update VLLM_USE_LIGHTOP_FILL_MOE_ALIGN

vllm/envs.py

@@ -179,7 +179,7 @@ if TYPE_CHECKING:
     VLLM_SCHED_ENABLE_MINIMAL_INJECTION: bool = False
     VLLM_USE_PD_SPLIT: bool = False
     VLLM_USE_PP_SYNC: bool = False
-    VLLM_USE_LIGHTOP_FILL_MOE_ALIN: bool = False
+    VLLM_USE_LIGHTOP_FILL_MOE_ALIGN: bool = False
 
 def get_default_cache_root():
     return os.getenv(
@@ -1163,8 +1163,8 @@ environment_variables: dict[str, Callable[[], Any]] = {
         lambda: (os.environ.get("VLLM_USE_PP_SYNC", "False").lower() in
                  ("true", "1")), 
     # vLLM will use lightop to fuse fill and moe align
-    "VLLM_USE_LIGHTOP_FILL_MOE_ALIN":
-        lambda: (os.environ.get("VLLM_USE_LIGHTOP_FILL_MOE_ALIN", "False").lower() in
+    "VLLM_USE_LIGHTOP_FILL_MOE_ALIGN":
+        lambda: (os.environ.get("VLLM_USE_LIGHTOP_FILL_MOE_ALIGN", "False").lower() in
                  ("true", "1")), 
 }
 

vllm/model_executor/layers/fused_moe/moe_align_block_size.py

@@ -216,7 +216,7 @@ def moe_align_block_size(
         sorted_ids = torch.empty((max_num_tokens_padded, ),
                                  dtype=torch.int32,
                                  device=topk_ids.device)
-        if not envs.VLLM_USE_LIGHTOP_FILL_MOE_ALIN:
+        if not envs.VLLM_USE_LIGHTOP_FILL_MOE_ALIGN:
             sorted_ids.fill_(topk_ids.numel())
             
     max_num_m_blocks = triton.cdiv(max_num_tokens_padded, block_size)

vllm/model_executor/model_loader/utils.py

@@ -251,8 +251,8 @@ def get_model_architecture(
                     os.environ['VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD'] = '1'
                 if not envs.is_set("VLLM_USE_OPT_CAT"):
                     os.environ['VLLM_USE_OPT_CAT'] = '1'
-                # if not envs.is_set("VLLM_USE_LIGHTOP_FILL_MOE_ALIN"):
-                #     os.environ['VLLM_USE_LIGHTOP_FILL_MOE_ALIN'] = '1'
+                if not envs.is_set("VLLM_USE_LIGHTOP_FILL_MOE_ALIGN"):
+                    os.environ['VLLM_USE_LIGHTOP_FILL_MOE_ALIGN'] = '1'
                 
             if os.getenv('GEMM_PAD') != '1': 
                 os.environ['GEMM_PAD'] = '0'
@@ -266,8 +266,8 @@ def get_model_architecture(
                     os.environ['VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD'] = '1'
                 if not envs.is_set("VLLM_USE_OPT_CAT"):
                     os.environ['VLLM_USE_OPT_CAT'] = '1'
-                # if not envs.is_set("VLLM_USE_LIGHTOP_FILL_MOE_ALIN"):
-                #     os.environ['VLLM_USE_LIGHTOP_FILL_MOE_ALIN'] = '1'
+                if not envs.is_set("VLLM_USE_LIGHTOP_FILL_MOE_ALIGN"):
+                    os.environ['VLLM_USE_LIGHTOP_FILL_MOE_ALIGN'] = '1'
                     
         # awq相关配置
         try: