Skip to content

fix(serve): hide incomplete GGUF shard sets from model list - #8

Merged
Vasyl-Trefilov merged 2 commits into
RustingStudio:mainfrom
bettercallcaleb:fix/complete-gguf-shards-model-list
Nov 6, 2026
Merged

Vasyl-Trefilov merged 2 commits into
RustingStudio:mainfrom
bettercallcaleb:fix/complete-gguf-shards-model-list

Conversation

@bettercallcaleb

Copy link
Copy Markdown
Contributor

Problem

The model list treats a split GGUF as loadable as soon as shard 1 exists, even if later shards are missing. It also sums shards with the same stem but different -of-N counts. This can show a model that cannot load, with an incorrect size.

Change

Require every numbered shard from 1 through the declared total, matching that exact total. Incomplete groups are omitted from the model list; regular single-file models remain unchanged.

Verification

  • New regression test failed before the fix and passed afterward, covering incomplete sets and matching stems with different shard totals.
  • cargo test --locked -- --skip spirv_is_fresh: passed.
  • git diff --check: passed.
  • The unfiltered local suite has a separate spirv_is_fresh failure because the installed GLSL compiler rejects an existing shader; neither shaders nor SPIR-V were modified.

Signed-off-by: bettercallcaleb <121191527+bettercallcaleb@users.noreply.github.com>
@Vasyl-Trefilov
Vasyl-Trefilov self-requested a review November 6, 2026 14:44
@Vasyl-Trefilov

Copy link
Copy Markdown
Member

Thanks for contribute, this is right. A model with a missing shard should not be offered, and summing shards across different -of-N totals was a real bug. The total <= files.len() bound is a good touch too.

@Vasyl-Trefilov
Vasyl-Trefilov merged commit bb03b0e into RustingStudio:main Nov 6, 2026
Vasyl-Trefilov pushed a commit that referenced this pull request Nov 8, 2026
Vasyl-Trefilov pushed a commit that referenced this pull request Nov 8, 2026
Vasyl-Trefilov pushed a commit that referenced this pull request Nov 8, 2026
…in cache (#9)

Closes #8. Arch/NVRTC bring-up for sm_70 (M1-T1): version detection, accurate 12.x guidance, a hard error for NVRTC >= 13 on pre-Ampere devices, Volta-marked startup line, and an NVRTC-tagged cubin cache key. A6 (version-scoped cache hit) stays deferred until M1-T2..M1-T6 make the sm_70 compile succeed.
Vasyl-Trefilov pushed a commit that referenced this pull request Nov 8, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants