$ the-wire · showcase
Shrink keeps every layer kind, MQA and ResNet's identity shortcut
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
transformers-mic closes the shrink rule bugs a 192-fixture fleet sweep turned up, and adds a command that shows you exactly what a fixture dropped from its checkpoint.
Shrink keeps every layer kind, MQA and ResNet's identity shortcut huggingface/transformers-mic
A fleet-wide `tmic shrink` pass over 192 fixtures surfaced 22 with findings, and this lands the rule-side fixes. The builder now constructs the original and shrunk stacks on the meta device and compares layer kinds (module classes plus `layer_type`), keeping the prefix up to each kind's first layer capped at 8 layers; that rescues kinds the config derives from the index, taking DeepSeek-V4 hash...
tmic shrink: show how a fixture was shrunk from its checkpoint huggingface/transformers-mic
`tmic shrink [FIXTURE] [--open] [--json]` builds the source checkpoint and the fixture on the meta device from their configs, no weights downloaded, and writes a self-contained HTML page. It shows one square per layer colored by kind, parameter share per component, and which properties survived (attention style MHA/GQA/MQA, MoE routing, vocab, RoPE), plus changed config fields, so you can audit...
[Bot] Update Inference Providers documentation huggingface/hub-docs
The automated workflow bumped `@huggingface/tasks` and `@huggingface/inference` to latest and regenerated the Inference Providers documentation with `pnpm run generate`; merged as its own commit with Wauplin credited as co-author.
Change 'Paper' section to 'Tech reports' and add links huggingface/carbon
The project's "Paper" section is now "Tech reports", with links added to the tech reports and a blog post.
