed_kv)
+        k_pe = k_pe.view(bsz, q_len, 1, self.qk_rope_head_dim)
+        compressed_kv = compressed_kv.view(bsz, q_len, 1, self.kv_lora_rank)
+        
+        cos, sin = self.rotary_emb(q_pe, position_ids)
+        q_pe, k_pe = apply_rotary_pos_emb(q_pe, k_pe, cos, sin, unsqueeze_dim=2)
+        # q_pe [bsz, q_len, self.num_heads, self.qk_rope_head_dim] k_pe [bsz, q_len, 1, self.qk_rope_head_dim]
+        
+        # decode
+        if q_len == 1:
+            if past_key_value is not None:
+                cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
+                compressed_kv_with_k_pe, page_table = past_key_value.update(compressed_kv, k_pe, self.layer_idx, cache_kwargs)
+                compressed_kv = compressed_kv_with_k_pe [:, :, :, :self.kv_lora_rank].view(-1, past_key_value.page_size, self.kv_lora_rank)
+                k_pe = compressed_kv_with_k_pe [:, :, :, self.kv_lora_rank:].view(-1, past_key_value.page_size, self.qk_rope_head_dim)
+                # k_pe [max_pages, page_size, self.qk_rope_head_dim]
+                # compressed_kv [max_pages, page_size, self.kv_lora_rank]
+
+            # q_nope [bsz, q_len, self.num_heads, self.qk_nope_head_dim]
+            # q_absorb [self.num_heads, self.qk_nope_head_dim, self.kv_lora_rank]
+            q_absorb, out_absorb = self.get_absorbed()
+            q_nope = q_nope.transpose(1, 2) # q_len is 1, no GPU overhead, same below
+            q_nope = torch.matmul(q_nope, q_absorb) # batched MM
+            q_nope = q_nope.transpose(1, 2)
+            assert q_nope.is_contiguous()
+            
+            # q_nope [bsz, q_len, self.num_heads, self.kv_lora_rank]
+            # q_pe [bsz, q_len, self.num_heads, self.qk_rope_head_dim]
+            q_nope.squeeze_(1)
+            q_pe.squeeze_(1)
+
+            # flash attn doesn't support head_dim bigger than 256, use flashinfer
+            if self.mla_wrapper is None:
+                self.mla_wrapper = MLAWrapperSingleton.get_instance(self.device, 1, past_key_value.max_pages, use_cuda_graph = True)
+                if self.mla_wrapper.need_plan:
+                    self.mla_wrapper.need_plan = False
+                    self.mla_wrapper.plan(None,None,None,
+                                      position_ids.squeeze(1)+1,
+                                      self.num_heads,
+                                      self.kv_lora_rank,
+                                      self.qk_rope_head_dim,
+                                      past_key_value.page_size,
+                                      self.softmax_scale,
+                                      q_nope.dtype,
+                                      compressed_kv.dtype)
+            attn_output = self.mla_wrapper.run(q_nope, q_pe, compressed_kv, k_pe).view(bsz, q_len, self.num_heads, self.kv_lora_rank)
+
+            """
+            k = (
+                torch.cat([compressed_kv, k_pe], dim=-1)
+                .view(-1, 1, 512 + 64)
+                .repeat_interleave(self.num_heads, dim=1)
+            )
+            v = compressed_kv.view(-1, 1, 512).repeat_interleave(self.num_heads, dim=1)
+            lens = position_ids.item() + 1
+            #print("lens", lens)
+            attn_ref, lse_ref = attention_ref(
+                1,
+                torch.cat([q_nope, q_pe], dim=-1),
+                k[:lens],
+                v[:lens],
+                False,
+                self.softmax_scale
+            )
+            attn_output = attn_ref.view(bsz, q_len, self.num_heads, self.kv_lora_rank)
+            """
+            
+            # mla_wrapper run output: [tokens, self.num_heads, self.kv_lora_rank]
+            # attn_output [bsz, q_len, self.num_heads, self.kv_lora_rank]
+            # out_absorb [self.num_heads, self.v_head_dim, self.kv_lora_rank]
+            attn_output = attn_output.transpose(1, 2)
+            attn_output = torch.matmul(attn_output, out_absorb.mT)
+            
+            attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.v_head_dim)
+            attn_output = self.o_proj(attn_output)
+
+            return attn_output, None, past_key_value
+        else:
+            if past_key_value is not None:
+                cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
+                k_pe.squeeze(0)
+                compressed_kv.squeeze(0)
+                past_key_value.update(compressed_kv, k_pe, self.layer_idx, cache_kwargs)
+                k_pe.unsqueeze(0)
+                compressed_kv.unsqueeze(0)
+        
+            k_pe = k_pe[:, :q_len]
+            compressed_kv = compressed_kv[:, :q_len]
+            kv = (
+                self.kv_b_proj(compressed_kv)
+                .view(bsz, q_len, self.num_heads, self.qk_nope_head_dim + self.v_head_dim)
+            )
+            k_nope, value_states = torch.split(kv, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+            query_states = k_pe.new_empty(bsz, q_len, self.num_heads, self.q_head_dim)
+            query_states[:, :, :, : self.qk_nope_head_dim] = q_nope
+            query_states[:, :, :, self.qk_nope_head_dim :] = q_pe
+
+            key_states = k_pe.new_empty(bsz, q_len, self.num_heads, self.q_head_dim)
+            key_states[:, :, :, :self.qk_nope_head_dim] = k_nope
+            key_states[:, :, :, self.qk_nope_head_dim:] = k_pe
+            
+            value_states = value_states.view(bsz, q_len, self.num_heads, self.v_head_dim)
+            value_states_padded = torch.nn.functional.pad(value_states, [0, query_states.shape[-1] - value_states.shape[-1]], value=0)
+
+            attn_output = flash_attn_func(
+                query_states,
+                key_states,
+                value_states_padded,
+                softmax_scale=self.softmax_scale,
+                causal=True,
+            )
+
+            if self.q_head_dim != self.v_head_dim:
+                attn_output = attn_output[:, :, :, : self.v_head_dim]
+
+            attn_output = attn_output.reshape(
+                bsz, q_len, self.num_heads * self.v_head_dim
+            ).contiguous()
+            attn_output = self.o_proj(attn_output)
+            return attn_output, None, past_key_value
+        
     def forward_windows(
         self,
         hidden_states: torch.Tensor,
@@ -415,7 +569,7 @@ class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
         cache_position: Optional[torch.LongTensor] = None,
         **kwargs,
     ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-        if os.name == 'nt' or hidden_states.shape[1] == 1: # Use in decode
+        if os.name == 'nt':
             return self.forward_windows(
                 hidden_states,
                 attention_mask,
@@ -427,16 +581,28 @@ class KDeepseekV2Attention(BaseInjectedModule, DeepseekV2Attention):
                 **kwargs,
             )
         else:
-            return self.forward_linux(
+            if flashinfer_enabled:
-                hidden_states,
+                return self.forward_linux_flashinfer(
-                attention_mask,
+                    hidden_states,
-                position_ids,
+                    attention_mask,
-                past_key_value,
+                    position_ids,
-                output_attentions,
+                    past_key_value,
-                use_cache,
+                    output_attentions,
-                cache_position,
+                    use_cache,
-                **kwargs,
+                    cache_position,
-            )
+                    **kwargs,
+                )
+            else:
+                return self.forward_linux_triton(
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    past_key_value,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                    **kwargs,
+                )
 
