remove unused impl and envs

vllm/envs.py

@@ -292,8 +292,6 @@ if TYPE_CHECKING:
     VLLM_USE_FUSED_FILL_RMS_CAT: bool = False
     VLLM_W8A8_BACKEND: int = 3
     VLLM_REJECT_SAMPLE_OPT: bool = False
-    VLLM_USE_OPT_MOE_SUM: bool = False
-    VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD: bool = False
     VLLM_USE_MOE_W16A16_TRITON: bool = False
 
 
@@ -1845,14 +1843,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_REJECT_SAMPLE_OPT":
         lambda: (os.getenv('VLLM_REJECT_SAMPLE_OPT', 'True').lower() in
                  ("true", "1")),
-    # vLLM will use triton moe_sum
-    "VLLM_USE_OPT_MOE_SUM":
-        lambda: (os.environ.get("VLLM_USE_OPT_MOE_SUM", "False").lower() in
-                 ("true", "1")),
-    # vLLM will use lightop moe_sum_mul_add
-    "VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD":
-        lambda: (os.environ.get("VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD", "False").lower() in
-                 ("true", "1")),
     # Force using Triton MoE path (disable Marlin W16A16 MoE).
     "VLLM_USE_MOE_W16A16_TRITON":
         lambda: (os.environ.get("VLLM_USE_MOE_W16A16_TRITON", "0").lower() in

vllm/model_executor/layers/fused_moe/fuse_moe_w16a16_marlin.py

@@ -397,21 +397,13 @@ def fused_experts_impl_w16a16_marlin(hidden_states: torch.Tensor,
         )
         intermediate_cache3 = intermediate_cache3.view(-1, top_k_num, K)
 
-        if envs.VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD: 
+        if envs.VLLM_USE_LIGHTOP_MOE_SUM:
             from lightop import op as op
             op.moe_sum(input=intermediate_cache3.view(*intermediate_cache3.size()),
-                    output=out_hidden_states[begin_chunk_idx:end_chunk_idx], bias=shared_output[begin_chunk_idx:end_chunk_idx], 
-                    expert_mask=None, num_local_tokens=None, factor=routed_scaling_factor)
+                output=out_hidden_states[begin_chunk_idx:end_chunk_idx], bias=None, 
+                expert_mask=None, num_local_tokens=None, factor=1.0)
         else:
-            if envs.VLLM_USE_LIGHTOP_MOE_SUM:
-                from lightop import op as op
-                op.moe_sum(input=intermediate_cache3.view(*intermediate_cache3.size()),
-                    output=out_hidden_states[begin_chunk_idx:end_chunk_idx], bias=None, 
-                    expert_mask=None, num_local_tokens=None, factor=1.0)
-            elif envs.VLLM_USE_OPT_MOE_SUM:
-                moe_reduce_dispatch(intermediate_cache3.view(*intermediate_cache3.size()), out_hidden_states[begin_chunk_idx:end_chunk_idx], begin_chunk_idx, end_chunk_idx)
-            else:
-                ops.moe_sum(intermediate_cache3.view(*intermediate_cache3.size()),
-                                out_hidden_states[begin_chunk_idx:end_chunk_idx])
+            ops.moe_sum(intermediate_cache3.view(*intermediate_cache3.size()),
+                            out_hidden_states[begin_chunk_idx:end_chunk_idx])
 
     return out_hidden_states
\ No newline at end of file