From 7cee07a06759c7f7e6ee4ff0f2c5b71d79f598be Mon Sep 17 00:00:00 2001 From: Ke Bao Date: Tue, 4 Nov 2025 23:27:46 +0800 Subject: [PATCH] Fix skip layer in get_quant_method (#12632) --- .../quantization/compressed_tensors/compressed_tensors.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/python/sglang/srt/layers/quantization/compressed_tensors/compressed_tensors.py b/python/sglang/srt/layers/quantization/compressed_tensors/compressed_tensors.py index faab1ebd5..3960e568a 100644 --- a/python/sglang/srt/layers/quantization/compressed_tensors/compressed_tensors.py +++ b/python/sglang/srt/layers/quantization/compressed_tensors/compressed_tensors.py @@ -123,7 +123,10 @@ class CompressedTensorsConfig(QuantizationConfig): if should_ignore_layer( prefix, ignore=self.ignore, fused_mapping=self.packed_modules_mapping ): - return UnquantizedLinearMethod() + if isinstance(layer, LinearBase): + return UnquantizedLinearMethod() + return None + if isinstance(layer, LinearBase): if CompressedTensorsConfig.DeepSeekFP8Config is not None: return Fp8LinearMethod(CompressedTensorsConfig.DeepSeekFP8Config)