 
 class KLlamaAttention(BaseInjectedModule):
@@ -447,9 +613,10 @@ class KLlamaAttention(BaseInjectedModule):
                  gguf_loader : GGUFLoader,
                  config: PretrainedConfig,
                  orig_module: nn.Module,
-                 device: str = "cuda",
+                 prefill_device: str = "cuda",
+                 generate_device: str = "cuda",
                  **kwargs):
-        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, device, **kwargs)
+        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, prefill_device, generate_device, **kwargs)
         self.orig_module.__init__(orig_module.config,
             orig_module.layer_idx)
     def apply_rotary_pos_emb(self, q, k, cos, sin, position_ids=None, unsqueeze_dim=1):

ktransformers/operators/base_operator.py

@@ -16,14 +16,17 @@ class BaseInjectedModule(nn.Module):
                  gguf_loader : GGUFLoader,
                  config: PretrainedConfig,
                  orig_module: nn.Module,
-                 device: str = "cuda",
+                 prefill_device: str = "cuda",
+                 generate_device: str = "cuda",
                  **kwargs):
         nn.Module.__init__(self)
         nn.Module.__setattr__(self, "orig_module", orig_module)
         object.__setattr__(self, "key", key)
         object.__setattr__(self, "gguf_loader", gguf_loader)
         object.__setattr__(self, "config", config)
-        object.__setattr__(self, "device", device)
+        object.__setattr__(self, "prefill_device", prefill_device)
+        object.__setattr__(self, "generate_device", generate_device)
+        object.__setattr__(self, "device", generate_device)
         
