vllm.model_executor.layers.quantization.online.requantization
¶
Classes:
-
OnlineRequantizationMixin–Shared lifecycle for converting serialized weights before quantization.
Functions:
-
skip_checkpoint_backend_selection–Skip backend selection for checkpoint quantization methods in this scope.
OnlineRequantizationMixin
¶
Shared lifecycle for converting serialized weights before quantization.
Methods:
-
create_requantization_source_weights–Delegate weight creation to the source and track its parameters.
-
release_requantization_source_weights–Release serialized parameters after successful requantization.
-
resolve_activation_quant_key–Final activation key for a requested (or default) key.
-
set_requantization_source–Configure serialized-weight conversion before online quantization.
Source code in vllm/model_executor/layers/quantization/online/requantization.py
create_requantization_source_weights(layer, *weight_args, **extra_weight_attrs)
¶
Delegate weight creation to the source and track its parameters.
Source code in vllm/model_executor/layers/quantization/online/requantization.py
release_requantization_source_weights(layer)
¶
Release serialized parameters after successful requantization.
Source code in vllm/model_executor/layers/quantization/online/requantization.py
resolve_activation_quant_key(activation_quant_key, explicit)
classmethod
¶
Final activation key for a requested (or default) key.
Source code in vllm/model_executor/layers/quantization/online/requantization.py
set_requantization_source(source_method)
¶
Configure serialized-weight conversion before online quantization.
Source code in vllm/model_executor/layers/quantization/online/requantization.py
skip_checkpoint_backend_selection()
¶
Skip backend selection for checkpoint quantization methods in this scope.
Covers checkpoint methods used as requantization sources (their weights are only dequantized) and unquantized checkpoint methods discarded when an online target replaces them, so oracles need not probe a compute backend.