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

debug and fix tbo error in mtp

See merge request dcutoolkit/deeplearing/vllm!124

vllm/two_batch_overlap/model_input_split.py

@@ -102,7 +102,7 @@ def split_model_input(model_input, self_device, batch_size_left, batch_size_righ
     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)
+        split_previous_hidden_states = torch.split(model_input.previous_hidden_states, query_tokens_split, dim=0)
         previous_hidden_states_left = split_previous_hidden_states[0]
         previous_hidden_states_right = split_previous_hidden_states[1]    
     

vllm/two_batch_overlap/two_batch_overlap.py

@@ -250,12 +250,12 @@ def tbo_model_executable(
     
     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
+    model_kwargs_left = model_kwargs.copy()
+    model_kwargs_right = model_kwargs.copy()
     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)
+        query_tokens_split = [sum(model_input.query_lens[0:batch_size_left]), sum(model_input.query_lens[batch_size_left:])]
+        split_previous_hidden_states = torch.split(previous_hidden_states, query_tokens_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]