An independently implemented FP8 conversion of Motif-3, reducing stored tensor bytes by 48.9%. The checkpoint is approximately 322 GB. No calibration data, retraining or benchmark-driven quantization tuning was used.
A Python/PyTorch converter processed the original BF16 safetensors shards. Only the routed-expert gate/up and down projections were quantized: 102 tensors across 51 layers. All other weight tensors were preserved bit-for-bit.
Each expert matrix was split into 128 × 128 blocks, using FP8 E4M3FN values and FP32 power-of-two scales:
amax = max(abs(block))
scale = 1 if amax == 0 else 2**max(-126, ceil(log2(amax / 448)))
q = clamp(block.float() / scale, -448, 448).to(float8_e4m3fn)
reconstructed_block = q.float() * scale
Every output shard was checked; retained tensors and independently recomputed
quantized values/scales matched exactly. The BF16 source revision was
2ed2ed5cfabffa10fdabb2fc0d0288f8e6de893a.
| Benchmark | Result |
|---|---|
| GPQA Diamond | 84.85% (168/198) |
| IFEval, local prompt-strict | 91.87% (497/541) |
| IFBench, prompt-loose | 79.00% (237/300) |
| HumanEval, pass@1 | 96.34% (158/164) |
One answer per task, temperature 1.0, top-p 0.95. Output limits: 98,304 tokens for GPQA, 32,768 for the others; context limit 131,072. Capped/empty finals were retained in scoring. IFEval used a local deterministic verifier. Saved-answer re-scoring matched; these are not repeated generation trials or official leaderboard submissions. No matched full-suite BF16 control was run, so these results do not establish superiority over BF16 or another quantization.
Tested with a pinned Motif vLLM fork plus custom runtime integration, using four H200 NVL GPUs, TP1/PP4 and eager execution. The tested runtime also performs dynamic activation quantization. Stock vLLM/Transformers loading is not validated. Reproducing the run requires the reviewed adapter and pipeline-buffer fix; see runtime notes and the included conversion code. This checkpoint is not advertised as a drop-in stock-runtime release.
Original model and licence: Motif Technologies. This is an independent community conversion, not an official Motif Technologies release.
model-00105-of-00155.safetensors 4.0 GBmodel-00106-of-00155.safetensors 4.0 GBmodel-00107-of-00155.safetensors 4.0 GBmodel-00108-of-00155.safetensors 4.0 GBmodel-00109-of-00155.safetensors 4.0 GBmodel-00110-of-00155.safetensors 4.0 GBmodel-00111-of-00155.safetensors 4.0 GBmodel-00112-of-00155.safetensors 4.0 GBmodel-00113-of-00155.safetensors 4.0 GBmodel-00114-of-00155.safetensors 4.0 GBmodel-00115-of-00155.safetensors 4.0 GBmodel-00116-of-00155.safetensors 4.0 GBmodel-00117-of-00155.safetensors 4.0 GBmodel-00118-of-00155.safetensors 4.0 GBmodel-00119-of-00155.safetensors 4.0 GBmodel-00120-of-00155.safetensors 4.0 GBmodel-00121-of-00155.safetensors 4.0 GBmodel-00122-of-00155.safetensors 4.0 GBmodel-00123-of-00155.safetensors 4.0 GBmodel-00124-of-00155.safetensors 4.0 GBmodel-00125-of-00155.safetensors 4.0 GBmodel-00126-of-00155.safetensors 4.0 GBmodel-00127-of-00155.safetensors 4.0 GBmodel-00128-of-00155.safetensors 4.0 GBmodel-00129-of-00155.safetensors 4.0 GBmodel-00130-of-00155.safetensors 4.0 GBmodel-00131-of-00155.safetensors 4.0 GBmodel-00132-of-00155.safetensors 4.0 GBmodel-00133-of-00155.safetensors 4.0 GBmodel-00134-of-00155.safetensors 4.0 GBmodel-00135-of-00155.safetensors 4.0 GBmodel-00136-of-00155.safetensors 4.0 GBmodel-00137-of-00155.safetensors 4.0 GBmodel-00138-of-00155.safetensors 4.0 GBmodel-00139-of-00155.safetensors 4.0 GBmodel-00140-of-00155.safetensors 4.0 GBmodel-00141-of-00155.safetensors 4.0 GBmodel-00142-of-00155.safetensors 4.0 GBmodel-00143-of-00155.safetensors 4.0 GBmodel-00144-of-00155.safetensors 4.0 GBmodel-00145-of-00155.safetensors 4.0 GBmodel-00146-of-00155.safetensors 4.0 GBmodel-00147-of-00155.safetensors 4.0 GBmodel-00148-of-00155.safetensors 4.0 GBmodel-00149-of-00155.safetensors 4.0 GBmodel-00150-of-00155.safetensors 4.0 GBmodel-00151-of-00155.safetensors 4.0 GBmodel-00152-of-00155.safetensors 4.0 GBmodel-00153-of-00155.safetensors 4.0 GBmodel-00154-of-00155.safetensors 