From 058cc0a8b6e33523b1ed75db933726959df43791 Mon Sep 17 00:00:00 2001 From: Yuwen Zhou Date: Fri, 19 Jun 2026 00:20:29 +0800 Subject: [PATCH] [Bugfix] Restore is_sym guard for zp in GPTQ/CT MoE to fix symmetric quant regression (#45656) Signed-off-by: yuwenzho --- vllm/model_executor/layers/quantization/auto_gptq.py | 10 ++++++++-- .../compressed_tensors_moe_wna16_marlin.py | 4 ++-- 2 files changed, 10 insertions(+), 4 deletions(-) diff --git a/vllm/model_executor/layers/quantization/auto_gptq.py b/vllm/model_executor/layers/quantization/auto_gptq.py index 459a615832..f7fe7f6e9e 100644 --- a/vllm/model_executor/layers/quantization/auto_gptq.py +++ b/vllm/model_executor/layers/quantization/auto_gptq.py @@ -25,6 +25,7 @@ from vllm.model_executor.layers.fused_moe import ( UnquantizedFusedMoEMethod, ) from vllm.model_executor.layers.fused_moe.oracle.int_wna16 import ( + WNA16MoEBackend, convert_to_wna16_moe_kernel_format, make_wna16_moe_kernel, select_wna16_moe_backend, @@ -753,13 +754,18 @@ class AutoGPTQMoEMethod(FusedMoEMethodBase): gptq_marlin_moe_quant_config, ) + # CPU fused_experts_cpu requires zero points even for symmetric quant + use_zp = ( + not self.quant_config.is_sym + or self.wna16_moe_backend == WNA16MoEBackend.CPU + ) return gptq_marlin_moe_quant_config( w1_scale=layer.w13_scales, w2_scale=layer.w2_scales, weight_bits=self.quant_config.weight_bits, group_size=self.quant_config.group_size, - w1_zp=getattr(layer, "w13_qzeros", None), - w2_zp=getattr(layer, "w2_qzeros", None), + w1_zp=getattr(layer, "w13_qzeros", None) if use_zp else None, + w2_zp=getattr(layer, "w2_qzeros", None) if use_zp else None, w1_bias=getattr(layer, "w13_bias", None), w2_bias=getattr(layer, "w2_bias", None), ) diff --git a/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_wna16_marlin.py b/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_wna16_marlin.py index a69d2a594a..8273410391 100644 --- a/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_wna16_marlin.py +++ b/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_wna16_marlin.py @@ -415,9 +415,9 @@ class CompressedTensorsWNA16MarlinMoEMethod(CompressedTensorsMoEMethod): replace_parameter(layer, "w13_weight_scale", w13_scales) replace_parameter(layer, "w2_weight_scale", w2_scales) - if w13_qzeros is not None: + # CPU fused_experts_cpu requires zero points even for symmetric quant + if not self.symmetric or self.wna16_backend == WNA16MoEBackend.CPU: replace_parameter(layer, "w13_weight_zero_point", w13_qzeros) - if w2_qzeros is not None: replace_parameter(layer, "w2_weight_zero_point", w2_qzeros) # Marlin-specific parameters (not needed for Flashinfer)