From 8980001c9333c617f2bc3b2c60926b02434067a7 Mon Sep 17 00:00:00 2001
From: caozuoba <44251931+caozuoba@users.noreply.github.com>
Date: Sat, 31 Jan 2026 17:51:26 +0800
Subject: [PATCH] [perf] v1/spec_decode: skip softmax for all-greedy rejection
 sampling (#32852)

Signed-off-by: hdj <1293066020@qq.com>
---
 vllm/v1/sample/rejection_sampler.py | 21 +++++++++++----------
 1 file changed, 11 insertions(+), 10 deletions(-)

diff --git a/vllm/v1/sample/rejection_sampler.py b/vllm/v1/sample/rejection_sampler.py
index 16d01d518..b57c93e29 100644
--- a/vllm/v1/sample/rejection_sampler.py
+++ b/vllm/v1/sample/rejection_sampler.py
@@ -136,8 +136,6 @@ class RejectionSampler(nn.Module):
             metadata.cu_num_draft_tokens,
             sampling_metadata,
         )
-        # Compute probability distribution from target logits.
-        target_probs = target_logits.softmax(dim=-1, dtype=torch.float32)
 
         output_token_ids = rejection_sample(
             metadata.draft_token_ids,
@@ -145,7 +143,7 @@ class RejectionSampler(nn.Module):
             metadata.max_spec_len,
             metadata.cu_num_draft_tokens,
             draft_probs,
-            target_probs,
+            target_logits,
             bonus_token_ids,
             sampling_metadata,
         )
@@ -353,7 +351,7 @@ def rejection_sample(
     # [num_tokens, vocab_size]
     draft_probs: torch.Tensor | None,
     # [num_tokens, vocab_size]
-    target_probs: torch.Tensor,
+    target_logits: torch.Tensor,
     # [batch_size, 1]
     bonus_token_ids: torch.Tensor,
     sampling_metadata: SamplingMetadata,
@@ -361,17 +359,16 @@ def rejection_sample(
     assert draft_token_ids.ndim == 1
     assert draft_probs is None or draft_probs.ndim == 2
     assert cu_num_draft_tokens.ndim == 1
-    assert target_probs.ndim == 2
+    assert target_logits.ndim == 2
 
     batch_size = len(num_draft_tokens)
     num_tokens = draft_token_ids.shape[0]
-    vocab_size = target_probs.shape[-1]
-    device = target_probs.device
+    vocab_size = target_logits.shape[-1]
+    device = target_logits.device
     assert draft_token_ids.is_contiguous()
     assert draft_probs is None or draft_probs.is_contiguous()
-    assert target_probs.is_contiguous()
     assert bonus_token_ids.is_contiguous()
-    assert target_probs.shape == (num_tokens, vocab_size)
+    assert target_logits.shape == (num_tokens, vocab_size)
 
     # Create output buffer.
     output_token_ids = torch.full(
@@ -387,7 +384,7 @@ def rejection_sample(
         is_greedy = sampling_metadata.temperature == GREEDY_TEMPERATURE
     if not sampling_metadata.all_random:
         # Rejection sampling for greedy sampling requests.
-        target_argmax = target_probs.argmax(dim=-1)
+        target_argmax = target_logits.argmax(dim=-1)
         rejection_greedy_sample_kernel[(batch_size,)](
             output_token_ids,
             cu_num_draft_tokens,
@@ -400,6 +397,10 @@ def rejection_sample(
         if sampling_metadata.all_greedy:
             return output_token_ids
 
+    # Compute probability distribution from target logits.
+    target_probs = target_logits.softmax(dim=-1, dtype=torch.float32)
+    assert target_probs.is_contiguous()
+
     # Generate uniform probabilities for rejection sampling.
     # [num_tokens]
     uniform_probs = generate_uniform_probs(