Fibo-2.0-dev-OrbitQuant-W4A8 / transformer /quantization_config.json
WaveCut's picture
Fibo-2.0-dev-OrbitQuant-W4A8: OrbitQuant W4A8 transformer and text encoder
e290fd7 verified
Raw History Blame Contribute Delete
1.33 kB
{
"quant_method": "orbitquant",
"weight_bits": 4,
"activation_bits": 4,
"rotation": "rpbh",
"rotation_seed": 0,
"block_size": "paper",
"codebook": "lloyd_max",
"codebook_version": 2,
"codebook_dtype": "float32",
"row_norm_dtype": "bfloat16",
"activation_norm_dtype": "float32",
"activation_eps": 1e-10,
"weight_pack_dtype": "uint8",
"target_policy": "universal",
"adaln_policy": "int4_rtn",
"adaln_group_size": 64,
"modules_to_convert": [
"noise_refiner.*.attention.to_*",
"noise_refiner.*.feed_forward.w*",
"layers.*.attention.to_*",
"layers.*.feed_forward.w*",
"layers.*.cross_attention.to_*",
"perceiver.layers.*.cross_attention.to_*",
"perceiver.layers.*.self_attention.to_*",
"perceiver.layers.*.ffn.w*"
],
"modules_to_use_adaln": [],
"modules_to_not_convert": [],
"modules_dtype_dict": {},
"artifact_format_version": 1,
"runtime_mode": "auto_fused",
"activation_kernel_backend": "auto",
"packed_matmul_block_m": 64,
"packed_matmul_block_n": 64,
"packed_matmul_block_k": 128,
"packed_matmul_num_warps": 4,
"weight_row_tile_size": 256,
"w4a4_int8_weight_cache": false,
"lowbit_boundary_protection": "auto",
"lowbit_protected_blocks": 4,
"lowbit_protected_bits": 4,
"lowbit_interior_protection": "auto",
"fused_layout": null
}