Motif-3-FP8-Independent

Motif-3 FP8 — Independent

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.

How it was made

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.

Local evaluation

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.

Runtime caveat

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.