[bong-water-water-bong]

Brings 1bit.gg up to date with this week's work.

• Blog post 2026-09-26-zyphra-family-old-architectures-experts-from-the-drive:

  • Zyphra's whole family: ZAYA1-74B, the legacy base models, ZAYA1-VL-8B (100/101 against Zyphra's FP32 code), Zamba, Zamba2 and BlackMamba;
  • the four older architectures;
  • --moe-slots streaming;
  • the deterministic HRX decode default;
  • ONNX on the GPU, the matrix-unit reference, and the issue/RFC/CI work.

  Every number comes from a committed doc or a merged PR, with links.

• docs/vulkan.md: a section on OPT, GPT-Neo, CodeGen and GPT-J from our llama.cpp (Pin llama.cpp 83e1c41: OPT, CodeGen, GPT-Neo, GPT-J and 48 sibling HF names #138). These had no docs; the FP32 agreement table comes from the PR.

• docs/registry.md: the counts said 265 HF architectures (vulkan 265, hrx 246, zinc 44). At the current pins they are 323 / 323 / 300 / 64. The line now also says what "mapped" means, per CONTRIBUTING rule 3.

• registry/architectures.json: regenerated for ZINC 29bc350, bumped in Bump ZINC: zolotukhin/zinc 29bc350ac4cb #132 without a regeneration. The counts are unchanged; the gap check passes.

tools/site.py builds 34 pages. The post's links and the four anchors it uses resolve.

🤖 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 d240dec

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit d240dec)

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 138 - Partially compliant Compliant requirements: Move third_party/llama.cpp from 706ad1a to 83e1c41 Add support for OPT, CodeGen, GPT-Neo, GPT-J architectures Fix issues in the new ports Update serve to cover the new architectures in fork_only_arch Update registry_build to count architectures only when llama_model_mapping builds a model for it Update registry/architectures.json to reflect 321 HF architectures mapped Non-compliant requirements: Drop BertForCausalLM/BertLMHeadModel aliases (not visible in diff) Update docs/vulkan.md with new architecture support (not fully visible in diff) Requires further human verification: Verification that BertForCausalLM/BertLMHeadModel aliases were actually dropped Verification that docs/vulkan.md fully reflects the new architecture support 132 - Partially compliant Compliant requirements: Move third_party/zinc from 3a35e76d64eb to 29bc350ac4cb Update ZINC with upstream commits Non-compliant requirements: Verify on Strix Halo with specific command (not visible in diff) Requires further human verification: Verification that Strix Halo command was run and results confirmed

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Documentation Inconsistency The blog post mentions that the model registry maps 323 Hugging Face architectures to a backend, but the documentation in docs/registry.md still refers to the old count of 265 architectures. This inconsistency should be resolved to ensure clarity for users. Outdated Count Reference The documentation still references the old count of 265 architectures instead of the updated count of 323 architectures. This should be updated to reflect the current state. At the current pins (llama.cpp `8dd75eb`, llama.cpp-vulkan `546800c`, ZINC `29bc350`): 323 HF architectures mapped, vulkan 323, hrx 300, zinc 64, npu 1. Mapped means a backend's code accepts the architecture; the census below reports how many were checked.

[github-actions]

Persistent review updated to latest commit d240dec