Merge branch 'v0.8.5-zero_overhead' into 'v0.8.5.post1-dev'

fix tbo support deepseek mtp

See merge request dcutoolkit/deeplearing/vllm!123

vllm/two_batch_overlap/model_input_split.py

@@ -98,6 +98,13 @@ def split_model_input(model_input, self_device, batch_size_left, batch_size_righ
         seq_groups_right = model_input.sampling_metadata.seq_groups[batch_size_left:]
     selected_token_indices_left = split_seq_lens_tensor[0].cumsum(dim=0) - 1
     selected_token_indices_right = split_seq_lens_tensor[1].cumsum(dim=0) - 1
+
+    previous_hidden_states_left = None
+    previous_hidden_states_right = None
+    if model_input.previous_hidden_states != None:
+        split_previous_hidden_states = torch.split(model_input.previous_hidden_states, batch_size_split, dim=0)
+        previous_hidden_states_left = split_previous_hidden_states[0]
+        previous_hidden_states_right = split_previous_hidden_states[1]    
     
     if isinstance(model_input.attn_metadata, MLACommonMetadata):
         attn_metadata_left = MLACommonMetadata(
@@ -302,7 +309,7 @@ def split_model_input(model_input, self_device, batch_size_left, batch_size_righ
         virtual_engine=model_input.virtual_engine,
         async_callback=model_input.async_callback,
         scheduler_outputs=model_input.scheduler_outputs,
-        previous_hidden_states=model_input.previous_hidden_states,
+        previous_hidden_states=previous_hidden_states_left,
         sampling_metadata=SamplingMetadata(
             seq_groups=seq_groups_left,
             selected_token_indices=selected_token_indices_left,
@@ -330,7 +337,7 @@ def split_model_input(model_input, self_device, batch_size_left, batch_size_righ
         virtual_engine=model_input.virtual_engine,
         async_callback=model_input.async_callback,
         scheduler_outputs=model_input.scheduler_outputs,
-        previous_hidden_states=model_input.previous_hidden_states,
+        previous_hidden_states=previous_hidden_states_right,
         sampling_metadata=SamplingMetadata(
             seq_groups=seq_groups_right,
             selected_token_indices=selected_token_indices_right,

vllm/two_batch_overlap/two_batch_overlap.py

@@ -86,6 +86,11 @@ class TwoBatchOverlap():
                     break
                 profile.ProfRangePush('start')
                 self.tbo_thread_synchronize(tid)
+                model_kwargs = None
+                if is_left_thread:
+                    model_kwargs = self.model_kwargs_left
+                else:
+                    model_kwargs = self.model_kwargs_right
                 with set_forward_context(model_input.attn_metadata,
                                         self.vllm_config, self.virtual_engine):
                     hidden_or_intermediate_states = self.model_executable(
@@ -95,7 +100,7 @@ class TwoBatchOverlap():
                         **MultiModalKwargs.as_kwargs(self.multi_modal_kwargs,
                                                         device=self.self_device),
                         **self.seqlen_agnostic_kwargs,
-                        **self.model_kwargs,
+                        **model_kwargs,
                     )
                 if is_left_thread:
                     self.sem_right.release()
@@ -131,7 +136,8 @@ class TwoBatchOverlap():
                         multi_modal_kwargs,
                         self_device,
                         seqlen_agnostic_kwargs,
-                        model_kwargs):
+                        model_kwargs_left,
+                        model_kwargs_right):
         if self.left_thread == None:
             self.init_tbo_thread()
         self.vllm_config = vllm_config
@@ -141,7 +147,8 @@ class TwoBatchOverlap():
         self.multi_modal_kwargs = multi_modal_kwargs
         self.self_device = self_device
         self.seqlen_agnostic_kwargs = seqlen_agnostic_kwargs
-        self.model_kwargs = model_kwargs
+        self.model_kwargs_left = model_kwargs_left
+        self.model_kwargs_right = model_kwargs_right
         self.model_input_left_queue.put(model_input_left)
         self.model_input_right_queue.put(model_input_right)
 
@@ -242,6 +249,16 @@ def tbo_model_executable(
         batch_size_right += 1
     
     model_input_left, model_input_right = split_model_input(model_input, self_device, batch_size_left, batch_size_right)
+    
+    model_kwargs_left = model_kwargs
+    model_kwargs_right = model_kwargs
+    if "previous_hidden_states" in model_kwargs:
+        previous_hidden_states = model_kwargs["previous_hidden_states"]
+        batch_size_split = [batch_size_left, batch_size_right]
+        split_previous_hidden_states = torch.split(previous_hidden_states, batch_size_split, dim=0)
+        model_kwargs_left["previous_hidden_states"] = split_previous_hidden_states[0]
+        model_kwargs_right["previous_hidden_states"] = split_previous_hidden_states[1]
+
     tbo_obj.step_event.record()
     current_stream = torch.cuda.current_stream()
     with torch.cuda.stream(tbo_step_stream):
@@ -255,7 +272,8 @@ def tbo_model_executable(
                                 multi_modal_kwargs,
                                 self_device,
                                 seqlen_agnostic_kwargs,
-                                model_kwargs)
+                                model_kwargs_left,
+                                model_kwargs_right)
         tbo_obj.all_reduce()
         states_left, states_right = tbo_obj.get_model_output()