enable int8 bmm

2026-01-14 14:30:59 +08:00
parent f0bf384e2e
commit 115eb32068
3 changed files with 164 additions and 84 deletions
--- a/vllm_kunlun/platforms/envs.py
+++ b/vllm_kunlun/platforms/envs.py
@@ -68,6 +68,11 @@ xvllm_environment_variables: dict[str, Callable[[], Any]] = {
    "ENABLE_VLLM_FUSED_QKV_SPLIT_NORM_ROPE":
    lambda: (os.environ.get("ENABLE_VLLM_FUSED_QKV_SPLIT_NORM_ROPE", "False").lower() in 
             ("true", "1")),
+
+    # use int8 bmm
+    "VLLM_KUNLUN_ENABLE_INT8_BMM":
+    lambda: (os.environ.get("VLLM_KUNLUN_ENABLE_INT8_BMM", "False").lower() in 
+             ("true", "1")),
 }

 # end-env-vars-definition