Move the wiki site generator to onnx-genai-wiki

[justinchuby]

The wiki site is now published from justinchuby/onnx-genai-wiki, so the static site generator comes out of this repository.

What moves

site/ (Quartz, its plugin lockfiles, the build and validation scripts) and .github/workflows/wiki-pages.yml. All of it is now in the publishing repository, unchanged except for what the move required.

What stays

wiki/ — the notes themselves. This repository remains the source of truth for them, and they are still edited only here.

How publishing works now

onnx-genai/wiki/  ──sync──▶  onnx-genai-wiki/content/zh/  ──translate──▶  content/en/
   (source of truth)              (mirror)                                 (derived)

A scheduled workflow in the publishing repository mirrors wiki/ into content/zh/ and writes a manifest of this repository's tracked paths so links into crates/ and docs/ can still be resolved and checked without a checkout. An English edition is derived from the Chinese pages, and both are published at /zh/ and /en/ with a language switcher.

The mirror is replaced wholesale on each sync, so a page deleted here disappears there. The consequence is worth stating plainly, and it is recorded in both wiki/README.md and the wiki maintenance note: an edit made directly to the mirror is silently overwritten.

The site is at https://www.justinchuby.com/onnx-genai-wiki/.

Why a sync rather than a submodule

Measured, not assumed: this repository's .git is about 229 MB and wiki/ is about 212 KB. A submodule would clone all of it on every build to reach a thousandth of it, and could only reference the whole repository rather than one directory. Copying also gives content/zh a real commit history in the publishing repository, which is what the translation staleness check reads.

Other changes here

• .gitignore loses the site/quartz build-state entries.
• ci.yml no longer counts site/* as documentation-only. A path filter for a directory that cannot exist reads as intent to whoever changes it next.

Checked

Wikilink validation passes on wiki/ — 145 links across 25 notes — using the validator that moved with the generator. The publishing repository builds both languages from these exact pages: 4159 links across 146 pages per locale, identical page sets in the two editions, and every note carrying the lang from its frontmatter.

One defect turned up in that build and is worth knowing about, because no check of the Markdown could have found it. Obsidian reads #word as an inline tag when it follows whitespace. The Chinese text writes 、#864/#874(WDDM 回退), where the ideographic comma stops it being a tag; the natural English rendering is , #864/#874 (WDDM fallback), which produced a tag page called 864/874 in the English site and nowhere else. The publishing build now requires both locales to emit the same set of pages, since every page derives from something identical across the two trees.

[github-actions]

🔴 Benchmark Regression Detected

Comparison of criterion micro-benchmarks: PR head vs merge-base, measured on the same runner in the same job (base first → PR second).

ℹ️ Absolute times are informational only — they vary with runner load. The % change column is the reliable signal because both sides ran under identical conditions.

