Qwen3.8-27B-INT8-W8A8 / recipe.yaml
Avesed's picture
Qwen3.8-27B W8A8 (int8 channel weights + int8 dynamic acts) + bf16 MTP head
c264fc9 verified
Raw History Blame Contribute Delete
1.16 kB
default_stage:
default_modifiers:
QuantizationModifier:
config_groups:
group_0:
targets: [Linear]
weights:
num_bits: 8
type: int
symmetric: true
group_size: null
strategy: channel
block_structure: null
dynamic: false
actorder: null
scale_dtype: null
zp_dtype: null
observer: mse
observer_kwargs: {}
input_activations:
num_bits: 8
type: int
symmetric: true
group_size: null
strategy: token
block_structure: null
dynamic: true
actorder: null
scale_dtype: null
zp_dtype: null
observer: null
observer_kwargs: {}
output_activations: null
format: null
targets: [Linear]
ignore: [lm_head, 're:.*embed_tokens', 're:.*linear_attn[.]in_proj_a$', 're:.*linear_attn[.]in_proj_b$',
're:.*visual.*', 're:.*mtp.*']
bypass_divisibility_checks: false
requires_calibration_data: false