Repository navigation
Arc Llama 1.0.0 - #91
Merged
Merged
Conversation
Includes download destinations, partial-output retention, stream deadlines, and the ruff formatting pass across src and tests.
Project and jobs move to a left rail, task entry becomes a centered composer, project settings open in a slide-over sheet, and job detail is grouped into cards with activity collapsed by default.
Add an opt-in staged sweep for MTP draft length and confidence, with repeated measurements, acceptance telemetry, controlled sampling, output comparison, and fresh confirmation before retaining a faster recipe. Preserve legacy confidence flags and restore original settings on failure or interruption. Credit DeepSeek's DeepInfer and its confidence-scheduled speculative decoding for inspiring this tuning approach: https://github.com/deepseek-ai/DeepInfer Local Qwen validation did not establish an additional verified speedup, so the original recipe was restored and no numeric gain is claimed. Validation: scripts/check.sh passed, including 1,548 Python tests and all 79 browser tests. Live validation exercised real MTP telemetry and rollback; the service is healthy and the previously loaded Muse model was restored.
…ops and config rollback safety Per-model runtimes - Models can set their own llama-server executable and environment (TOML, Model settings, `model-runtime`); launch flags, tuning, compatibility and status follow each model's runtime. Docs: docs/per-model-runtimes.md. Work - Documents work type: read, create and review Word, Excel, PowerPoint and PDF files, with page images for vision-capable models only (the UI and server refuse it for text-only models). Optional `documents` extra. Paths and file types are validated before the Office libraries are imported. - Native OpenCode and Hermes history in Work Sessions, a generation speed counter, and tool-profile/session fixes. Safety and lifecycle - POST /admin/stop and stop-all refuse with 409 while the model is serving a request (`force=true` overrides) and give llama-server 60 s after SIGTERM before SIGKILL (ARC_LLAMA_STOP_GRACE). A SIGKILL mid-request can wedge Intel Arc GPUs. Upgrades and rollbacks (#78) - Config keys this version does not recognise are kept on save, so rolling back to this version no longer drops settings written by a newer one. - Upgrade and rollback tests use configs written by the published 0.8.0, 0.9.0 and 0.9.1.2 packages; scripts/check_config_versions.py repeats the round trips against PyPI. Docs: docs/upgrading.md. Docs - Supported hardware narrowed to what has native results (Battlemage, B60); Alchemist and Lunar Lake detected and expected to work but outside the contract until reports arrive (#71). - 1.0 plan: the soak is the maintainer's daily use of the published package, and the 1.0 candidate includes the current work (0.9.2+).
arc-llama install-runtime --optimized installs the optimized llama.cpp build for Intel Arc B-series (Linux x86-64, oneAPI), from a GitHub release or a local --archive. The Ed25519 signature over SHA256SUMS (ssh-keygen -Y verify, release key shipped in the package) and the archive checksum are verified before extraction; unsupported hosts are refused with reasons and keep stock llama.cpp. Per-model settings shipped in the runtime's arc-llama-runtime.json (matched on GGUF architecture, optionally MoE only, ARC_RT_* keys only) apply to models launched with that runtime; a model's own env still wins.
Version 1.0.0, dated changelog, README for 1.0 (Arc Llama Runtime, per-model runtimes, upgrade safety, recorded Windows validation), lockfile brought up to date, and the 1.0 plan's stabilization note records how 1.0 shipped.
The check treated the oneAPI runtime as present when any of its libraries resolved, and GPU driver packages install the Level Zero loader system-wide. A SYCL llama-server then failed with "libsvml.so: cannot open shared object file" whenever Arc Llama ran without oneAPI sourced. Require SVML or the SYCL runtime in the ldconfig cache or on LD_LIBRARY_PATH; files elsewhere on disk no longer count.
LICENSE: Copyright (c) 2026 offbyonebit and arc-llama contributors (was only "arc-llama contributors"). pyproject: authors = offbyonebit, so PyPI credits the author.
- Hermes: catch asyncio.TimeoutError. On Python 3.10 asyncio.wait_for raises it and it is not the builtin TimeoutError, so startup and stop timeouts escaped their handlers. - Sessions: keep saved timestamps strictly increasing; Windows' coarse clock gave two quick saves the same time and the newest-first list fell back to file order. - Tests: close SQLite connections explicitly (the connection context manager only commits; Python 3.13+ warns about the late close), and give the probe test a full environment (Windows Python 3.10 cannot start without SYSTEMROOT).
Outside a oneAPI shell, a SYCL llama-server's --help fails to load libsvml.so and the probe fell back to assumed capabilities. Retry once with setvars applied, as launches do.
Hide install-runtime --optimized/--archive and drop the runtime from the README, changelog and docs until its own launch. The inactive integration stays (it applies only to installs of that runtime); the 1.1 wording is kept on a local branch.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Arc Llama 1.0.0. The planned 0.9.2 release is folded into 1.0.
Highlights
compaction and context budgets, per-model runtimes, safe model stops, and config upgrade/rollback safety.
Validation
scripts/check.shon the release commit: ruff, mypy, 1670 tests passed, browser suite 88/88; CI on Linux and Windows, Python 3.10/3.12/3.14.docs/release-validation-windows-2026-10-22.md.