keep-warm pulse: hold GPU boost clock while typing so casual chat decodes at boosted rate

warmup uses decode() to precompile batched pipelines (cut first-msg TTFT)

fast first turn: warmup boosts clock + primes system-prompt KV (12s TTFT to ~0.5s) + static grounded greeting

discrete-GPU validation ?bench=discrete: bandwidth + live + spec-flip in one page

per-pass GPU trace ?bench=trace: name the non-weight overhead

spec bench: warmup + short prompt + 192-tok decode-dominated (fix prefill confound)

live decode profile ?bench=perf: boosted-clock steady tok/s vs roofline

spec-decode for BitNet: subNorm+f32-KV batched verify, ?spec + ?bench=spec

Upload index.html with huggingface_hub

Upload index.html with huggingface_hub

Upload index.html with huggingface_hub

Upload index.html with huggingface_hub

Upload index.html with huggingface_hub

Upload folder using huggingface_hub

initial commit