     def __getattr__(self, name: str) -> Any:
         # __getattr__ in nn.Module doesn't call super().__getattribute__ when name is not in nn.Module.__dict__,

ktransformers/operators/experts.py

@@ -119,6 +119,7 @@ class KExpertsCPU(KExpertsBase):
     output_cpu:Tensor = None
     output_gpu_map:dict = {} # Manage output tensor buffer on different gpu
     #stream_map:dict = {} # Manage cuda stream on different gpu
+    #gguf_loader:GGUFLoader = None
     CPU_INFER = CPUInfer(Config().cpu_infer)
     def __init__(
         self,
@@ -132,6 +133,9 @@ class KExpertsCPU(KExpertsBase):
         **kwargs
     ):
         super().__init__(key, gguf_loader, config, orig_module, device, **kwargs)
+        #if KExpertsCPU.gguf_loader is None:
+        #    KExpertsCPU.gguf_loader = GGUFLoader("/mnt/data/model/DeepseekV3-q4km-gguf")
+        self.gguf_loader = gguf_loader
         assert device.lower() == "cpu", "KExpertsCPU can only be loaded on CPU"
         self.n_routed_experts = n_routed_experts
         self.out_device = out_device
@@ -532,7 +536,7 @@ class KTransformersExperts(BaseInjectedModule, KExpertsBase):
                  generate_device: str = "cpu",
                  generate_op: str | None = "KExpertsCPU",
                  **kwargs):
-        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, generate_device, **kwargs)
+        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, prefill_device, generate_device, **kwargs)
         KExpertsBase.__init__(self, key, gguf_loader, config, orig_module, generate_device, **kwargs)
         if generate_op is not None:
             self.generate_experts = EXPERTS_MAP[generate_op](key, gguf_loader, config, len(orig_module), device=generate_device, **kwargs)

