[V1][Perf] Reduce scheduling overhead in model runner after cuda sync (…

…vllm-project#12094)

Signed-off-by: Keyun Tong <tongkeyun@gmail.com>

vllm/v1/outputs.py

@@ -8,7 +8,7 @@
 class SamplerOutput:
 
     # [num_reqs]
-    sampled_token_ids: List[int]
+    sampled_token_ids: torch.Tensor
 
     # [num_reqs, max_num_logprobs + 1]
     logprob_token_ids: Optional[torch.Tensor]

vllm/v1/sample/sampler.py

@@ -50,9 +50,8 @@ def forward(
         # Use int32 to reduce the tensor size.
         sampled = sampled.to(torch.int32)
 
-        # NOTE: CPU-GPU synchronization happens here.
         sampler_output = SamplerOutput(
-            sampled_token_ids=sampled.tolist(),
+            sampled_token_ids=sampled,
             logprob_token_ids=topk_indices,
             logprobs=topk_logprobs,
             prompt_logprob_token_ids=None,

vllm/v1/worker/gpu_model_runner.py

@@ -775,10 +775,10 @@ def execute_model(
             sampling_metadata=sampling_metadata,
         )
 
-        sampled_token_ids = sampler_output.sampled_token_ids
         # TODO(woosuk): The following loop can be slow since it iterates over
         # the requests one by one. Optimize.
         num_reqs = self.input_batch.num_reqs
+        request_seq_lens: List[Tuple[int, CachedRequestState, int]] = []
         for i, req_id in enumerate(self.input_batch.req_ids[:num_reqs]):
             assert req_id is not None
             req_state = self.requests[req_id]
@@ -787,10 +787,10 @@ def execute_model(
             assert seq_len <= req_state.num_tokens
             if seq_len == req_state.num_tokens:
                 # Append the sampled token to the output token ids.
-                token_id = sampled_token_ids[i]
-                self.input_batch.token_ids_cpu[i, seq_len] = token_id
                 self.input_batch.num_tokens[i] += 1
-                req_state.output_token_ids.append(token_id)
+                # OPTIMIZATION: Priming the state updates for later updates.
+                req_state.output_token_ids.append(0)
+                request_seq_lens.append((i, req_state, seq_len))
             else:
                 # Ignore the sampled token from the partial request.
                 # Rewind the generator state as if the token was not sampled.
@@ -799,6 +799,21 @@ def execute_model(
                     # This relies on cuda-specific torch-internal impl details
                     generator.set_offset(generator.get_offset() - 4)
 
+        # num_reqs entries should be non-None
+        assert all(
+            req_id is not None for req_id in
+            self.input_batch.req_ids[:num_reqs]), "req_ids contains None"
+        req_ids = cast(List[str], self.input_batch.req_ids[:num_reqs])
+
+        # NOTE: GPU -> CPU Sync happens here.
+        # Move as many CPU operations as possible before this sync point.
+        sampled_token_ids = sampler_output.sampled_token_ids.tolist()
+        # Update with the actual token ids
+        for i, req_state, seq_len in request_seq_lens:
+            token_id = sampled_token_ids[i]
+            self.input_batch.token_ids_cpu[i, seq_len] = token_id
+            req_state.output_token_ids[-1] = token_id
+
         if sampler_output.logprob_token_ids is None:
             logprob_token_ids = None
         else:
@@ -808,12 +823,6 @@ def execute_model(
         else:
             logprobs = sampler_output.logprobs.cpu()
 
-        # num_reqs entries should be non-None
-        assert all(
-            req_id is not None for req_id in
-            self.input_batch.req_ids[:num_reqs]), "req_ids contains None"
-        req_ids = cast(List[str], self.input_batch.req_ids[:num_reqs])
-
         model_runner_output = ModelRunnerOutput(
             req_ids=req_ids,
             req_id_to_index=self.input_batch.req_id_to_index,