Merge pull request #62 from Azure-Tang/main

[Fix] Fix problem that ktransformers cannot offload whole layer in cpu

Dockerfile

@@ -25,7 +25,7 @@ rm -rf /var/lib/apt/lists/* &&
 cd ktransformers &&
 git submodule init &&
 git submodule update &&
-pip install ninja pyproject numpy &&
+pip install ninja pyproject numpy cpufeature &&
 pip install flash-attn &&
 CPU_INSTRUCT=NATIVE  KTRANSFORMERS_FORCE_BUILD=TRUE TORCH_CUDA_ARCH_LIST="8.0;8.6;8.7;8.9" pip install . --no-build-isolation --verbose &&
 pip cache purge

ktransformers/__init__.py

@@ -5,7 +5,7 @@ Description  :
 Author       : kkk1nak0
 Date         : 2024-08-15 07:34:46
 Version      : 1.0.0
-LastEditors  : chenxl 
-LastEditTime : 2024-08-28 15:19:03
+LastEditors  : Azure-Tang 
+LastEditTime : 2024-08-29 22:35:51
 '''
-__version__ = "0.1.3"
+__version__ = "0.1.4"
\ No newline at end of file

ktransformers/local_chat.py

@@ -67,6 +67,7 @@ def local_chat(
     tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
     config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
     if mode == 'long_context':
+        assert config.architectures[0] == "LlamaForCausalLM", "only LlamaForCausalLM support long_context mode"
         torch.set_default_dtype(torch.float16)
     else:
         torch.set_default_dtype(config.torch_dtype)
@@ -143,6 +144,7 @@ def local_chat(
         input_tensor = tokenizer.apply_chat_template(
             messages, add_generation_prompt=True, return_tensors="pt"
         )
+        if mode == 'long_context':
             assert Config().long_context_config['max_seq_len'] > input_tensor.shape[1] + max_new_tokens, \
             "please change max_seq_len in  ~/.ktransformers/config.yaml"
         torch.set_default_dtype(

ktransformers/operators/experts.py

@@ -6,7 +6,7 @@ Author       : Azure-Tang, Boxin Zhang, chenht2022
 Date         : 2024-07-25 11:25:24
 Version      : 0.1.0
 LastEditors  : Azure 
-LastEditTime : 2024-08-27 03:50:23
+LastEditTime : 2024-08-29 09:41:10
 Copyright (c) 2024 by KVCache.AI, All Rights Reserved. 
 '''
 
@@ -202,7 +202,7 @@ class KExpertsCPU(KExpertsBase):
     def forward(self, input_tensor, expert_ids, weights):
         # generate, capture and run cuda graph
         # print(expert_ids)
-        if input_tensor.size(0)==1:
+        if input_tensor.size(0)==1 and torch.cuda.is_current_stream_capturing():
             # TODO: this branch is unreachable, but the shape of input_tensor([1,hidden_size]) and input_tensor_cpu([hidden_size]) is not compatible
             #print("capturing experts")
             KExpertsCPU.input_tensor_cpu.copy_(input_tensor, non_blocking=True)
@@ -636,7 +636,7 @@ class KDeepseekV2MoE(BaseInjectedModule, DeepseekV2MoE):
         topk_idx, topk_weight, aux_loss = self.gate(hidden_states)
         hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
         
-        if sequence_length == 1 and hasattr(self.experts.generate_experts, "submit_for_one_decode"):
+        if sequence_length == 1 and hasattr(self.experts.generate_experts, "submit_for_one_decode") and torch.cuda.is_current_stream_capturing():
             self.experts.generate_experts.submit_for_one_decode(hidden_states[0], topk_idx[0], topk_weight[0])
             if self.config.n_shared_experts is not None:
                 y_ = self.shared_experts(identity).squeeze(0)

ktransformers/operators/linear.py

@@ -6,7 +6,7 @@ Author       : Azure-Tang, Boxin Zhang
 Date         : 2024-07-25 11:25:24
 Version      : 0.1.0
 LastEditors  : Azure 
-LastEditTime : 2024-08-14 14:57:04
+LastEditTime : 2024-08-29 09:11:16
 Copyright (c) 2024 by KVCache.AI, All Rights Reserved. 
 '''
 