ktransformers/operators/flashinfer_wrapper.py

@@ -0,0 +1,240 @@
+'''
+Description  : flashinfer MLA wrapper
+Author       : Boxin Zhang
+Version      : 0.2.2
+'''
+import torch
+
+flashinfer_enabled = False
+
+try:
+    import flashinfer
+    flashinfer_enabled = True
+    print("found flashinfer")
+    
+except ImportError:
+    print("flashinfer not found, use triton for linux")
+
+import math
+
+def attention_ref(
+    batch_size,
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    causal: bool,
+    sm_scale: float,
+) -> torch.Tensor:
+    qo_len = q.shape[0] // batch_size
+    kv_len = k.shape[0] // batch_size
+    num_qo_heads = q.shape[1]
+    head_dim_qk = q.shape[2]
+    head_dim_vo = v.shape[2]
+    logits = (
+        torch.einsum(
+            "bmhd,bnhd->bhmn",
+            q.view(batch_size, qo_len, num_qo_heads, head_dim_qk).float(),
+            k.view(batch_size, kv_len, num_qo_heads, head_dim_qk).float(),
+        )
+        * sm_scale
+    )
+
+    #print("attn weights", logits)
+
+    if causal:
+        mask = (
+            torch.arange(kv_len - qo_len, kv_len).unsqueeze(1)
+            >= torch.arange(0, kv_len).unsqueeze(0)
+        ).to(q.device)
+    else:
+        mask = torch.ones(qo_len, kv_len).to(q.device)
+
+    logits = logits.masked_fill(mask.unsqueeze(0).unsqueeze(0) == 0, float("-inf"))
+    lse_ref = torch.logsumexp(logits, -1).transpose(-1, -2)
+    p = torch.softmax(logits, dim=-1)
+    o_ref = (
+        torch.einsum(
+            "bhmn,bnhd->bmhd",
+            p,
+            v.view(batch_size, kv_len, num_qo_heads, head_dim_vo).float(),
+        )
+        .contiguous()
+        .view(batch_size * qo_len, num_qo_heads, head_dim_vo)
+        .to(q)
+    )
+
+    return o_ref, lse_ref * math.log2(math.e)
+
+class MLAWrapper():
+    def __init__(self,
+                 max_batch_size,
+                 max_pages,
+                 use_cuda_graph = True,
+                 device = "cuda",
+                 ):
+        self.float_workspace_buffer = torch.empty(128*1024*1024, dtype=torch.int8, device=device)
+        self.max_batch_size = max_batch_size
+        self.max_pages = max_pages
+        if use_cuda_graph:
+            if self.max_batch_size == 1:
+                self.qo_indptr_buf = torch.arange(0, max_batch_size+1, dtype=torch.int32, device=device)
+                self.kv_indptr_buf = torch.tensor([0, max_pages], dtype=torch.int32, device=device)
+                self.kv_indices_buf = torch.arange(0, max_pages, dtype=torch.int32, device=device)
+            else:
+                self.qo_indptr_buf = torch.empty(max_batch_size+1, dtype=torch.int32, device=device)
+                self.kv_indptr_buf = torch.empty(max_batch_size+1, dtype=torch.int32, device=device)
+                self.kv_indices_buf = torch.empty(max_pages, dtype=torch.int32, device=device)
+            self.kv_len_arr_buf = torch.empty(max_batch_size, dtype=torch.int32, device=device)
+        else:
+            self.qo_indptr_buf = None
+            self.kv_indptr_buf = None
+            self.kv_indices_buf = None
+            self.kv_len_arr_buf = None
+        self.wrapper = flashinfer.mla.BatchMLAPagedAttentionWrapper(
+            self.float_workspace_buffer,
+            use_cuda_graph=False,
+            qo_indptr=self.qo_indptr_buf,
+            kv_indptr=self.kv_indptr_buf,
+            kv_indices=self.kv_indices_buf,
+            kv_len_arr=self.kv_len_arr_buf,
+        )
+        self.need_plan = True
+    
+    def plan(self,
+             qo_indptr,
+             kv_indptr,
+             kv_indices,
+             kv_len_arr,
+             num_heads,
+             head_dim_ckv,
+             head_dim_kpe,
