Publish selected TP4 reference, full speed evidence and matched FP8/NVFP4 cache KLD
43a635b verified | # Reference-stack KV-cache KLD — September 9, 2026 | |
| | Cache | Mean KL(teacher || student) | Window BCa 95% interval | | |
| | --- | ---: | --- | | |
| | Current r27 DCP4 / NVFP4 MLA KV | 0.0354562238 | [0.0295558848, 0.0434620368] | | |
| | Current r27 DCP4 / FP8 MLA KV | 0.0319451732 | [0.0268267482, 0.0387948613] | | |
| FP8 minus NVFP4: -0.0035110506; paired-window BCa 95% interval | |
| [-0.0081718772, -0.0013675941]. | |
| FP8 has lower observed KLD in 22/32 windows. | |
| Exact same 32 previously opened conditional-fit windows and BF16 teacher as the | |
| September 8 measurement: 2048 input tokens, 2047 captured predictions, exclude | |
| row zero, leaving 2046 true-decode rows/window. CPU FP64 KL over vocabulary154880; | |
| equal mean of window means; BCa20000 resamples with seed20260902. Teacher stored F32. | |
| TP4/DCP4, MTP off, maxseq1, batch4096, GMU0.97, maxlen1M; prefix caching enabled, | |
| all prefix-hit counters zero. FP8 first then NVFP4, one server preparation each. | |
| Window uncertainty does not estimate server-run variability. These are development | |
| measurements, not untouched-final qualification or independent reproduction. | |
| Selected image: verdictai/trellismx@sha256:ca6b80188dce154b91f49108b7d87792d2ba6328935afc71b44d1c0e6f6a1adf. | |
| Capture-only derivative image, seal, source identities and all64 window scores are | |
| in comparison.json. Capture seam preserves pre-mask logits and forces exact histories. | |
| NCCL8, plain one-shot cutoff131072, fused cutoff86016, shared expert threshold4096. | |
| Checkpoint and FP8 weight/activation math held fixed; KV dtype and native cache | |
| specialization vary. Capture timing is not a throughput measurement. | |
| Raw logits retired only after hashes, durable numerical scores and receipts. | |
| Audit recomputes aggregates from retained FP64 scores and verifies metadata, | |
| zero prefix hits, row masks and receipt hashes; it is not independent recapture. | |
| No measured windows excluded or rerun. Prior image/cache values remain historical. | |
| The separate production recipe uses MTP3 and24slots; its KV capacity must not be | |
| confused with capacity from this MTP-disabled correctness profile. | |