Status	Scenario	Base	PR	Change
🔴	block_quantized_matmul_cached_dense/mxfp4_cached_dense_repeated_call/1x1024x1024	43.23 µs	123.95 µs	+186.7%
🔴	block_quantized_moe_cached_dense/mxfp4_uncached_expert_dequant_each_call/rows=1,H=256,I=256,E=4,top_k=1	359.39 µs	572.43 µs	+59.3%
🔴	matmul/medium_generic_bf16_threads=8/32x512x512	373.38 µs	588.11 µs	+57.5%
🔴	matmul/large_generic_bf16_threads=8/32x1024x1024	1.32 ms	2.03 ms	+53.8%
🔴	qwen3_sampling_processors/top_k_top_p_full_sort_baseline	5.27 ms	8.04 ms	+52.7%
🔴	sampling_latency/top_k_per_token	48.58 µs	69.34 µs	+42.7%
🔴	block_quantized_matmul_cached_dense/mxfp4_preexpanded_dense_oncelock_like_proxy/1x1024x1024	50.61 µs	71.62 µs	+41.5%
🔴	block_quantized_moe_cached_dense/mxfp4_cached_dense_expert_repeated_call/rows=1,H=256,I=256,E=4,top_k=1	52.69 µs	72.58 µs	+37.8%
🔴	matmul/large_generic_bf16_threads=1/32x1024x1024	2.15 ms	2.83 ms	+31.7%
🔴	tokenization/decode_tokens_per_second	5.81 ms	7.63 ms	+31.2%
⚠️	gather/medium_f32_threads=1-internal/32768	4.12 µs	5.25 µs	+27.6%
⚠️	grammar_masking/llguidance_compute_mask/32	72.56 µs	91.96 µs	+26.7%
⚠️	qwen3_sampling_processors/top_p_full_sort_after_top_k_baseline	3.27 ms	4.14 ms	+26.3%
⚠️	matmul/medium_generic_f16_threads=8/32x512x512	37.07 µs	45.29 µs	+22.2%
⚠️	matmul/large_generic_f16_threads=8/32x1024x1024	91.06 µs	110.40 µs	+21.2%
⚠️	sampling_latency/greedy_per_token	3.02 µs	3.67 µs	+21.2%
⚠️	gather/small_f16_threads=1-internal/4096	494.9 ns	596.7 ns	+20.6%
⚠️	matmul/large_generic_f16_threads=1/32x1024x1024	78.46 µs	94.07 µs	+19.9%
⚠️	kv_cache/alloc_dealloc_pages	36.45 µs	43.64 µs	+19.7%
⚠️	sampling_latency/top_p_per_token	389.28 µs	461.09 µs	+18.4%
⚠️	block_quantized_matmul_cached_dense/mxfp4_uncached_dequant_each_call/1x1024x1024	567.90 µs	665.64 µs	+17.2%
⚠️	matmul/medium_generic_bf16_threads=1/32x512x512	527.42 µs	618.06 µs	+17.2%
⚠️	qwen3_sampling_processors/top_k_full_sort_baseline	1.96 ms	2.27 ms	+15.8%
⚠️	matmul/small_generic_f32_threads=1/1x256x256	39.23 µs	45.27 µs	+15.4%
✅	matmul/medium_generic_f16_threads=1/32x512x512	37.09 µs	42.56 µs	+14.8%
✅	qwen3_sampling_processors/top_k_top_p_fast	633.04 µs	719.63 µs	+13.7%
✅	gather/small_bf16_threads=1-internal/4096	485.1 ns	548.2 ns	+13.0%
✅	gather/large_bf16_threads=1-internal/131072	15.78 µs	17.82 µs	+13.0%
✅	qwen3_sampling_processors/top_p_fast_after_top_k	498.86 µs	562.33 µs	+12.7%
✅	sampling_latency/min_p_per_token	194.57 µs	217.33 µs	+11.7%
✅	matmul/small_generic_bf16_threads=8/1x256x256	41.77 µs	46.59 µs	+11.5%
✅	tokenization/encode_tokens_per_second	353.08 µs	389.76 µs	+10.4%
✅	gather/large_f32_threads=1-internal/131072	39.58 µs	43.59 µs	+10.1%
✅	gather/medium_f16_threads=1-internal/32768	2.80 µs	3.04 µs	+8.7%
✅	matmul/large_generic_f32_threads=1/32x1024x1024	10.08 ms	10.85 ms	+7.6%
✅	gather/medium_bf16_threads=1-internal/32768	2.81 µs	3.00 µs	+6.9%
✅	qwen3_sampling_processors/top_k_partial_selection	133.80 µs	142.53 µs	+6.5%
✅	matmul/medium_generic_f32_threads=8/32x512x512	1.24 ms	1.32 ms	+6.5%
✅	matmul/large_generic_f32_threads=8/32x1024x1024	4.25 ms	4.51 ms	+6.0%
✅	logit_processing/seven_processor_chain_per_step	352.68 µs	367.36 µs	+4.2%
✅	gather/small_f32_threads=1-internal/4096	752.3 ns	773.3 ns	+2.8%
✅	gather/large_f16_threads=1-internal/131072	15.42 µs	15.75 µs	+2.1%
✅	matmul/medium_generic_f32_threads=1/32x512x512	2.50 ms	2.54 ms	+1.8%
✅	matmul/small_generic_f16_threads=8/1x256x256	43.97 µs	43.98 µs	+0.0%
✅	matmul/small_generic_bf16_threads=1/1x256x256	49.51 µs	48.97 µs	-1.1%
✅	matmul/small_generic_f16_threads=1/1x256x256	36.77 µs	36.07 µs	-1.9%
✅	add/large_bf16_threads=1-internal/4194304	1.96 ms	1.85 ms	-5.9%
✅	reduce_mean/medium_f32_threads=1-internal/65536	307.36 µs	286.04 µs	-6.9%
✅	add/medium_f16_threads=1-internal/262144	141.76 µs	130.53 µs	-7.9%
✅	reduce_mean/small_f32_threads=1-internal/4096	17.81 µs	16.35 µs	-8.2%
🟢	reduce_mean/large_f32_threads=1-internal/262144	1.25 ms	1.04 ms	-16.2%
🟢	add/medium_bf16_threads=1-internal/262144	155.15 µs	124.71 µs	-19.6%
🟢	add/large_f32_threads=1-internal/4194304	1.16 ms	902.54 µs	-22.1%
🟢	matmul/small_generic_f32_threads=8/1x256x256	62.18 µs	47.02 µs	-24.4%
🟢	add/large_f16_threads=1-internal/4194304	2.76 ms	1.99 ms	-28.0%
🟢	add/small_bf16_threads=1-internal/1024	767.4 ns	549.0 ns	-28.5%
🟢	add/small_f32_threads=1-internal/1024	321.5 ns	186.8 ns	-41.9%
🟢	add/medium_f32_threads=1-internal/262144	52.19 µs	30.21 µs	-42.1%
🟢	add/small_f16_threads=1-internal/1024	785.4 ns	436.3 ns	-44.4%

Visual flags: ⚠️ ≥ 15% slower, 🔴 ≥ 30% slower — calibrated against measured runner noise (~27% worst-case on multi-threaded matmul)

Host info

CPU: Apple M1 (Virtual)
Cores: 3
OS: Darwin 25.5.0 arm64
Rust: rustc 1.97.1 (8bab26f4f 2026-07-14)
Load avg: { 4.16 3.79 6.78 }

What this cannot catch

• Regressions in code paths not covered by these benchmarks (e.g., end-to-end decode with a real model)
• Sub-threshold regressions that compound over multiple PRs
• Performance changes that only manifest under GPU execution
• Latency changes in the ORT integration path (these benchmarks exercise the native Rust kernels)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 80.12%. Comparing base (4a9f4ec) to head (b31e363).
⚠️ Report is 67 commits behind head on main.

Additional details and impacted files

@@             Coverage Diff             @@
##             main    #1488       +/-   ##
===========================================
- Coverage   82.10%   80.12%    -1.99%     
===========================================
  Files          12      376      +364     
  Lines        5471   164127   +158656     
  Branches     5471   164127   +158656     
===========================================
+ Hits         4492   131507   +127015     
- Misses        780    27793    +27013     
- Partials      199     4827     +4628     

Flag	Coverage Δ
cli-ort-linux	82.60% <ø> (?)
cli-ort-windows	82.10% <ø> (ø)
offline	80.03% <ø> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.
see 365 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.