[Perf] Remove hardcoded num_warps=1 (vllm-project#26183)

Signed-off-by: Corey Lowman <clowman1993@gmail.com>

Author: coreylowman

vllm/v1/sample/rejection_sampler.py

@@ -164,12 +164,12 @@ def rejection_sample(
     assert target_probs.shape == (num_tokens, vocab_size)
 
     # Create output buffer.
-    output_token_ids = torch.empty(
+    output_token_ids = torch.full(
         (batch_size, max_spec_len + 1),
+        PLACEHOLDER_TOKEN_ID,
         dtype=torch.int32,  # Consistent with SamplerOutput.sampled_token_ids.
         device=device,
     )
-    output_token_ids.fill_(PLACEHOLDER_TOKEN_ID)
 
     if sampling_metadata.all_greedy:
         is_greedy = None
@@ -186,7 +186,6 @@ def rejection_sample(
             bonus_token_ids,
             is_greedy,
             max_spec_len,
-            num_warps=1,
         )
         if sampling_metadata.all_greedy:
             return output_token_ids
@@ -227,7 +226,6 @@ def rejection_sample(
         max_spec_len,
         vocab_size,
         NO_DRAFT_PROBS=draft_probs is None,
-        num_warps=1,
     )
     return output_token_ids
 
@@ -329,7 +327,6 @@ def expand_batch_to_tokens(
         replace_from,
         replace_to,
         MAX_NUM_TOKENS=MAX_SPEC_LEN,  # To avoid recompilation.
-        num_warps=1,
     )
     return expanded_x