4.0 GBmodel-00155-of-00155.safetensors 4.0 GBmodel-00001-of-00155.safetensors 2.5 GBmodel-00104-of-00155.safetensors 2.3 GBmodel-00003-of-00155.safetensors 2.0 GBmodel-00005-of-00155.safetensors 2.0 GBmodel-00007-of-00155.safetensors 2.0 GBmodel-00009-of-00155.safetensors 2.0 GBmodel-00011-of-00155.safetensors 2.0 GBmodel-00013-of-00155.safetensors 2.0 GBmodel-00015-of-00155.safetensors 2.0 GBmodel-00017-of-00155.safetensors 2.0 GBmodel-00019-of-00155.safetensors 2.0 GBmodel-00021-of-00155.safetensors 2.0 GBmodel-00023-of-00155.safetensors 2.0 GBmodel-00025-of-00155.safetensors 2.0 GBmodel-00027-of-00155.safetensors 2.0 GBmodel-00029-of-00155.safetensors 2.0 GBmodel-00031-of-00155.safetensors 2.0 GBmodel-00033-of-00155.safetensors 2.0 GBmodel-00035-of-00155.safetensors 2.0 GBmodel-00037-of-00155.safetensors 2.0 GBmodel-00039-of-00155.safetensors 2.0 GBmodel-00041-of-00155.safetensors 2.0 GBmodel-00043-of-00155.safetensors 2.0 GBmodel-00045-of-00155.safetensors 2.0 GBmodel-00047-of-00155.safetensors 2.0 GBmodel-00049-of-00155.safetensors 2.0 GBmodel-00051-of-00155.safetensors 2.0 GBmodel-00053-of-00155.safetensors 2.0 GBmodel-00055-of-00155.safetensors 2.0 GBmodel-00057-of-00155.safetensors 2.0 GBmodel-00059-of-00155.safetensors 2.0 GBmodel-00061-of-00155.safetensors 2.0 GBmodel-00063-of-00155.safetensors 2.0 GBmodel-00065-of-00155.safetensors 2.0 GBmodel-00067-of-00155.safetensors 2.0 GBmodel-00069-of-00155.safetensors 2.0 GBmodel-00071-of-00155.safetensors 2.0 GBmodel-00073-of-00155.safetensors 2.0 GBmodel-00075-of-00155.safetensors 2.0 GBmodel-00077-of-00155.safetensors 2.0 GBmodel-00079-of-00155.safetensors 2.0 GBmodel-00081-of-00155.safetensors 2.0 GBmodel-00083-of-00155.safetensors 2.0 GBmodel-00085-of-00155.safetensors 2.0 GBmodel-00087-of-00155.safetensors 2.0 GBmodel-00089-of-00155.safetensors 2.0 GBmodel-00091-of-00155.safetensors 2.0 GBmodel-00093-of-00155.safetensors 2.0 GBmodel-00095-of-00155.safetensors 2.0 GBmodel-00097-of-00155.safetensors 2.0 GBmodel-00099-of-00155.safetensors 2.0 GBmodel-00101-of-00155.safetensors 2.0 GBmodel-00103-of-00155.safetensors 2.0 GBmodel-00002-of-00155.safetensors 336 MBmodel-00018-of-00155.safetensors 169 MBmodel-00020-of-00155.safetensors 169 MBmodel-00022-of-00155.safetensors 169 MBmodel-00024-of-00155.safetensors 169 MBmodel-00026-of-00155.safetensors 169 MBmodel-00028-of-00155.safetensors 169 MBmodel-00030-of-00155.safetensors 169 MBmodel-00032-of-00155.safetensors 169 MBmodel-00034-of-00155.safetensors 169 MBmodel-00036-of-00155.safetensors 169 MBmodel-00038-of-00155.safetensors 169 MBmodel-00040-of-00155.safetensors 169 MBmodel-00042-of-00155.safetensors 169 MBmodel-00044-of-00155.safetensors 169 MBmodel-00046-of-00155.safetensors 169 MBmodel-00048-of-00155.safetensors 169 MBmodel-00050-of-00155.safetensors 169 MBmodel-00052-of-00155.safetensors 169 MBmodel-00054-of-00155.safetensors 169 MBmodel-00056-of-00155.safetensors 169 MBmodel-00058-of-00155.safetensors 169 MBmodel-00060-of-00155.safetensors 169 MBmodel-00062-of-00155.safetensors 169 MBmodel-00064-of-00155.safetensors 169 MBmodel-00066-of-00155.safetensors 169 MBmodel-00068-of-00155.safetensors 169 MBmodel-00070-of-00155.safetensors 169 MBmodel-00072-of-00155.safetensors 169 MBmodel-00074-of-00155.safetensors 169 MBmodel-00076-of-00155.safetensors 169 MBmodel-00078-of-00155.safetensors 169 MBmodel-00080-of-00155.safetensors 169 MBmodel-00082-of-00155.safetensors 169 MBmodel-00084-of-00155.safetensors 169 MBmodel-00086-of-00155.safetensors 169 MBmodel-00088-of-00155.safetensors 169 MBmodel-00090-of-00155.safetensors 169 MBmodel-00092-of-00155.safetensors 169 MBmodel-00094-of-00155.safetensors 169 MBmodel-00096-of-00155.safetensors 169 MBmodel-00098-of-00155.safetensors 169 MBmodel-00100-of-00155.safetensors 169 MBmodel-00102-of-00155.safetensors 169 MBmodel-00004-of-00155.safetensors 169 MBmodel-00006-of-00155.safetensors 169 MBmodel-00008-of-00155.safetensors 169 MBmodel-00010-of-00155.safetensors 169 MBmodel-00012-of-00155.safetensors 169 MBmodel-00014-of-00155.safetensors 169 MBmodel-00016-of-00155.safetensors 169 MB有文件?提交种子。
提交种子