velaia/ncm-reproduction

Next Concept Models: reproduction results

Raw results from reproducing Next Concept Models with the released code (HazyResearch/ncm) and checkpoints (hazyresearch/ncm-checkpoints). The runs were on an Apple M1 Max, a rented RTX 4090 and a rented H100.

In short:

• The quality numbers reproduce. Every task with a released checkpoint lands within run-to-run noise of the paper.
• The KV cache gives identical outputs once fixed. With the fix (#5), cached and uncached generation give identical outputs, and the cached path is 2.1–5.1× faster on an H100.
• The paper's wall-clock times are only reached with reduced precision. With the released code in fp32, XSum takes 44.7 s for the paper's timing protocol; Figure 1 shows about 3.4 s. Only bf16 with the cache and CUDA graphs comes close (3.45 s). See issue #9.

Code: velaia/ncm, branch mac-mps. Changes sent upstream:

• #4: free generate()'s tensors on return (memory leak)
• #5: fix the KV-cached generate() path
• #6: run evaluation on Apple silicon (MPS) and CPU

Quality

Greedy decoding with the eval configs shipped with the checkpoints. Mac runs used the uncached path. The 4090 and H100 runs used the KV cache.

Task	Metric	M1 Max	RTX 4090	H100	Paper
WMT14 De→En, 3,003 sentences	BLEU / lowercased	21.72 / 22.61	21.73 / 22.62	21.72 / 22.61	22.6
WMT14 + CFG 1.0	BLEU / lowercased	22.95 / 23.84	22.96 / 23.85	22.96 / 23.84	23.9
XSum, 11,334 articles	ROUGE-1 / 2 / L		34.92 / 12.71 / 27.93	34.92 / 12.71 / 27.93	35.0 / 12.8 / 28.0
Sudoku easy / medium / hard, 2,000 each	% solved	99.4 / 96.4 / 86.3	99.40 / 96.45 / 86.25	99.4 / 96.4 / 86.3	99.5 / 96.5 / 86.1
TinyStories, 5 seeds × 256	gen. PPL (entropy)	15.98 ± 0.28 (3.97)			15.41 (3.98)
LM1B small, 5 seeds × 256	gen. PPL (entropy)	92.60 ± 1.41 (4.31)			93.05 (4.31)
LM1B medium	gen. PPL (entropy)	88.96 ± 1.68 (4.34)			plot only
LM1B large	gen. PPL (entropy)	81.41 ± 1.95 (4.34)			plot only

The paper's WMT14 numbers match lowercased BLEU. The Mac and the 4090 differ by about 0.01 BLEU, and by one Sudoku puzzle on medium and on hard. These differences come from floating-point differences between the GPUs. On each machine, cached and uncached outputs are identical.

The repo has no OpenWebText checkpoint, MBR decoding, baselines or training code, so those parts of the paper are out of scope.

Speed

scripts/bench_kv.py runs the same inputs through each mode, with the uncached fp32 run (eager) first. "Identical" is the share of outputs that match eager token for token.

Timing follows Section 4.6 of the paper: 1,024 inputs at batch 16. It covers generation plus VQ decoding to token ids, after one warmup batch, with the device synchronised. It excludes model loading, detokenisation and scoring.

H100 SXM 80 GB, at b744792 (h100/reuse/):

Task (NFE)	eager	kv	kv + CUDA graphs	kv + bf16 + CUDA graphs
WMT14 (11)	13.73 s	6.52 s, 100%	5.15 s, 100%	2.20 s, 27.6%, BLEU 20.43 vs 20.59
WMT14 + CFG 1.0 (22)	24.64 s	10.47 s, 100%	8.54 s, 100%	2.87 s, 11.2%, BLEU 21.96 vs 22.09
XSum (9)	44.71 s	10.04 s, 100%	8.85 s, 100%	3.45 s, 53.5%, ROUGE-L 27.74 vs 27.77
XSum + CFG 0.5 (18)	86.42 s	16.98 s, 100%	15.44 s, 100%	5.55 s, 50.9%, ROUGE-L 27.88 vs 27.95
Sudoku ×3, 2,000 at batch 128 (12)	8.48–8.52 s	3.60–3.63 s, 100%

NFE counts generator forward passes per generate() call. We measured it with a hook on the first transformer block. The paper's Table 2 lists 8 for XSum.

Batch size barely matters without the cache: XSum takes 44.16 s at batch 64. Figure 1 of the paper shows about 3.4 s for XSum and about 4.9 s with CFG (read off the log axis).

benchmarks.csv has every bench row, with its machine, run and commit, including:

• TF32, bf16 and torch.compile modes in the sweep runs
• the 4090 and Mac numbers

Memory

Before #4, generate() kept its tensors alive after returning, until Python's cyclic garbage collector ran. The cause was a reference cycle from a getattr on a missing OmegaConf key.

• Uncached path, RTX 4090, 1,024 sequences: peak memory dropped from 5.82 GB to 2.52 GB on XSum at batch 64 (rtx4090/pre-fix/summary.txt vs rtx4090/bench/).
• KV-cached path: before the fix and the cache memory trimming in eb6da99, XSum at batch 16 peaked at 11.54 GB (2.35 GB after), and batch 64 ran out of memory on the 24 GB card.

Layout

Path	What	Commit
h100/bench/, h100/eval/, h100/outputs/	bench JSONs, full-test-set scores and generated outputs (cached path)	df2e848
h100/sweep/	precision / compile / CUDA-graph sweep	df2e848
h100/reuse/	re-timing after the cache reuses the level-0 pass's prefix K/V	b744792
rtx4090/bench/, rtx4090/eval/, rtx4090/outputs/	as for the H100	eb6da99
rtx4090/pre-fix/	uncached path before the memory fix	2b8c50f
rtx4090/sweep/	precision / compile / CUDA-graph sweep	c30081e
mac/bench/	eager vs kv, stateful cache with the mask fix	00184d7 (Sudoku: 2b8c50f)
mac/bench-pr-branch/, mac/eval/wmt14-kv-pr-branch.txt	the #5 branch on top of the #6 commit	60c947b + 1281b68
mac/eval/, mac/outputs/	full evaluations, uncached path	57233ec (WMT14), 00184d7 (TinyStories, LM1B)
*/nvidia-smi.txt, */torch.txt, */commit.txt	GPU, driver, PyTorch version, commit at the start of the run

Commits are on velaia/ncm, branch mac-mps, except 60c947b (branch kv-cache) and 1281b68 (branch mps-cpu). df2e848 has the same model code as c30081e; only the cloud runner scripts differ.

outputs/*.jsonl hold the generated text only, one object per input: idx, the index into the test set, and output; for the unconditional tasks, also seed. References and source texts are not included; they are in the original datasets.

Machines and caveats

• M1 Max, 64 GB, macOS 26.5, PyTorch 2.14.1, MPS. Needs the changes in #6: float32 instead of float64 for the frequency penalty, and a replacement for area interpolation.
• RTX 4090 24 GB (vast.ai), PyTorch 2.14.1+cu126.
• H100 SXM 80 GB HBM3 (vast.ai), PyTorch 2.14.1+cu126.
  • This host was power-capped at 500 W instead of the stock 700 W (h100/nvidia-smi.txt).
  • Absolute H100 times are therefore likely slower than on a stock H100 like the paper's.
• cuDNN TF32 is off in bench_kv.py. With it on, 1 of 1,024 guided WMT14 translations differed between cached and uncached in one early 4090 run.
• The 4090 and Mac speed numbers predate b744792, so their KV-cache columns are slightly pessimistic. So is everything in h100/bench/ and h100/sweep/.