@@ -277,7 +277,7 @@ class KLinearCPUInfer(KLinearBase):
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         origin_shape = x.shape # [batch_size, q_len, hidden_size]
-        if origin_shape[1] == 1:
+        if origin_shape[1] == 1 and torch.cuda.is_current_stream_capturing():
             out_device = x.device
             self.input_tensor_cpu.copy_(x, non_blocking=True)
             qlen = origin_shape[1]

ktransformers/operators/models.py

@@ -670,8 +670,9 @@ class KDeepseekV2Model(BaseInjectedModule):
             if self.transfer_map is not None and i in self.transfer_map:
                 prev_stream = torch.cuda.current_stream()
                 cur_device = self.transfer_map[i]
-                if cur_device not in self.stream_device_map:
+                if cur_device not in self.stream_device_map and cur_device.lower() != "cpu":
                     self.stream_device_map[cur_device] = torch.cuda.Stream(cur_device)
+                if cur_device.lower() != "cpu":
                     torch.cuda.set_device(cur_device)
                     self.stream_device_map[cur_device].wait_stream(prev_stream)
                     torch.cuda.set_stream(self.stream_device_map[cur_device])

ktransformers/optimize/optimize_rules/DeepSeek-V2-Chat-multi-gpu-4.yaml

@@ -7,7 +7,7 @@
         prefill_device: "cpu"
 
 - match:
-    name: "^model\\.layers\\.([0-9])\\."
+    name: "^model\\.layers\\.([0-9]|[1][0-4])\\."
     class: ktransformers.models.modeling_deepseek.DeepseekV2YarnRotaryEmbedding
   replace:
     class: ktransformers.operators.RoPE.YarnRotaryEmbedding
@@ -15,7 +15,7 @@
       generate_device: "cuda:0"
       prefill_device: "cuda:0"
 - match:
-    name: "^model\\.layers\\.([1][0-9])\\."
+    name: "^model\\.layers\\.([2][0-9]|[1][5-9])\\."
     class: ktransformers.models.modeling_deepseek.DeepseekV2YarnRotaryEmbedding
   replace:
     class: ktransformers.operators.RoPE.YarnRotaryEmbedding
@@ -23,7 +23,7 @@
       generate_device: "cuda:1"
       prefill_device: "cuda:1"
 - match:
-    name: "^model\\.layers\\.([2][0-9])\\."
+    name: "^model\\.layers\\.([3][0-9]|[4][0-4])\\."
     class: ktransformers.models.modeling_deepseek.DeepseekV2YarnRotaryEmbedding
   replace:
     class: ktransformers.operators.RoPE.YarnRotaryEmbedding
@@ -31,7 +31,7 @@
       generate_device: "cuda:2"
       prefill_device: "cuda:2"
 - match:
-    name: "^model\\.layers\\.([345][0-9])\\."
+    name: "^model\\.layers\\.([5][0-9]|[4][5-9])\\."
     class: ktransformers.models.modeling_deepseek.DeepseekV2YarnRotaryEmbedding
   replace:
     class: ktransformers.operators.RoPE.YarnRotaryEmbedding
@@ -40,7 +40,7 @@
       prefill_device: "cuda:3"
 
 - match:
-    name: "^model\\.layers\\.([0-9])\\.(?!self_attn).*$"  # regular expression 
+    name: "^model\\.layers\\.([0-9]|[1][0-4])\\.(?!self_attn\\.kv_b_proj).*$"  # regular expression 
     class: torch.nn.Linear  # only match modules matching name and class simultaneously
   replace:
     class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
@@ -50,7 +50,7 @@
       generate_op: "KLinearMarlin"
       prefill_op: "KLinearTorch"
 - match:
-    name: "^model\\.layers\\.([1][0-9])\\.(?!self_attn).*$"  # regular expression 
+    name: "^model\\.layers\\.([2][0-9]|[1][5-9])\\.(?!self_attn\\.kv_b_proj).*$"  # regular expression 
     class: torch.nn.Linear  # only match modules matching name and class simultaneously
   replace:
     class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
@@ -60,7 +60,7 @@
       generate_op: "KLinearMarlin"
       prefill_op: "KLinearTorch"
 - match:
-    name: "^model\\.layers\\.([2][0-9])\\.(?!self_attn).*$"  # regular expression 
+    name: "^model\\.layers\\.([3][0-9]|[4][0-4])\\.(?!self_attn\\.kv_b_proj).*$"  # regular expression 
     class: torch.nn.Linear  # only match modules matching name and class simultaneously
   replace:
     class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
@@ -70,7 +70,7 @@
       generate_op: "KLinearMarlin"
       prefill_op: "KLinearTorch"
 - match:
-    name: "^model\\.layers\\.([345][0-9])\\.(?!self_attn).*$"  # regular expression 
+    name: "^model\\.layers\\.([5][0-9]|[4][5-9])\\.(?!self_attn\\.kv_b_proj).*$"  # regular expression 
     class: torch.nn.Linear  # only match modules matching name and class simultaneously
   replace:
     class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
@@ -81,7 +81,7 @@
       prefill_op: "KLinearTorch"
 
