HF Chinchilla Benchmark - hf-chinchilla-local-mac-20260504T095713Z
TL;DR
Two Chinchilla models were sampled once each on the Mac CPU. The table shows load time, generation latency and output speed for those single samples. The two saved reports vary sharply for the 1B model, so neither establishes a stable performance estimate.
- Timestamp (UTC): 2026-05-04T09:57:13Z - Device: cpu - Python: 3.11.15 (main, Mar 3 2026, 15:47:15) [Clang 21.1.4 ] - Torch: 2.11.0 - Transformers: 5.7.0 - Tokenizer: mistralai/Mistral-7B-v0.1
Model Summary
| Model | Avg load (s) | Avg gen latency (s) | Avg tok/s | Prompt runs |
|---|---|---|---|---|
| mlnomad/yatnmn-softplus-d22-chinchilla-1B-pytorch | 0.978 | 6.682 | 2.394 | 1 |
| mlnomad/yatnmn-softplus-d12-chinchilla-261M-pytorch | 161.337 | 1.276 | 12.534 | 1 |
Notes
One prompt run is recorded per model, with sampling enabled. These are individual samples, not repeated-trial averages. The original results.json is not bundled on this page; the measurements shown above are the retained public summary.
Raw data: GitHub archive ·
Source: ~/Documents/MyServers/instances/slobodans-macbook-air/reports/hf-chinchilla-benchmarks/hf-chinchilla-local-mac-20260504T095713Z/