[Perf] Avoid pageable HtoD transfer in MinTokensLogitsProcessor (#29826)

Signed-off-by: jthomson04 <[email protected]>
(cherry picked from commit 1528e07)

Author: jthomson04

vllm/v1/sample/logits_processor/builtin.py

@@ -110,7 +110,7 @@ def apply(self, logits: torch.Tensor) -> torch.Tensor:
         # Identify valid tokens using threshold comparison
         invalid_token_mask = probability_values < adjusted_min_p
         # Apply mask using boolean indexing
-        logits[invalid_token_mask] = -float("inf")
+        logits.masked_fill_(invalid_token_mask, -float("inf"))
         return logits
 
 
@@ -178,6 +178,10 @@ def __init__(
             self._device_tensor([], torch.int32),
         )
 
+        self.neg_inf_tensor = torch.tensor(
+            -float("inf"), dtype=torch.float32, device=self.device
+        )
+
     def is_argmax_invariant(self) -> bool:
         """By censoring stop tokens, min-tokens can change the outcome
         of the argmax operation in greedy sampling."""
@@ -229,7 +233,7 @@ def _device_tensor(self, data: list, dtype: torch.dtype) -> torch.Tensor:
     def apply(self, logits: torch.Tensor) -> torch.Tensor:
         if self.min_toks:
             # Inhibit EOS token for requests which have not reached min length
-            logits[self.logits_slice] = -float("inf")
+            logits.index_put_(self.logits_slice, self.neg_inf_tensor)
         return logits