+             page_size,
+             sm_scale,
+             q_data_type,
+             kv_data_type,
+             ):
+        if qo_indptr is None:
+            assert self.max_batch_size == 1
+            qo_indptr = self.qo_indptr_buf
+        if kv_indptr is None:
+            assert self.max_batch_size == 1
+            kv_indptr = self.kv_indptr_buf
+        if kv_indices is None:
+            assert self.max_batch_size == 1
+            kv_indices = self.kv_indices_buf
+
+        self.wrapper.plan(
+            qo_indptr,
+            kv_indptr,
+            kv_indices,
+            kv_len_arr,
+            num_heads,
+            head_dim_ckv,
+            head_dim_kpe,
+            page_size,
+            False, # causal is False for decoding
+            sm_scale,
+            q_data_type,
+            kv_data_type,
+        )
+
+    def run(self, q_nope, q_pe, ckv, k_pe, return_lse = False):
+        return self.wrapper.run(q_nope, q_pe, ckv, k_pe, return_lse)
+
+class MLAWrapperSingleton():
+    wrappers:dict = {}
+
+    @classmethod
+    def get_instance(cls, device, *args, **kwargs)->MLAWrapper:
+        if device not in cls.wrappers:
+            cls.make_instance(device, *args, **kwargs)
+        return cls.wrappers[device]
+    
+    @classmethod
+    def make_instance(cls, device, *args, **kwargs):
+        cls.wrappers[device] = MLAWrapper(*args, **kwargs, device=device)
+
+    @classmethod
+    def plan_all(cls, qo_indptr,
+             kv_indptr,
+             kv_indices,
+             kv_len_arr,
+             num_heads,
+             head_dim_ckv,
+             head_dim_kpe,
+             page_size,
+             sm_scale,
+             q_data_type,
+             kv_data_type,):
+        for device, wrapper in cls.wrappers.items():
+            kv_len_arr_cur_device = kv_len_arr.to(device)
+            wrapper.plan(qo_indptr,
+                kv_indptr,
+                kv_indices,
+                kv_len_arr_cur_device,
+                num_heads,
+                head_dim_ckv,
+                head_dim_kpe,
+                page_size,
+                sm_scale,
+                q_data_type,
+                kv_data_type,)
+            
+
+if __name__ == "__main__":
+    max_batch_size = 1
+    max_pages = 1
+    page_size = 64
+    num_heads = 128
+
+    q_nope = torch.randn((1, num_heads, 512), dtype=torch.bfloat16, device="cuda")
+    q_pe = torch.randn((1, num_heads, 64), dtype=torch.bfloat16, device="cuda")
+    ckv = torch.randn((max_pages, page_size, 512), dtype=torch.bfloat16, device="cuda")
+    k_pe = torch.randn((max_pages, page_size, 64), dtype=torch.bfloat16, device="cuda")
+    
+
+    wrapper = MLAWrapperSingleton.get_instance(
+        "cuda",
+        max_batch_size,
+        max_pages,
+    )
+    
+    kv_len_arr = torch.tensor([10], dtype=torch.int32, device="cuda")
+
+    wrapper.plan(
+        None,
+        None,
+        None,
+        kv_len_arr,
+        128,
+        512,
+        64,
+        page_size,
+        192 ** (-0.5),
+        torch.bfloat16,
+        torch.bfloat16,
+    )
+
+    attn_output = wrapper.run(q_nope, q_pe, ckv, k_pe)
+
+    k = (
+        torch.cat([ckv, k_pe], dim=-1)
+        .view(-1, 1, 512 + 64)
+        .repeat_interleave(num_heads, dim=1)
+    )
+    v = ckv.view(-1, 1, 512).repeat_interleave(num_heads, dim=1)
+
+    print(k[:10].shape)
+    print(v[:10].shape)
+
+    attn_ref, lse_ref = attention_ref(
+        max_batch_size,
+        torch.cat([q_nope, q_pe], dim=-1),
+        k[:10],
+        v[:10],
+        False,
+        192 ** (-0.5)
+    )
+    
+    torch.testing.assert_close(attn_output, attn_ref, rtol=1e-3, atol=1e-3)
+    print("test past")

