S1-mini-Chinese-1106-FP8 / quantization-provenance.json
WIlfLin's picture
Release verified FP8 W8A16 Marlin checkpoint with Chinese regression results
2cf7deb verified
Raw History Blame Contribute Delete
1.2 kB
{
"format": "FP8",
"scheme": "FP8_WEIGHT_ONLY_PER_CHANNEL",
"calibration_rows": 16,
"calibration_scope": "Training split only, includes target continuations; no evaluation examples",
"kv_cache": "BF16 (not quantized)",
"embedding_and_lm_head": "BF16",
"modelopt_version": "0.48.0",
"seconds": 17.068106809994788,
"activation_mode": "bf16",
"source_candidate": "s1-stability-r2-fp8-dynamic-ptq",
"additional_training_steps": 0,
"method": "Lossless conversion of existing per-channel E4M3 weight bytes and scales to compressed-tensors W8A16; no new SGD or requantization",
"packed_weight_bits_verified": 196,
"scale_reshapes_verified": 196,
"layout": "compressed-tensors",
"runtime_backend": "marlin",
"full_weight_sha256": "29f28b50974183e314f296b3af7f5805b1185aeebce871386c59dc690ffadc81",
"verified_runtime_backend": "marlin",
"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"
}