From ce6a283c097cd573f6f371bbb376c4e5c4d07ef7 Mon Sep 17 00:00:00 2001 From: Micah Williamson Date: Tue, 1 Sep 2026 16:08:11 -0500 Subject: [PATCH] [Bugfix] Restore `weight_dtype` in `QuarkW8A8Fp8MoEMethod` to fix GPT-OSS FP8 MoE weight loading (#54824) Signed-off-by: Micah Williamson --- vllm/model_executor/layers/quantization/quark/quark_moe.py | 1 + 1 file changed, 1 insertion(+) diff --git a/vllm/model_executor/layers/quantization/quark/quark_moe.py b/vllm/model_executor/layers/quantization/quark/quark_moe.py index 74a64041cb..fd3fcc7e04 100644 --- a/vllm/model_executor/layers/quantization/quark/quark_moe.py +++ b/vllm/model_executor/layers/quantization/quark/quark_moe.py @@ -243,6 +243,7 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod): activation_quant_key: QuantKey | None, ): super().__init__(moe, weight_quant_key, activation_quant_key) + self.weight_dtype = "fp8" self.weight_qscheme = ( "per_channel" if weight_quant_key == kFp8StaticChannelSym else "per_tensor" )