Yuan3.0-Flash-4bit / recipe.yaml
Yuan-lab's picture
Upload 20 files
ed8a2e0 verified
raw
history blame contribute delete
392 Bytes
default_stage:
default_modifiers:
GPTQModifier:
targets: [Linear]
ignore: ['re:.*lm_head$', 're:.*qkv$', 're:.*fc1$', 're:.*fc2$', 're:.*attn.proj$',
're:.*up_proj$', 're:.*gate_proj$', 're:.*down_proj$', 're:.*router.query_key_value$']
scheme: W4A16
sequential_update: true
block_size: 128
dampening_frac: 0.01
offload_hessians: false