From 9d0fe9bac89d0bda98f977770f5ae88b386e981a Mon Sep 17 00:00:00 2001 From: Aditya Jha <118803119+ima-helikoptaaa@users.noreply.github.com> Date: Sun, 30 Aug 2026 22:32:15 +0530 Subject: [PATCH] [Bugfix][Quantization][MoE] Route weight only NVFP4 checkpoints through W4A16 (#54427) Signed-off-by: Aditya Jha <4adityajha@gmail.com> Co-authored-by: Claude --- .../layers/quantization/modelopt.py | 16 ++++++++++++++++ 1 file changed, 16 insertions(+) diff --git a/vllm/model_executor/layers/quantization/modelopt.py b/vllm/model_executor/layers/quantization/modelopt.py index 241a7c8ada..575c122d6b 100644 --- a/vllm/model_executor/layers/quantization/modelopt.py +++ b/vllm/model_executor/layers/quantization/modelopt.py @@ -1081,6 +1081,22 @@ class ModelOptNvFp4Config(ModelOptQuantConfigBase): ) -> "ModelOptNvFp4Config": is_checkpoint_nvfp4_serialized = "NVFP4" in quant_method + # A checkpoint can declare quant_algo "NVFP4" yet quantize weights + # only (input_activations is null in every config group). The W4A4 + # MoE path then folds an uninitialized input_scale that was never + # loaded and can zero every expert, so route it through W4A16. + if quant_method == "NVFP4": + config_groups = original_config.get("config_groups") + if ( + isinstance(config_groups, dict) + and config_groups + and all( + isinstance(group, dict) and group.get("input_activations") is None + for group in config_groups.values() + ) + ): + quant_method = "W4A16_NVFP4" + if group_size is None: group_size = 16 # Default value