Model registry and the daily HF census (PORTING step 5)

[bong-water-water-bong]

Step 5: which HF models this engine can run, kept current daily. The two counts are reported separately and never added together.

Registry. registry/architectures.json is generated by tools/registry_build.py from the pinned sources. Nothing in it is typed in by hand:

• llama.cpp's converter (HF -> GGUF architecture);
• src/llama-arch.cpp in the upstream pin (vulkan) and in the HRX fork (hrx);
• ZINC's parseArchitecture;
• the NPU fast lane's qwen3.

At the current pins: 265 HF architectures. vulkan maps 265, hrx 246, zinc 44, npu 1.

Checked. tools/registry_check.py runs tests/serve_e2e.sh over registry/check_models.tsv and writes registry/checked.json, failures included. On Strix Halo, 17 of 20 backend-model runs pass across seven GGUF architectures. The three failures:

• HRX on Qwen3-Coder-30B-A3B and GLM-4.7-Flash: HTTP 500 "Compute error.", reproduced on a second run;
• ZINC on Qwen2.5-7B and MiniCPM5-1B: no answer.

Census. tools/census.py sweeps every HF text-generation page, and census.yml runs it daily at 05:17 UTC and opens a PR when anything changed. First sweep:

	models	share of those with an architecture
text-generation models on HF	415,414
with an architecture	332,565 (2,610 architectures)
mapped	310,221	93.28%
checked	213,456	64.18%

"Checked" counts a model when a model of its architecture passed. It does not mean every counted model was run.

Docs: docs/registry.md (in the site nav), plus PORTING row 5 and a README status line.

The workflow has not run in CI yet. Its steps were run locally: registry_build.py, the full census.py sweep, and census.py --pr-body.

🤖 Generated with Claude Code

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	➖ Did not run. This site has no agent runs available this month. Wait for the monthly reset or check your Docs7 plan.	—

Commit c298263

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 4 🔵🔵🔵🔵⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Incorrect GGUF architecture resolution The hf_to_gguf function attempts to resolve GGUF architectures from class inheritance, but it does not correctly handle cases where a class inherits from multiple base classes. This can lead to incorrect architecture mappings when the resolution logic picks an unintended base class. The function should be more robust in handling complex inheritance hierarchies. def resolve(name, seen=()): if name not in classes or name in seen: return None bases, arch, _ = classes[name] if arch: return arch for b in bases: a = resolve(b, seen + (name,)) if a: return a return None out = {} for cname, (_, _, hf) in classes.items(): arch = resolve(cname) if not arch or arch == "MMPROJ" or arch not in names: continue # vision/audio projector classes carry no text architecture for h in hf: out.setdefault(h, names[arch]) return out Incomplete error handling for API requests The sweep function retries API requests up to 6 times with exponential backoff, but it does not handle cases where the API returns a non-200 status code that is not a transient error. This could lead to silent failures or incorrect model counts if the API returns an error that should be treated as a permanent failure. for attempt in range(6): try: req = urllib.request.Request(url, headers={"User-Agent": "1bit-engine-census"}) with urllib.request.urlopen(req, timeout=120) as r: batch = json.load(r) link = r.headers.get("Link", "") break except Exception as e: # rate limits and transient errors: back off and retry if attempt == 5: raise print(f"page {pages + 1}: {e}; retrying", file=sys.stderr) time.sleep(30 * (attempt + 1)) for m in batch: Potential race condition in result updates The registry_check.py script updates registry/checked.json after each model check, but it does not use file locking or atomic operations. If multiple instances of the script run concurrently, it could lead to data corruption or loss of results. print(f"{'PASS' if ok else 'FAIL'} {backend:6} {arch:10} {model} said: {said[:60]!r}") data["results"] = sorted(results.values(), key=lambda r: (r["arch"], r["backend"], r["model"])) with open(OUT, "w") as f: # after every run, so an interrupted sweep keeps what finished json.dump(data, f, indent=1) f.write("\n")

---

⚠️ Review coverage: The following files were not included in this review because of the token budget:

• registry/census.json
• .github/workflows/census.yml
• README.md