inferrsi.js

'use strict';
(() => {
 const zh = Object.fromEntries([...document.querySelectorAll('[data-rsi-i18n]')].map(el=>[el.dataset.rsiI18n,el.textContent]));
 Object.assign(zh,{metric:'实测音频吞吐 · 相对基线倍数',oneRequest:'单并发',sixtyFour:'64 并发',benchmarkSource:'查看固定版本数据与测试条件 ↗',ttsBench:'作者报告 · 单张 910B1 / BF16 · 3 轮均值，每轮 128 请求；共享主机顺序测试，基线含兼容性修补。',dramaBench:'作者报告 · 单张 B300 · 3 轮中位数；vLLM-Omni C64 仅完成 2 轮，第 3 轮超时，≈1.00× 表示基本持平。配置与 codec 精度不同。'});
 const en = {metric:'Measured audio throughput · vs baseline',oneRequest:'concurrency',sixtyFour:'concurrent requests',benchmarkSource:'Pinned results & benchmark conditions ↗',ttsBench:'Author-reported · 1× 910B1 / BF16 · 3-run means, 128 requests/run. Shared-host sequential trials; baselines include compatibility patches.',dramaBench:'Author-reported · 1× B300 · 3-run medians; vLLM-Omni C64 completed 2 runs, with a third timeout. ≈1.00× means parity. Configuration and codec precision differ.', 
 navRecipe:'Recipes ↗',kicker:'FROM BOTTLENECKS TO VERIFIABLE PROGRESS',title:'Make every lesson reusable.',intro:'A CLI workbench for inference optimization with coding agents. Connect profiling, repeated experiments, quality gates and reusable knowledge.',profile:'01 / Profile',compare:'02 / Experiment & verify',learn:'03 / Learn & reuse',repo:'Explore InferRSI',template:'Engine delivery template ↗',scope:'You drive profiling and code changes; InferRSI runs and evaluates the benchmark and validation commands you provide.',examples:'Dedicated engines · In practice',ttsStatus:'Scoped campaign passed · Author-reported',ttsDesc:'Independent NPU speech inference for Qwen3-TTS-12Hz-1.7B-CustomVoice, with continuous batching, NPU graphs and streaming audio.',ttsLimit:'One Ascend 910B1, a fixed model and workload. The repository records three performance trials and audio-quality acceptance; results do not generalize across hardware.',dramaStatus:'Experimental · Original target unmet',dramaDesc:'A dedicated single-B300 streaming TTS engine combining native FP8 MoE, persistent GPU state, CUDA Graphs and incremental audio decoding.',dramaLimit:'C1–C64 performance, quality and failed-run records are retained. The original all-concurrency throughput target remains unmet and optimization is paused; configurations and repeat counts differ.',viewExample:'Explore example & reproduction ↗',recipesTitle:'Find a starting point for your model and hardware.',recipesDesc:'Visit the dedicated Recipes catalog for deployment recipes, supported environments and linked evidence. Open to everyone, with no sign-in required.'
 };
 Object.assign(en,{"thirtyTwo":"concurrent requests","minicpmStatus":"H200 configuration · Unverified","minicpmMetric":"L20X source measurements · vs baseline","minicpmDesc":"Single-GPU streaming speech targeting H200: Thinker and Talker share paged KV and CUDA graphs; a separate Token2Wav process decodes under CUDA MPS, streaming 24 kHz audio.","minicpmBench":"Source measurements · L20X (143 GB), not H200 results · track_s_fast_r11 · 5-run means, 320/640 English Seed-TTS requests per run at C1/C32. Throughput: 397.43/2437.10 frames/s; TTFA p50: 126.2/654.6 ms.","minicpmLimit":"The 2× target across all metrics and concurrency levels is unmet. The r11 HiFT error exceeds its bound on 1/190 chunks; speech quality gates are pending. About 95 GB used on the source host; H200 execution remains unverified. The default track_s_fast differs from this measured profile."});
 Object.assign(en,{"h3Status": "Video / audio · Author-reported", "h3Desc": "Metal-native FastH3 / MiniMax-H3 video and audio inference on Apple GPUs. Block-sparse attention, int8 TensorOps and resident DiT / decoder state, with InferRSI-style frozen baselines and alternating candidate trials.", "h3Metric": "Warm latency per clip · vs h3.c at 20 steps", "h3Twenty": "Original model · 20 steps", "h3Four": "Distilled model · 4 steps", "h3Warm": "Reused in-process state", "h3Different": "Different checkpoint / steps", "h3Bench": "Author-reported · M5 Max 40-core GPU / 128 GB / macOS 27; 768×448, 49 frames, seed 0, one engine at a time. Cold: 54.29 / 48.23 s. Warm: clips 2–4 mean / two A/B/B/A run means, respectively.", "h3Limit": "FastH3 uses distilled weights and sparse attention: its ratio is not an equal-model, equal-quality engine-only speedup. Not independently rerun; unattended model-backed search remains on the roadmap."});
 function translate(){const copy=document.documentElement.lang.startsWith('zh')?zh:en;document.querySelectorAll('[data-rsi-i18n]').forEach(el=>{el.textContent=copy[el.dataset.rsiI18n]||zh[el.dataset.rsiI18n];});}
 new MutationObserver(translate).observe(document.documentElement,{attributes:true,attributeFilter:['lang']});translate();
})();