From a8c86eeb1695a3d35d0c748a68a1451379ea497a Mon Sep 17 00:00:00 2001 From: Mike G <180722391+mikekg@users.noreply.github.com> Date: Tue, 16 Jun 2026 05:43:44 -0700 Subject: [PATCH] [Quant] Support modelopt_mixed on Ampere (SM80/SM86) (#45306) Signed-off-by: Mike G <180722391+mikekg@users.noreply.github.com> --- vllm/model_executor/layers/quantization/modelopt.py | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/vllm/model_executor/layers/quantization/modelopt.py b/vllm/model_executor/layers/quantization/modelopt.py index 2bdb26e1a8..24ec55e400 100644 --- a/vllm/model_executor/layers/quantization/modelopt.py +++ b/vllm/model_executor/layers/quantization/modelopt.py @@ -2299,7 +2299,13 @@ class ModelOptMixedPrecisionConfig(ModelOptQuantConfigBase): @classmethod def get_min_capability(cls) -> int: - return 89 + # Ampere (SM80/SM86): NVFP4 routed experts run via Marlin W4A16, and FP8 + # weight-only dense layers run via MarlinFP8 (W8A16, compute in + # bf16/fp16). FP8 MoE, if present, also routes to Marlin because + # TritonExperts gates its FP8 schemes behind supports_fp8() (cc>=89). + # None of these paths require native FP8 tensor cores, so SM80 is + # sufficient. + return 80 @classmethod def override_quantization_method(