Feat: reimplement vllm backend beam search using logprobs (#293)

Co-authored-by: Ilyas Moutawwakil <[email protected]>
huggingface · Nov 25, 2024 · 1c20082 · 1c20082
1 parent 31aa662
commit 1c20082
Showing 1 changed file with 18 additions and 20 deletions.
diff --git a/optimum_benchmark/backends/vllm/backend.py b/optimum_benchmark/backends/vllm/backend.py
@@ -117,35 +117,33 @@ def batch_offline_engine_generate(self, inputs: Dict[str, Any], kwargs: Dict[str
             self.pretrained_model.add_request(
                 inputs=prompt,
                 request_id=str(i),
-                params=SamplingParams(
-                    ignore_eos=True,
-                    detokenize=True,
-                    seed=self.config.seed,
-                    n=kwargs.get("num_return_sequences"),
-                    max_tokens=kwargs.get("max_new_tokens"),
-                    min_tokens=kwargs.get("min_new_tokens"),
-                    use_beam_search=kwargs.get("num_beams") > 1,
-                    logits_processors=kwargs.get("logits_processors", None),
-                ),
+                params=self.get_sampling_params(kwargs),
             )
 
         while self.pretrained_model.has_unfinished_requests():
             self.pretrained_model.step()
 
+    def get_sampling_params(self, kwargs: Dict[str, Any]) -> SamplingParams:
+        params = SamplingParams(
+            ignore_eos=True,
+            detokenize=True,
+            seed=self.config.seed,
+            n=kwargs.get("num_return_sequences"),
+            max_tokens=kwargs.get("max_new_tokens"),
+            min_tokens=kwargs.get("min_new_tokens"),
+            logits_processors=kwargs.get("logits_processors", None),
+        )
+        # following huggingface transformers implementation
+        # https://github.com/huggingface/transformers/blob/v4.45.2/src/transformers/generation/beam_search.py#L534
+        if kwargs.get("num_beams") > 1:
+            params.logprobs = 2 * kwargs.get("num_beams")
+        return params
+
     async def single_online_engine_generate(self, prompt: str, request_id: str, kwargs: Dict[str, Any]) -> Any:
         stream = await self.pretrained_model.add_request(
             inputs=prompt,
             request_id=request_id,
-            params=SamplingParams(
-                ignore_eos=True,
-                detokenize=True,
-                seed=self.config.seed,
-                n=kwargs.get("num_return_sequences"),
-                max_tokens=kwargs.get("max_new_tokens"),
-                min_tokens=kwargs.get("min_new_tokens"),
-                use_beam_search=kwargs.get("num_beams") > 1,
-                logits_processors=kwargs.get("logits_processors", None),
-            ),
+            params=self.get_sampling_params(kwargs),
         )
 
         async for _ in stream: