[MRV2][Performance] Skip no-op FP32 logits materialization (#47711)

Signed-off-by: jesse <szxfml@gmail.com>
Signed-off-by: Song Zhixin <szxfml@gmail.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>

vllm/v1/worker/gpu/sample/sampler.py

@@ -152,6 +152,9 @@ class Sampler:
         expanded_local_pos: torch.Tensor,
         skip_top_k_top_p: bool = False,
     ) -> torch.Tensor:
+        if not self._requires_logits_processing(idx_mapping_np):
+            return logits
+
         # Copy logits to a new FP32 tensor.
         logits = torch.empty_like(logits, dtype=torch.float32).copy_(logits)
 
@@ -194,6 +197,24 @@ class Sampler:
             logits, expanded_idx_mapping, idx_mapping_np
         )
 
+    def _requires_logits_processing(self, idx_mapping_np: np.ndarray) -> bool:
+        if np.any(self.logit_bias_state.use_logit_bias[idx_mapping_np]):
+            return True
+        if np.any(self.penalties_state.use_penalty[idx_mapping_np]):
+            return True
+        if np.any(self.bad_words_state.num_bad_words.np[idx_mapping_np] > 0):
+            return True
+
+        states = self.sampling_states
+        temperatures = states.temperature.np[idx_mapping_np]
+        if np.any((temperatures != 0.0) & (temperatures != 1.0)):
+            return True
+        if np.any(states.min_p.np[idx_mapping_np] != 0.0):
+            return True
+        if np.any(states.top_k.np[idx_mapping_np] != states.vocab_size):
+            return True
+        return bool(np.any(states.top_p.np[idx_mapping_np] != 1.0))
+
     def sample(
         self,
         logits: torch.Tensor,