A high-throughput and memory-efficient inference and serving engine for LLMs
[Quant] Support modelopt_mixed on Ampere (SM80/SM86) (#45306)
Signed-off-by: Mike G <180722391+mikekg@users.noreply.github.com>
Mike G committed
Jun 16, 2026 at 05:43 UTC
a8c86eeb1695a3d35d0c748a68a1451379ea497a
1 file changed
+7
-1
vllm/model_executor/layers/quantization/modelopt.py
+7
-1
@@ -2299,7 +2299,13 @@ class ModelOptMixedPrecisionConfig(ModelOptQuantConfigBase):
2299
2300
@classmethod
2301
def get_min_capability(cls) -> int:
2302
- return 89
2302
+ # Ampere (SM80/SM86): NVFP4 routed experts run via Marlin W4A16, and FP8
2303
+ # weight-only dense layers run via MarlinFP8 (W8A16, compute in
2304
+ # bf16/fp16). FP8 MoE, if present, also routes to Marlin because
2305
+ # TritonExperts gates its FP8 schemes behind supports_fp8() (cc>=89).
2306
+ # None of these paths require native FP8 tensor cores, so SM80 is
2307
+ # sufficient.
2308
+ return 80
2309
2310
@classmethod
2311
def override_quantization_method(