gpt_oss_converter

GPT-OSS-20B converter for AnyModel compression.

Classes

GptOssConverter

Converter for GPT-OSS models to AnyModel format.

class GptOssConverter

Bases: Converter

Converter for GPT-OSS models to AnyModel format.

GPT-OSS is a pure MoE model with 32/128 experts per layer and 4/16 active experts. All layers use MoE FFN (no standard dense FFN layers).

static create_block_configs_from_main_config(config)

Create block configs for GPT-OSS layers.

GPT-OSS uses MoE for all FFN layers with: - 32/128 local experts (num_local_experts) - 4/16 active experts per token (experts_per_token) - No dense/standard FFN layers

Parameters:

config (PreTrainedConfig)

Return type:

List[BlockConfig]

quantized = 'mxfp4'