Verified accelerator results

The machine-readable provenance index pins the current record digests, source revision, identity evidence, and exact values used by public summaries. Tests validate it against the committed files.

Current acceptance record (12bfd8522f9a4dff46f05157108eb63159240882)

Kaggle kernel version 12 checked out the exact revision above and passed every stage on one TPU v5 Lite host with eight devices. The compact evidence is archived as five immutable records:

The suite passed all eleven stages, including the complete test suite and tokenizer → pretrain → SFT → preference → evaluation → inference pipeline, checkpointing, attention/speculative benchmarks, and device scaling. This is one physical host with eight devices; it is not multi-host evidence. The 589,902-parameter case remains an overhead regression test. The 29,360,454-parameter representative case reached 1.170M tokens/sec at eight devices under fixed global work and 1.158M tokens/sec under fixed per-device work. Against a predeclared 0.50 floor, eight-device strong-scaling efficiency was 0.5138 and weak-scaling efficiency was 0.7412; both passed. Timing dispersion, compile time, checkpoint latency, MFU estimate, and compiled-memory semantics are retained in the records. These are device-scaling measurements, not multi-host claims.

Previous acceptance record (7df9fe85817acf27fa61f5feb46e7f2a0774a3b1)

The prior five-file kernel-version-10 record remains available:

Earlier acceptance record (97ba1333bf46b51f85ca3d9099c8c2717438ce91)

The previous three-file record remains available for historical comparison:

Earlier characterization records

The raw structured record for the successful acceptance bundle is benchmarks/results/kaggle-tpu-v5e-8-cde72cd.json. It identifies the exact source and TinyStories revisions, TPU kind/count, stage-loss series, evaluation loss, compile times, attention throughput, compiled memory estimates, speculative correctness/speedup, and limitations.

At cde72cd the bundled Kaggle TPU v5e-8 run passed every check. The pinned TinyStories path completed tokenizer → pretrain → SFT → preference training → evaluation → inference and published its complete Orbax/tokenizer/manifest bundle. Splash attention measured 2.81M, 4.00M, 4.08M, and 2.46M tokens/sec at 1K, 2K, 4K, and 8K respectively. Greedy speculative output matched cached generation exactly and measured 1.77× speedup for that aligned synthetic upper-bound pair. That revision still shared initialization internally; the result is mechanism evidence, not a claim about an independently trained draft.

The independent GPU bundle in benchmarks/results/kaggle-gpu-p100-d33726e.json also passed: 236 tests on a Tesla P100, plus the complete pinned TinyStories pipeline, pretraining smoke test, XLA attention benchmark, and speculative correctness check. It measured 2.78M and 2.50M attention tokens/sec at 1K and 2K. Its randomly initialized draft model accepted no proposals, correctly making speculative decoding slower; the record preserves that negative result rather than presenting it as a speedup.

These numbers are acceptance and systems measurements, not quality or scaling claims. In particular, compiled XLA memory is not a whole-device HBM watermark, and no nanochat/MaxText result is presented until the controlled fields in benchmarks/baselines/ can be matched on the same hardware.

Matched trainer comparison

Before spending GPU quota, Kaggle CPU kernel version 4 checked out all three pinned sources, installed the isolated adapter dependencies, constructed each native model, and verified trainable counts of 589,902 (flaxchat), 589,902 (nanochat), and 606,848 (MaxText). The machine-readable preflight record passed the 600,000 ±5% gate and a native nanochat forward/backward AdamW update without enabling a GPU or TPU. Comparative performance is still withheld: bundled P100 versions 2 and 3 each completed flaxchat and MaxText but exposed successive nanochat runtime-compatibility and adapter-dtype failures. Both negative records are retained rather than selectively publishing the two successful framework measurements. The dtype correction is CPU-validated; another GPU measurement should only be run with an explicit quota decision.