@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / d5ab28511c
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 14, 05:30 UTC

[Bugfix] Use correct key "ignore" for config.json non-quantized layers (#25706)

Signed-off-by: Lee Nau <lnau@nvidia.com>

Lee Nau committed Sep 29, 2025 at 12:07 UTC d5ab28511c5fca0294d1b445b670e199f202193b
1 file changed +6 -2
vllm/model_executor/layers/quantization/modelopt.py
+6 -2
@@ -138,13 +138,15 @@ class ModelOptFp8Config(QuantizationConfig):
138 if not quant_method:
139 raise ValueError("Missing 'quant_algo' in quantization config")
140 kv_cache_quant_method = quant_config.get("kv_cache_quant_algo")
141 + # "exclude_modules" is the key in the legacy hf_quant_config.json
142 exclude_modules = quant_config.get("exclude_modules")
143 else:
144 # Compressed-tensors style format:
145 # {"quant_algo": "...", "quant_method": "modelopt"}
146 quant_method = config.get("quant_algo", "")
147 kv_cache_quant_method = config.get("kv_cache_quant_algo")
147 - exclude_modules = config.get("exclude_modules")
148 + # "ignore" is the key in config.json
149 + exclude_modules = config.get("ignore")
150
151 if quant_method not in QUANT_ALGOS:
152 raise ValueError(
@@ -723,6 +725,7 @@ class ModelOptNvFp4Config(QuantizationConfig):
725 raise ValueError(f"group_size must be an integer, got "
726 f"{type(group_size_raw)}") from None
727
728 + # "exclude_modules" is the key in the legacy hf_quant_config.json
729 exclude_modules = quant_config.get("exclude_modules", [])
730 if not isinstance(exclude_modules, list):
731 raise ValueError(f"exclude_modules must be a list, got "
@@ -756,7 +759,8 @@ class ModelOptNvFp4Config(QuantizationConfig):
759 raise ValueError(f"group_size must be an integer, got "
760 f"{type(group_size_raw)}") from None
761
759 - exclude_modules = config.get("exclude_modules", [])
762 + # "ignore" is the key in config.json
763 + exclude_modules = config.get("ignore", [])
764 if not isinstance(exclude_modules, list):
765 raise ValueError(f"exclude_modules must be a list, got "
766 f"{type(exclude_modules)}")