vllm.model_executor.layers.quantization.online.legacy_env
¶
Deprecated environment variables redirected to --quantization-config.
Functions:
-
fp8_moe_requant_args–Online quantization args equivalent to
VLLM_FP8_MOE_REQUANT[_ACT]. -
muse_moe_online_quant_args–Online quantization args equivalent to
-
reject_muse_env_with_quantization_config–VLLM_ROCM_MUSE_MOE_ONLINE_QUANTand--quantization-configare exclusive.
fp8_moe_requant_args()
¶
Online quantization args equivalent to VLLM_FP8_MOE_REQUANT[_ACT].
Only meaningful for block-FP8 checkpoints on ROCm, which the caller checks.
Source code in vllm/model_executor/layers/quantization/online/legacy_env.py
muse_moe_online_quant_args(num_hidden_layers)
¶
Online quantization args equivalent to
VLLM_ROCM_MUSE_MOE_ONLINE_QUANT=mxfp4: the first and last decoder layers
stay in BF16.
Source code in vllm/model_executor/layers/quantization/online/legacy_env.py
reject_muse_env_with_quantization_config()
¶
VLLM_ROCM_MUSE_MOE_ONLINE_QUANT and --quantization-config are exclusive.