[Bugfix][Quantization] Fix INT8 W8A8 MoE crash in TritonExperts (#51411)
Signed-off-by: Djordje Ramic <djoramic@amd.com>
djramic committed
Aug 8, 2026 at 05:06 UTC
5eaa70dac58102fe6e243291e2bf25ca192c46f2
1 file changed
+2
vllm/model_executor/layers/fused_moe/experts/triton_moe.py
+2
@@ -252,6 +252,7 @@ class TritonExperts(LoRAExpertsMixin, mk.FusedMoEExpertsModular):
252
torch.bfloat16,
253
torch.float8_e4m3fn,
254
torch.float8_e4m3fnuz,
255
+ torch.int8,
256
]
257
258
# We declared expects_unquantized_inputs (LoRA + DP/EP all2all), so the
@@ -298,6 +299,7 @@ class TritonExperts(LoRAExpertsMixin, mk.FusedMoEExpertsModular):
299
elif (
300
hidden_states.dtype == torch.float8_e4m3fn
301
or hidden_states.dtype == torch.float8_e4m3fnuz
302
+ or hidden_states.dtype == torch.int8
303
):
304
compute_type = tl.bfloat16
305
else: