[Kernel] optimize performance of gptq marlin kernel when n is small (#14138)

Signed-off-by: Jinzhen Lin <linjinzhen@hotmail.com>

vllm/envs.py

@@ -95,6 +95,7 @@ if TYPE_CHECKING:
     VLLM_DP_SIZE: int = 1
     VLLM_DP_MASTER_IP: str = ""
     VLLM_DP_MASTER_PORT: int = 0
+    VLLM_MARLIN_USE_ATOMIC_ADD: bool = False
 
 
 def get_default_cache_root():
@@ -630,6 +631,10 @@ environment_variables: dict[str, Callable[[], Any]] = {
     # Whether to use S3 path for model loading in CI via RunAI Streamer
     "VLLM_CI_USE_S3":
     lambda: os.environ.get("VLLM_CI_USE_S3", "0") == "1",
+
+    # Whether to use atomicAdd reduce in gptq/awq marlin kernel.
+    "VLLM_MARLIN_USE_ATOMIC_ADD":
+    lambda: os.environ.get("VLLM_MARLIN_USE_ATOMIC_ADD", "0") == "1",
 }
 
 # end-env-vars-definition