A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Bug] Fix Shape Validation for Fallback while Enabling E8M0 for DeepGEMM (#26322)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Oct 7, 2025 at 13:50 UTC
8f36850f732394648e36770bd6dfc325748d1104
1 file changed
+1
-1
vllm/model_executor/layers/fused_moe/triton_deep_gemm_moe.py
+1
-1
@@ -143,7 +143,7 @@ class TritonOrDeepGemmExperts(mk.FusedMoEPermuteExpertsUnpermute):
143
apply_router_weight_on_input: bool,
144
):
145
use_deep_gemm = self.allow_deep_gemm and (
146
- _valid_deep_gemm(hidden_states, w1, w2) or is_deep_gemm_e8m0_used()
146
+ is_deep_gemm_e8m0_used() or _valid_deep_gemm(hidden_states, w1, w2)
147
)
148
149
experts = self.deep_gemm_expert if use_deep_gemm else self.triton_expert