WaveCut's picture
Requantize W2A4 with mixed-width low-bit protection
23861cb verified
Raw
History Blame Contribute Delete
863 Bytes
{
"_class_name": "OrbitQuantComponentArtifact",
"artifact_format": "orbitquant-v1",
"quant_method": "orbitquant",
"source_model_id": "black-forest-labs/FLUX.2-klein-4B",
"source_revision": "e7b7dc27f91deacad38e78976d1f2b499d76a294",
"source_license": "apache-2.0",
"component": "transformer",
"weight_name": "model.safetensors",
"quantization_config": "quantization_config.json",
"manifest": "orbitquant_manifest.json",
"codebooks": "orbitquant_codebooks.safetensors",
"rotations": "orbitquant_rotations.safetensors",
"weight_bits": 2,
"activation_bits": 4,
"codebook_version": 2,
"target_policy": "flux2",
"runtime_mode": "auto_fused",
"activation_kernel_backend": "auto",
"activation_eps": 1e-10,
"quantization_device": "cuda",
"weight_quantization_backend": "triton_cuda",
"quantization_staging_mode": "component"
}