 - match:
-    name: "^model\\.layers\\.([0-9])\\.mlp$"
+    name: "^model\\.layers\\.([0-9]|[1][0-4])\\.mlp$"
     class: ktransformers.models.modeling_deepseek.DeepseekV2MoE
   replace:
     class: ktransformers.operators.experts.KDeepseekV2MoE     # mlp module with custom forward function
@@ -89,7 +89,7 @@
       generate_device: "cuda:0"
       prefill_device: "cuda:0"
 - match:
-    name: "^model\\.layers\\.([1][0-9])\\.mlp$"
+    name: "^model\\.layers\\.([2][0-9]|[1][5-9])\\.mlp$"
     class: ktransformers.models.modeling_deepseek.DeepseekV2MoE
   replace:
     class: ktransformers.operators.experts.KDeepseekV2MoE     # mlp module with custom forward function
@@ -97,7 +97,7 @@
       generate_device: "cuda:1"
       prefill_device: "cuda:1"
 - match:
-    name: "^model\\.layers\\.([2][0-9])\\.mlp$"
+    name: "^model\\.layers\\.([3][0-9]|[4][0-4])\\.mlp$"
     class: ktransformers.models.modeling_deepseek.DeepseekV2MoE
   replace:
     class: ktransformers.operators.experts.KDeepseekV2MoE     # mlp module with custom forward function
@@ -105,7 +105,7 @@
       generate_device: "cuda:2"
       prefill_device: "cuda:2"
 - match:
-    name: "^model\\.layers\\.([345][0-9])\\.mlp$"
+    name: "^model\\.layers\\.([5][0-9]|[4][5-9])\\.mlp$"
     class: ktransformers.models.modeling_deepseek.DeepseekV2MoE
   replace:
     class: ktransformers.operators.experts.KDeepseekV2MoE     # mlp module with custom forward function
@@ -114,7 +114,7 @@
       prefill_device: "cuda:3"
 
 - match:
-    name: "^model\\.layers\\.([0-9])\\.mlp\\.experts$"
+    name: "^model\\.layers\\.([0-9]|[1][0-4])\\.mlp\\.experts$"
   replace:
     class: ktransformers.operators.experts.KTransformersExperts     # custom MoE Kernel with expert paralleism
     kwargs:
@@ -125,7 +125,7 @@
       out_device: "cuda:0"
   recursive: False # don't recursively inject submodules of this module
 - match:
-    name: "^model\\.layers\\.([1][0-9])\\.mlp\\.experts$"
+    name: "^model\\.layers\\.([2][0-9]|[1][5-9])\\.mlp\\.experts$"
   replace:
     class: ktransformers.operators.experts.KTransformersExperts     # custom MoE Kernel with expert paralleism
     kwargs:
@@ -136,7 +136,7 @@
       out_device: "cuda:1"
   recursive: False # don't recursively inject submodules of this module
 - match:
-    name: "^model\\.layers\\.([2][0-9])\\.mlp\\.experts$"
+    name: "^model\\.layers\\.([3][0-9]|[4][0-4])\\.mlp\\.experts$"
   replace:
     class: ktransformers.operators.experts.KTransformersExperts     # custom MoE Kernel with expert paralleism
     kwargs:
@@ -147,7 +147,7 @@
       out_device: "cuda:2"
   recursive: False # don't recursively inject submodules of this module
 - match:
-    name: "^model\\.layers\\.([345][0-9])\\.mlp\\.experts$"
+    name: "^model\\.layers\\.([5][0-9]|[4][5-9])\\.mlp\\.experts$"
   replace:
     class: ktransformers.operators.experts.KTransformersExperts     # custom MoE Kernel with expert paralleism
     kwargs:
@@ -159,28 +159,28 @@
   recursive: False # don't recursively inject submodules of this module
 
