| { |
| "producer": { |
| "name": "quantkit" |
| }, |
| "quantization": { |
| "quant_method": "modelopt", |
| "quant_algo": "NVFP4", |
| "kv_cache_quant_algo": null, |
| "group_size": 16, |
| "exclude_modules": [ |
| "audio_encoder.input_local_transformer.layers.*.input_layernorm", |
| "audio_encoder.input_local_transformer.layers.*.mlp.down_proj", |
| "audio_encoder.input_local_transformer.layers.*.mlp.gate_proj", |
| "audio_encoder.input_local_transformer.layers.*.mlp.up_proj", |
| "audio_encoder.input_local_transformer.layers.*.post_attention_layernorm", |
| "audio_encoder.input_local_transformer.layers.*.self_attn.k_proj", |
| "audio_encoder.input_local_transformer.layers.*.self_attn.o_proj", |
| "audio_encoder.input_local_transformer.layers.*.self_attn.q_proj", |
| "audio_encoder.input_local_transformer.layers.*.self_attn.v_proj", |
| "audio_encoder.input_local_transformer.norm", |
| "audio_encoder.projection.mlp.0", |
| "audio_encoder.projection.mlp.2", |
| "lm_head", |
| "model.embed_tokens", |
| "model.layers.*.input_layernorm", |
| "model.layers.*.mlp.down_proj", |
| "model.layers.*.mlp.gate", |
| "model.layers.*.mlp.gate_proj", |
| "model.layers.*.mlp.up_proj", |
| "model.layers.*.post_attention_layernorm", |
| "model.layers.*.self_attn", |
| "model.layers.*.self_attn.o_proj", |
| "model.layers.*.self_attn.qkv_proj", |
| "model.mtp.layers.*.eh_proj", |
| "model.mtp.layers.*.enorm", |
| "model.mtp.layers.*.final_layernorm", |
| "model.mtp.layers.*.hnorm", |
| "model.mtp.layers.*.input_layernorm", |
| "model.mtp.layers.*.mlp.down_proj", |
| "model.mtp.layers.*.mlp.gate_proj", |
| "model.mtp.layers.*.mlp.up_proj", |
| "model.mtp.layers.*.pre_mlp_layernorm", |
| "model.mtp.layers.*.self_attn", |
| "model.mtp.layers.*.self_attn.o_proj", |
| "model.mtp.layers.*.self_attn.qkv_proj", |
| "model.norm", |
| "speech_embeddings.0", |
| "speech_embeddings.1", |
| "speech_embeddings.10", |
| "speech_embeddings.11", |
| "speech_embeddings.12", |
| "speech_embeddings.13", |
| "speech_embeddings.14", |
| "speech_embeddings.15", |
| "speech_embeddings.16", |
| "speech_embeddings.17", |
| "speech_embeddings.18", |
| "speech_embeddings.19", |
| "speech_embeddings.2", |
| "speech_embeddings.3", |
| "speech_embeddings.4", |
| "speech_embeddings.5", |
| "speech_embeddings.6", |
| "speech_embeddings.7", |
| "speech_embeddings.8", |
| "speech_embeddings.9", |
| "visual.blocks.*.attn", |
| "visual.blocks.*.attn.proj", |
| "visual.blocks.*.attn.qkv", |
| "visual.blocks.*.mlp.down_proj", |
| "visual.blocks.*.mlp.gate_proj", |
| "visual.blocks.*.mlp.up_proj", |
| "visual.blocks.*.norm1", |
| "visual.blocks.*.norm2", |
| "visual.merger.ln_q", |
| "visual.merger.mlp.0", |
| "visual.merger.mlp.2", |
| "visual.patch_embed.proj" |
| ] |
| } |
| } |
|
|
