From 766e1678210d797757dcfe28f05184a251685dfe Mon Sep 17 00:00:00 2001 From: xuebwang-amd Date: Fri, 13 Feb 2026 01:40:19 +0800 Subject: [PATCH] [ROCm][quantization] improve OCP weight quant parser robust (#34431) Signed-off-by: xuebwang-amd Co-authored-by: TJian --- vllm/model_executor/layers/quantization/quark/quark.py | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/vllm/model_executor/layers/quantization/quark/quark.py b/vllm/model_executor/layers/quantization/quark/quark.py index 2e75a3de5..36f20c89f 100644 --- a/vllm/model_executor/layers/quantization/quark/quark.py +++ b/vllm/model_executor/layers/quantization/quark/quark.py @@ -337,6 +337,13 @@ class QuarkConfig(QuantizationConfig): ) return False + if isinstance(weight_quant, list): + logger.debug( + "Quark model's weight quantization is incompatible with OCP_MX format: " + "weight_quant is a list (e.g. fp8_w4a8), OCP_MX requires a single dict." + ) + return False + # Input and weight qscheme needs to be per group. if weight_quant.get("qscheme") != "per_group": logger.debug(