Skip to content

Weekly digest: every pin and this week's new models in changes.json - #122

Merged
bong-water-water-bong merged 1 commit into
mainfrom
weekly-pins-models
Sep 26, 2026
Merged

bong-water-water-bong merged 1 commit into
mainfrom
weekly-pins-models

Conversation

@bong-water-water-bong

Copy link
Copy Markdown
Collaborator

tools/weekly_changes.py now also writes, into changes.json and the release notes:

  • pins: every third_party/ upstream at the release: the pinned commit and its date, plus the upstream's latest release where it has one. This replaces the separate pinned Updates post on Discord, so the weekly #announcements post is the single digest.
  • models: what was added since the last release: repos under the 1bit-MONSTER Hugging Face org, new entries in Lemonade's server_models.json (between the two Lemonade pins), and architectures in registry/architectures.json that are new or gained a backend.

Tested on Strix Halo against main. HF picked up the 4 repos published this week. From 2f1c45f the diff finds BlackMamba, Zamba2 and Zamba v1 on Vulkan. A dry run of the Discord post renders the pins list correctly (4 messages, each under 2000 chars).

🤖 Generated with Claude Code

The weekly Discord post is now the one digest: the pinned-upstreams list that used to be its own
post, and new models (our Hugging Face repos, Lemonade catalog entries, registry architectures).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@context7

context7 Bot commented Sep 26, 2026

Copy link
Copy Markdown

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 8d2f123

@github-actions

Copy link
Copy Markdown

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

Potential Runtime Error in `pins` Function

The pins function attempts to fetch commit dates and release information for each pinned upstream. If the GitHub API call for commit data fails (e.g., due to rate limiting or network issues), the code catches the subprocess.CalledProcessError and sets the date to None. However, if the gh function raises a different exception or returns an unexpected structure, this could lead to a runtime error when trying to access r["commit"]["committer"]["date"]. This could cause the entire script to fail during execution.

try:
    p["date"] = gh(f"repos/{repo}/commits/{sha}")["commit"]["committer"]["date"][:10]
except subprocess.CalledProcessError:
    p["date"] = None
Incomplete Error Handling in `models` Function

The models function fetches Hugging Face models using urllib.request. If the request fails due to network issues or timeouts, the code catches OSError and silently ignores the error. This could lead to missing model data in the generated release notes, which may be important for users. A more robust approach would be to log the error or provide a fallback mechanism.

except OSError:

@bong-water-water-bong
bong-water-water-bong merged commit 2e94e82 into main Sep 26, 2026
5 checks passed
@bong-water-water-bong
bong-water-water-bong deleted the weekly-pins-models branch September 26, 2026 09:57
bong-water-water-bong added a commit that referenced this pull request Oct 21, 2026
…4 SwiGLU prefill opt-in (27B KLD base re-recorded) (#493)

llama.cpp 96f81977 (fork #120, #122, #121 on 5dcf6623):
- #120: prompt flash attention above 2048 query rows as query chunks (27B -ub 4096: 6.5 -> 588 tok/s).
- #122: ROPE above 2048 tokens as token chunks (Qwen3-4B now runs -ub 4096; KLD 0.000002).
- #121: the int4-activation SwiGLU prefill (symmetric_i4_q8_plane) is opt-in,
  GGML_HRX_SYMMETRIC_I4_PREFILL=1. 27B prefill speed -0.0..-0.7% (interleaved x3).

scripts/hw-gate.sh --bin on a private build (this llama.cpp + hrx-system 82804197 + gpu-kernels
08bb2241, strixhalo performance mode), check mode: PASS on every row (ops 667/667, kat, needle,
placement, kld). The 27B kld row moved, as expected from #121: mean KLD 0.000000 -> 0.002222
(p99 0.01868, same top 97.65%, limit 0.01) against the base recorded with the int4 path on.
The 27B base is re-recorded on that same build (scripts/hw-gate.sh --phases kld --models
qwen3.8-27b --record): manifest sha256 d2f9a8af -> a31cb56c; the previous base file is kept on
strixhalo as ~/.cache/hw-gate/kld/qwen3.8-27b.kld.pre-96f81977. Measured directly: the new
default is the exact path - 27B -ub 2048 against the -ub 4096 logits 0.001949 -> 0.000003,
PPL (c4096) 6.7183 -> 6.7097, PPL (c512, the gate args) 6.6835 -> 6.6381.

Co-authored-by: bong-water-water-bong <bong-water-water-bong@1bit.gg>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant