Hy4-preview-SF-FP8
SiliconFlow crossingbits block-FP8 (E4M3, block 128, FP32 scales) of Tencent Hy4-preview.
License: Apache-2.0 (inherited from upstream).
Quant Diff
Official Hy4-preview-FP8 |
This checkpoint | |
|---|---|---|
| Weight format | ModelOpt MXFP8 (UE8M0, group 32) | crossingbits block-FP8 E4M3, block 128, fp8_scale_w FP32 |
| Activations | MXFP8 dynamic | per-token FP8 (runtime fp8_scale_x) |
| BF16 keep | lm_head, embed_tokens, mlp.gate, linear_gate, indexer.weights_proj, eh_proj, iHC hc_fn |
same + MLA absorb kv_b/k_b/v_b |
| Experts | packed 3-D gate_up_proj/down_proj |
unfused per-expert gate_proj/up_proj/down_proj |
| MTP | 1 layer, no iHC | quantized with main (use_mtp: true); eh_proj BF16 |
Calibration: C4, no rotate/smooth/clip.
Reported during PTQ: mean snr_err 4.61e-4, KLD 9.73e-2, rejection 0.361, dead 0.0%.
Files
HuggingFace-style shards: model-common-*.safetensors + model-experts-*.safetensors, quantize_config.json.