 - match:
-    name: "^model\\.layers\\.([0-9])\\.self_attn$"
+    name: "^model\\.layers\\.([0-9]|[1][0-4])\\.self_attn$"
   replace:
     class: ktransformers.operators.attention.KDeepseekV2Attention # optimized MLA implementation
     kwargs:
       generate_device: "cuda:0"
       prefill_device: "cuda:0"
 - match:
-    name: "^model\\.layers\\.([1][0-9])\\.self_attn$"
+    name: "^model\\.layers\\.([2][0-9]|[1][5-9])\\.self_attn$"
   replace:
     class: ktransformers.operators.attention.KDeepseekV2Attention # optimized MLA implementation
     kwargs:
       generate_device: "cuda:1"
       prefill_device: "cuda:1"
 - match:
-    name: "^model\\.layers\\.([2][0-9])\\.self_attn$"
+    name: "^model\\.layers\\.([3][0-9]|[4][0-4])\\.self_attn$"
   replace:
     class: ktransformers.operators.attention.KDeepseekV2Attention # optimized MLA implementation
     kwargs:
       generate_device: "cuda:2"
       prefill_device: "cuda:2"
 - match:
-    name: "^model\\.layers\\.([345][0-9])\\.self_attn$"
+    name: "^model\\.layers\\.([5][0-9]|[4][5-9])\\.self_attn$"
   replace:
     class: ktransformers.operators.attention.KDeepseekV2Attention # optimized MLA implementation
     kwargs:
@@ -194,33 +194,33 @@
     kwargs:
       per_layer_prefill_intput_threshold: 0 # 0 is close layer wise prefill
       transfer_map: 
-        10: "cuda:1"
-        20: "cuda:2"
-        30: "cuda:3"
+        15: "cuda:1"
+        30: "cuda:2"
+        45: "cuda:3"
 
 - match:
-    name: "^model\\.layers\\.([0-9])\\."
+    name: "^model\\.layers\\.([0-9]|[1][0-4])\\."
   replace:
     class: "default"
     kwargs:
       generate_device: "cuda:0"
       prefill_device: "cuda:0"
 - match:
-    name: "(^model\\.layers\\.([1][0-9])\\.)"
+    name: "(^model\\.layers\\.([2][0-9]|[1][5-9])\\.)"
   replace:
     class: "default"
     kwargs:
       generate_device: "cuda:1"
       prefill_device: "cuda:1"
 - match:
-    name: "(^model\\.layers\\.([2][0-9])\\.)"
+    name: "(^model\\.layers\\.([3][0-9]|[4][0-4])\\.)"
   replace:
     class: "default"
     kwargs:
       generate_device: "cuda:2"
       prefill_device: "cuda:2"
 - match:
-    name: "(^model\\.layers\\.([345][0-9])\\.)|(^model.norm)|(^lm_head)"
+    name: "(^model\\.layers\\.([5][0-9]|[4][5-9])\\.)|(^model.norm)|(^lm_head)"
   replace:
     class: "default"
     kwargs:

ktransformers/optimize/optimize_rules/DeepSeek-V2-Chat-multi-gpu.yaml

@@ -24,7 +24,7 @@
       prefill_device: "cuda:1"
 
 - match:
-    name: "^model\\.layers\\.(0|[1-9]|[12][0-9])\\.(?!self_attn).*$"  # regular expression 
+    name: "^model\\.layers\\.(0|[1-9]|[12][0-9])\\.(?!self_attn\\.kv_b_proj).*$"  # regular expression 
     class: torch.nn.Linear  # only match modules matching name and class simultaneously
   replace:
     class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
@@ -35,7 +35,7 @@
       prefill_op: "KLinearTorch"
 
 - match:
-    name: "^model\\.layers\\.([345][0-9])\\.(?!self_attn).*$"  # regular expression 
+    name: "^model\\.layers\\.([345][0-9])\\.(?!self_attn\\.kv_b_proj).*$"  # regular expression 
     class: torch.nn.Linear  # only match modules matching name and class simultaneously
   replace:
     class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types