From aa7fb5d7073b70b35445ad87bcc1fbf6aca34df1 Mon Sep 17 00:00:00 2001
From: Zhu Yi Lin <116337067+GDzhu01@users.noreply.github.com>
Date: Wed, 25 Feb 2026 23:02:00 +0800
Subject: [PATCH] [Bugfix] Fix DeepseekV3.1 Accuracy issue (#6805)

### What this PR does / why we need it?
In order to adapt to the GLM model, logits were passed in the sample,
which can cause accuracy issues in version 0.15.0.


- vLLM version: v0.15.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/83b47f67b1dfad505606070ae4d9f83e50ad4ebd

Signed-off-by: GDzhu01 <809721801@qq.com>
---
 vllm_ascend/sample/rejection_sampler.py | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/vllm_ascend/sample/rejection_sampler.py b/vllm_ascend/sample/rejection_sampler.py
index 695ae649..04552b97 100644
--- a/vllm_ascend/sample/rejection_sampler.py
+++ b/vllm_ascend/sample/rejection_sampler.py
@@ -19,6 +19,7 @@ from vllm_ascend.ops.triton.reject_sample import (
     sample_recovered_tokens_kernel,
 )
 from vllm_ascend.sample.sampler import apply_top_k_top_p
+from vllm_ascend.utils import vllm_version_is
 
 
 def apply_sampling_constraints(
@@ -166,7 +167,10 @@ def rejection_sample(
             return output_token_ids
 
     # Compute probability distribution from target logits.
-    target_probs = target_logits.softmax(dim=-1, dtype=torch.float32)
+    if vllm_version_is("0.15.0"):
+        target_probs = target_logits
+    else:
+        target_probs = target_logits.softmax(dim=-1, dtype=torch.float32)
     assert target_probs.is_contiguous()
 
     # Generate uniform probabilities for rejection sampling.