quantized_weight

Non-mutating export helpers for quantized checkpoint weights.

Functions

build_hf_quantization_config	Build the canonical ModelOpt HF config from canonical module states.
capture_quantized_weight_export_state	Capture export state without changing the module or its quantizers.
export_quantized_weight	Pack one canonical weight without mutating it or its quantizers.
quantized_weight_export_states_compatible	Return whether two states share one deployment format and tensor layout.
quantized_weight_export_states_equal	Return whether two opaque export states describe the same checkpoint state.
replicate_quantized_weight_export_state	Authorize scalar state replication across a fused projection split.
synchronize_quantized_weight_export_state	Max-reduce scalar state for ranks that shard one logical weight.

build_hf_quantization_config(layer_states)

Build the canonical ModelOpt HF config from canonical module states.

Parameters:

layer_states (Mapping[str, object | None] | Iterable[tuple[str, object | None]])

Return type:

dict

capture_quantized_weight_export_state(module, weight_name='weight', *, weight_quantizer=None, input_quantizer=None)

Capture export state without changing the module or its quantizers.

Parameters:

• module (Module)

• weight_name (str)

• weight_quantizer (object | None)

• input_quantizer (object | None)

Return type:

_QuantizedWeightExportState

export_quantized_weight(weight, state, *, dtype=None)

Pack one canonical weight without mutating it or its quantizers.

Parameters:

• weight (Tensor)

• state (object)

• dtype (dtype | None)

Return type:

_QuantizedWeightExport

quantized_weight_export_states_compatible(left, right)

Return whether two states share one deployment format and tensor layout.

Parameters:

• left (object)

• right (object)

Return type:

bool

quantized_weight_export_states_equal(left, right)

Return whether two opaque export states describe the same checkpoint state.

Parameters:

• left (object)

• right (object)

Return type:

bool

replicate_quantized_weight_export_state(state, projection_count)

Authorize scalar state replication across a fused projection split.

Parameters:

• state (object)

• projection_count (int)

Return type:

tuple[object, …]

synchronize_quantized_weight_export_state(state, *, group=None, device=None)

Max-reduce scalar state for ranks that shard one logical weight.

Parameters:

• state (object)

• device (device | str | None)

Return type:

object