reformat

Author: sufubao

lightllm/common/basemodel/layer_weights/meta_weights/fused_moe_weight_ep.py

@@ -359,9 +359,7 @@ def prefilled_group_gemm(
             ######################################## warning ##################################################
             # here is used to match autotune feature, make moe model run same triton kernel in different rank.
             # in some special case, one rank will recv 0 token, so add a token to make it run triton kernel.
-            if (
-                get_triton_autotune_level() in [AutotuneLevel.ADAPTIVE_AUTOTUNE, AutotuneLevel.FORCE_AUTOTUNE]
-            ):
+            if get_triton_autotune_level() in [AutotuneLevel.ADAPTIVE_AUTOTUNE, AutotuneLevel.FORCE_AUTOTUNE]:
                 _gemm_out_a = torch.zeros((1, N), device=device, dtype=hidden_dtype)
                 _silu_out = torch.zeros((1, N // 2), device=device, dtype=hidden_dtype)
                 silu_and_mul_fwd(_gemm_out_a.view(-1, N), _silu_out)

lightllm/common/fused_moe/grouped_fused_moe_ep.py

@@ -192,9 +192,7 @@ def fused_experts_impl(
             ######################################## warning ##################################################
             # here is used to match autotune feature, make moe model run same triton kernel in different rank.
             # in some special case, one rank will recv 0 token, so add a token to make it run triton kernel.
-            if (
-                get_triton_autotune_level() in [AutotuneLevel.ADAPTIVE_AUTOTUNE, AutotuneLevel.FORCE_AUTOTUNE]
-            ):
+            if get_triton_autotune_level() in [AutotuneLevel.ADAPTIVE_AUTOTUNE, AutotuneLevel.FORCE_AUTOTUNE]:
                 _gemm_out_a = torch.zeros((1, N), device=hidden_states.device, dtype=hidden_states.dtype)
                 _silu_out = torch.zeros((1, N // 2), device=hidden_states.device, dtype=hidden_states.dtype)
                 silu_and_mul_fwd(_gemm_out_a.view(-1, N), _silu_out)