S1-mini-Chinese-1106-INT8 / quantization-provenance.json
WIlfLin's picture
Release actual W8A16 compressed weights with matched RoPE-corrected runtime regression
2d6d6bc verified
Raw History Blame Contribute Delete
656 Bytes
{
"format": "INT8",
"scheme": "W8A16",
"algorithm": "per-channel RTN",
"calibration_rows": 0,
"calibration_scope": "No activation calibration required for weight-only RTN",
"kv_cache": "BF16 (not quantized)",
"embeddings_lm_head_norm": "BF16",
"seconds": 7.70783578499686,
"verified_runtime_backend": "",
"rope_compatibility": "Both legacy rope_theta and rope_parameters.rope_theta explicitly 1000000 to preserve trained position encoding in Transformers4/vLLM0.17 and Transformers5",
"source_release": "WIlfLin/S1-mini-Chinese-1106-BF16",
"base_weights_sha256": "c0d2cf58468359e0a44970083f9ce38078208c291430d20a5c685c5ce6759312"
}