Add new model: Kimi2-6 (#45630)
* add kimi
* modular run conversion
* add processing for image for now
* video processor, maybe like this
* i think this is defi gemma4 rope
* add model tester
* convert weights
* fix processing
* looks fine with dummy weights
* chunk/permute rope dynamically
* fix repo
* fix some tests
* .
* rename to match remote code config
* rename tests as well
* oops
* update
* some comments
* more comments about model code
* fast tokenizer loaded, thasnk to moonshot
* .
* move my ugly code from cluster to GH, needs mocing to proper conversion
* easier to read, yet not in conversion mapping
* add as conversion
* small fix
* have to go, finish next week!
* resolve todos
* push the ep plan and copy from mixtral for now
* Fix compressed-tensors MoE decompression under tensor/expert parallelism
`DecompressExperts` unpacked each expert with the checkpoint's per-expert
`weight_shape`. Under TP/EP the loader shards every source sibling mapped to the
experts target (`weight_packed`/`weight_scale`/`weight_shape`) with the same op,
leaving the 2-element `weight_shape` empty on most ranks (stale on rank 0), so
`unpack_from_int32` raises `IndexError`. Rebuild each shard's `[out, in]` from the
packed tensor (`cols * pack_factor`): correct for every shard, and identical to the
stored shape in the non-distributed path (which is therefore unchanged).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix repo
* memory fragmentation
* tokenizer
* nit
* .
* even the official repo has this bug lol
* deepseek sybc with main, it was refactored
* docs and the last bits
* and this
* tester uses MoE layers
* fix reverse rename
* some comments
* video tests, flaky for now
* tests!
* fix tests
* fix repo
* no integration tests for now, too huge ckpts
* maybe? looks like same
* !
* more comments
* run check repo
* unused attr, delete
* dont import these unless torch is installed
* skip export tests
* fix modular nit
* aya
* erm, oke, mvoe it to another folder following fpquant
* use the correct ckpt for tests
* Update docs/source/en/model_doc/kimi_k25.md
Co-authored-by: Anton Vlasjuk <73884904+vasqu@users.noreply.github.com>
* fix kimi k26 export
* Revert "fix kimi k26 export"
This reverts commit 5952534f4684d1acd026ff689fe95516bf66e5fb.
* copy from main
* skip
---------
Co-authored-by: Arthur Zucker <arthur.zucker@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: vasqu <antonprogamer@gmail.com>
Co-authored-by: IlyasMoutawwakil <moutawwakil.ilyas.tsi@gmail.com>
Co-authored-by: Anton Vlasjuk <73884904+vasqu@users.noreply.github.com>
