[staging CI] unslothai/unsloth#8267 - #244
Open
danielhanchen wants to merge 19 commits into
Open
Conversation
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
…he UI A dit_block clears the family-level chip fields (params, qlora_vram_gb, gated, note) precisely so FamilyFacts falls through to vram_note, which carries the actionable reason: no CUDA, or no native bf16 on this GPU. base_specs was still published though, and the per-base entry wins in resolveDiffusionTrainingFacts, so selecting Klein base-9B on a blocked host put the 9B and 18 GB chips straight back. FamilyFacts renders vram_note only when there are no chips, so the user lost the explanation and got a generic refusal plus a size they cannot act on. Clear base_specs on the same condition. The overlay only ever feeds those chips (resolveDiffusionTrainingFacts is its sole consumer), so nothing functional depends on it being present. Covered by a test that pins a bf16-unsupported host and asserts both halves: the overlay is empty and vram_note still carries the reason. It also asserts a per-base overlay exists on an unblocked host first, so it cannot pass vacuously. Verified it fails with the fix reverted.
for more information, see https://pre-commit.ci
Two consequences of routing the fetch through prefer_ungated_mirror that the mirror was not meant to have. The gated-repo name check ran on the canonical id while the start route preflighted the fetch repo. For FLUX.1-dev or FLUX.2-dev without a token that means the route probes the public mirror, answers 200, frees the resident models, and only then does the child refuse the run by name. It now checks the repo the run will actually fetch, so the two agree: a mirror-backed run proceeds, and a genuinely gated fetch still fails in the route, before anything is evicted. base_revision was reading the fetch repo too. Which repo that is depends on local cache state, so the same base records a different rev- value once the upstream snapshot is evicted or the run moves to another machine, and mismatch_reason then refuses the checkpoint as a different base revision though the weights are byte identical. Every checkpoint written before mirrors existed holds the canonical value, so those would be refused as well. Both identity_for_config and the post-load re-read go back to the canonical id; a canonical repo with no local ref records "unresolved", which is already treated as not comparable. Tests, both mutation checked: the gate assertion fails when the canonical id is restored, and the revision assertion fails when the fetch repo is.
for more information, see https://pre-commit.ci
A mirror-backed run resolves the revision from the mirror, so recording the canonical repo's revision instead lost validation entirely: source_revision(canonical) is "unresolved" on exactly the hosts a mirror is selected for, and mismatch_reason skips non-comparable revisions, so the check silently never fired. Record both halves and compare them only when the repos match, which still catches a moved mirror and no longer refuses a legitimate cross-repo resume. Bundles written before the field exist keep comparing as they did. Also take the mirror on a token-less run even when the vendor repo is cached: only gated repos are in the mirror table, so a mirror existing means the upstream needs credentials this run does not have. UNSLOTH_DIFFUSION_NO_MIRROR still pins the vendor repo. And preselect the training base a family pairs with a loaded distilled checkpoint: the distilled half is never in base_repos, so opening Train with the 9B model loaded seeded the 4B base.
for more information, see https://pre-commit.ci
The mirror table stopped being gated-only when the redistributable bases joined it, so "a mirror exists" no longer implies "the upstream needs credentials". Klein base-4B is both a default trainable base and mirrored, and the Hub serves it anonymously, so a token-less run with a complete local cache was re-pulling it from the mirror and an offline run failed outright. Split the table rather than keep a second list in sync: the 12 genuinely gated pairs stay in _GATED_MIRROR_PAIRS, the redistributable ones move to _UNGATED_MIRROR_PAIRS, and mirror_repo and canonical_base build from the union so redirection and the base-keyed table normalisation are unchanged. upstream_is_gated reads the gated half, and the token-less override now asks it.
danielhanchen
force-pushed
the
pr-8267-ci
branch
from
August 10, 2026 12:21
14fdc34 to
c16ab8f
Compare
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
2 participants
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.
Disposable CI run for unslothai/unsloth#8267. Do not merge; closed after CI.