ktransformers/operators/gate.py

@@ -93,11 +93,11 @@ class KMoEGate(BaseInjectedModule, KMoEGateBase):
         gguf_loader: GGUFLoader,
         config: PretrainedConfig,
         orig_module: nn.Module = None,
-        generate_device: str = "cuda",
         prefill_device: str = "cuda",
+        generate_device: str = "cuda",
         **kwargs,
     ):
-        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, generate_device, **kwargs)
+        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, prefill_device, generate_device, **kwargs)
         KMoEGateBase.__init__(self, key, gguf_loader, config, orig_module, generate_device, **kwargs)
         self.generate_device = generate_device
         self.prefill_device = prefill_device

ktransformers/operators/linear.py

@@ -383,7 +383,7 @@ class KTransformersLinear(BaseInjectedModule, KLinearBase):
         prefill_op: str| None = "KLinearTorch",
         **kwargs,
     ):
-        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, generate_device, **kwargs)
+        BaseInjectedModule.__init__(self, key, gguf_loader, config, orig_module, prefill_device, generate_device, **kwargs)
         KLinearBase.__init__(self, key, gguf_loader, config, orig_module, generate_device, **kwargs)
         # build all the linear operators
         if prefill_op is not None:

ktransformers/util/custom_gguf.py

@@ -109,6 +109,7 @@ GGML_TYPES = {
     "Q5_K": 13,
     "Q6_K": 14,
     "IQ4_XS": 23,
+    "BF16": 30,
 }
 
 GGML_NAMES = {ggml_type: name for name, ggml_type in GGML_TYPES.items()}
@@ -116,6 +117,7 @@ GGML_NAMES = {ggml_type: name for name, ggml_type in GGML_TYPES.items()}
 GGML_BLOCK_SIZES = {
     "F32": 4,
     "F16": 2,
+    "BF16": 2,
     "Q4_0": 2 + 16,
     "Q5_0": 2 + 4 + 16,
     "Q8_0": 2 + 32,
@@ -130,6 +132,7 @@ GGML_BLOCK_SIZES = {
 GGML_ELEMENTS_PER_BLOCK = {
     "F32": 1,
     "F16": 1,
+    "BF16": 1,
     "Q4_0": 32,
     "Q5_0": 32,
     "Q8_0": 32,
@@ -333,6 +336,8 @@ class GGUFLoader:
         else:
             values = GGML_DEQUANTIZE[ggml_name](data)
             values = torch.from_numpy(values)
+        if ggml_name == "BF16":
+            values = values.view(torch.bfloat16)
         values = values.view(shape[::-1])
         if "attn_q" in name and self.gguf_file_meta['general.architecture'] in ["llama"]:
             n_head = self.gguf_file_meta['llama.attention.head_count']
@@ -764,6 +769,7 @@ def dequantize_f16_gpu(data, device):
 GGML_DEQUANTIZE = {
     "F32": dequantize_f32,
     "F16": dequantize_f16,
+    "BF16": dequantize_f16,
     "Q4_0": dequantize_q4_0,
     "Q5_0": dequantize_q5_0,
     "Q8_0": dequantize_q8_0,
@@ -778,6 +784,7 @@ GGML_DEQUANTIZE = {
 GGML_DEQUANTIZE_GPU = {
     "F32": dequantize_f32_gpu,
     "F16": dequantize_f16_gpu,
+    "BF16": dequantize_f16_gpu,
     "Q4_0": dequantize_q4_0_gpu,
     "Q5_0": dequantize_q5_0_gpu,
     "Q8_0": dequantize_q8_0_gpu,

ktransformers/util/utils.py

@@ -17,6 +17,7 @@ from ktransformers.operators import base_operator
 from ktransformers.models.custom_cache import StaticCache
 from ktransformers.util.cuda_graph_runner import CUDAGraphRunner
 from ktransformers.util.textstream import TextStreamer
+from ktransformers.operators.flashinfer_wrapper import MLAWrapperSingleton
 
 warm_uped = False
 
@@ -87,7 +88,8 @@ def load_weights(module:nn.Module, gguf_loader:GGUFLoader, prefix=''):
         module.load()
 
 def prefill_and_generate(model, tokenizer, inputs, max_new_tokens=10000, use_cuda_graph: bool = True,
-                         mode = 'normal', force_think: bool = False):
+                         mode = 'normal', force_think: bool = False, use_flashinfer_mla = False,
+                         num_heads = None, head_dim_ckv = None, head_dim_kpe = None, q_head_dim = None):
     import os
     os.environ["TOKENIZERS_PARALLELISM"] = "false"
     torch._dynamo.config.suppress_errors = True
@@ -137,7 +139,7 @@ def prefill_and_generate(model, tokenizer, inputs, max_new_tokens=10000, use_cud
             )
         else:
             past_key_values = None
-        cache_position = torch.arange(seq_length, device=torch_device, dtype=torch.long)
+        cache_position = torch.arange(seq_length, device=torch_device, dtype=torch.int32)
         generated_ids = torch.zeros(
             batch_size, seq_length + max_new_tokens + 1, dtype=torch.int, device=torch_device
         )
@@ -182,7 +184,7 @@ def prefill_and_generate(model, tokenizer, inputs, max_new_tokens=10000, use_cud
         generated_ids[:, seq_length] = next_token
         tokens.append(int(next_token))
         inputs = torch.cat((inputs, next_token.unsqueeze(0)), dim=-1)
-        cache_position = torch.tensor([seq_length], device=torch_device, dtype=torch.long)
+        cache_position = torch.tensor([seq_length], device=torch_device, dtype=torch.int32)
         position_ids = cache_position.unsqueeze(0)
         seq_length += 1
         
@@ -195,7 +197,10 @@ def prefill_and_generate(model, tokenizer, inputs, max_new_tokens=10000, use_cud
                 warm_uped = True
                 cuda_graph_runner = CUDAGraphRunner()
                 cuda_graph_runner.capture(model, next_token.unsqueeze(0), position_ids, cache_position, past_key_values, torch_device, return_dict=False, use_cache=True)
-                
+            if i > 1 and use_flashinfer_mla:
+                MLAWrapperSingleton.plan_all(None,None,None,position_ids.squeeze(1)+1,
+                                             num_heads, head_dim_ckv, head_dim_kpe, past_key_values.page_size,
+                                             q_head_dim ** (-0.5), torch.bfloat16, torch.bfloat16)
             next_token = decode_one_tokens(cuda_graph_runner, next_token.unsqueeze(0), position_ids, cache_position, past_key_values, use_cuda_graph).to(torch_device)
             inputs = torch.cat((inputs, next_token.unsqueeze(0)), dim=-1)
             generated_ids[:, cache_position] = next_token.int()