[fix]解决w8a8 低延迟 cudagraph启动报错

8ba8a855 · 王敏 · 0ae68da1 · 8ba8a855
Commit 8ba8a855 authored Dec 18, 2025 by 王敏
Hide whitespace changes
Inline Side-by-side

Showing with 2 additions and 2 deletions

vllm/model_executor/layers/quantization/utils/w8a8_utils.py vllm/model_executor/layers/quantization/utils/w8a8_utils.py +2 -2

No files found.
--- a/vllm/model_executor/layers/quantization/utils/w8a8_utils.py
+++ b/vllm/model_executor/layers/quantization/utils/w8a8_utils.py
@@ -471,9 +471,9 @@ def apply_int8_linear(
                m_=m
                #best_config=W8A8_TRITONJSON.triton_json_dict[0][f"{m}_{n}_{k}"]
            elif m<=64:
-                m_= (m + 3) & -4 #取值到最近的4的倍数
+                m_= (m + 3) // 4 * 4 #(m + 3) & -4 #取值到最近的4的倍数
            elif m<=160:
-                m_=(m + 7) & -8
+                m_= (m + 7) // 8 * 8 #(m + 7) & -8
                
            elif m<200: #256
                m_=160