[BugFix] Fix cuda graph for MLPSpeculator (vllm-project#5875)

Co-authored-by: Abhinav Goyal <abhinav.goyal@flipkart.com>
Signed-off-by: Alvant <alvasian@yandex.ru>

Author: 2 people

examples/offline_inference_mlpspeculator.py

@@ -52,7 +52,6 @@ def time_generation(llm: LLM, prompts: List[str],
         speculative_model="ibm-fms/llama-13b-accelerator",
         # These are currently required for MLPSpeculator decoding
         use_v2_block_manager=True,
-        enforce_eager=True,
     )
 
     print("With speculation")

vllm/worker/model_runner.py

@@ -1020,10 +1020,13 @@ def execute_model(
 
         if self.return_hidden_states:
             # we only need to pass hidden states of most recent token
+            assert model_input.sampling_metadata is not None
+            indices = model_input.sampling_metadata.selected_token_indices
             if model_input.is_prompt:
-                assert model_input.sampling_metadata is not None
-                hidden_states = hidden_states.index_select(
-                    0, model_input.sampling_metadata.selected_token_indices)
+                hidden_states = hidden_states.index_select(0, indices)
+            elif decode_meta.use_cuda_graph:
+                hidden_states = hidden_states[:len(indices)]
+
             output.hidden_states = hidden_states
 
         return output