Tags: Hal0ai/hal0
Tags
release: v0.9.6.1 (#1191) Patch release on the 0.9.6 line. Bumps pyproject.toml + manifest.json to 0.9.6.1 and promotes the accumulated [Unreleased] notes into a dated [0.9.6.1] section. Fixes bundled: FLM slot warm→ready sentinel probes the assigned model (#1171), graceful 409 handling in hal0 setup (#1158), unified --warn color token (#1156/#1155), and a dash-JSX no-undef lint guard (#1170). Claude-Session: https://claude.ai/code/session_01DDf84s6iKC4CuGUnyeH4kJ Co-authored-by: Claude <[email protected]>
v0.9.5.2 — close React #310 on Models → Downloads Hotfix: useStateM was called inside jobs.map() in the Downloads pane, causing React invariant 310 ("Rendered more hooks than during the previous render") the moment the downloads list changed length. Extracted DownloadRow component to own the per-row hook. Pinned by a new AST-based regression test under ui/src/dash/__tests__/.
release: v0.9.5 (vulkan-minicpm5 runner + slot.image override)
Highlights:
- New canonical ROCmFPX runner: ghcr.io/hal0ai/hal0-rocmfpx:vulkan-minicpm5
(Hal0ai/Hal0_ROCmFPX@5b395660 + upstream PR #23384 cherry-pick for minicpm5
pre-tokenizer support). Single image serves ROCm/HIP and Vulkan backends.
Unblocks loading the BRAINTRAIN-1B GGUFs on Strix Halo.
- Slot-level image override: slot.image (top-level string in the slot TOML)
now overrides profile.image. Resolution order:
slot.image -> profile.image -> DEFAULT_ROCMFPX_IMAGE.
Future image bumps will be a code-only release (one constant in
schema.py); operators no longer have to chase profile clones.
Schema:
- Add DEFAULT_ROCMFPX_IMAGE = "ghcr.io/hal0ai/hal0-rocmfpx:vulkan-minicpm5"
constant in src/hal0/config/schema.py.
- The three ROCmFPX runner seed profiles (rocmfpx-rocm, vkfpx-moe,
vkfpx-dense) now reference vulkan-minicpm5.
- Add _resolve_image_ref(slot_cfg, profile) helper in
src/hal0/providers/container.py implementing the priority chain.
- _profile_image_and_flags now accepts slot_cfg and delegates image
resolution to _resolve_image_ref.
Tests:
- 11 new tests in tests/providers/test_image_resolution.py covering:
slot-level override (top-level + nested [slot] section), profile
fallback, DEFAULT_ROCMFPX_IMAGE fallback, image_gen dict collision
guard, empty-string override, and SEED_PROFILES image pinning.
Breaking:
- SEED_PROFILES for ROCmFPX runner changed from :server -> :vulkan-minicpm5.
- profile.image deprecated; will be removed from SEED_PROFILES in 0.9.6.
Refs:
- handoff: /tmp/handoff-b2-plan-2026-07-08.md (minicpm5 B2 research)
- handoff: /tmp/handoff-braintrain-nano-slot-2026-07-08.md (BRAINTRAIN context)
- upstream PR: ggml-org/llama.cpp#23384
v0.9.3 — The Guided Setup hal0 setup gains a review-gated TUI, headless answer files, hardware preflight, and a pick-free first-run workflow. The installer no longer ships model recommendations — every box starts clean and the operator chooses models interactively.
PreviousNext