Skip to content

convert : accept "ExaoneMoeForCausalLM" arch spelling - #26660

Merged
taronaeo merged 1 commit into
ggml-org:masterfrom
csabakecskemeti:fix/exaone-moe-arch-name-alias
Aug 6, 2026
Merged

convert : accept "ExaoneMoeForCausalLM" arch spelling#26660
taronaeo merged 1 commit into
ggml-org:masterfrom
csabakecskemeti:fix/exaone-moe-arch-name-alias

Conversation

@csabakecskemeti

Copy link
Copy Markdown
Contributor

Overview

LG AI Research renamed the architecture from ExaoneMoEForCausalLM to ExaoneMoeForCausalLM when exaone_moe was upstreamed into Transformers v5, and updated the configs of already-released models accordingly (K-EXAONE-236B-A23B config.json, commit f7db87a, "Update README.md and config.json for Transformers v5").

Freshly downloaded K-EXAONE-236B-A23B and K-EXAONE-2.0-750B-A37B checkpoints therefore fail conversion with

ERROR:hf-to-gguf:Model ExaoneMoeForCausalLM is not supported

even though the architecture is implemented. The spelling comes from the checkpoint's own config.json, so it is unaffected by the installed Transformers version.

Register both spellings so that older local copies keep working,

Additional information

Tested works, fresh GGUF
DevQuasar/LGAI-EXAONE.K-EXAONE-236B-A23B-GGUF

Requirements

  • I have read and agree with the contributing guidelines
    YES
  • AI usage disclosure:
    YES - Opus did the edit. I not even realized the slight change in the architecture name...

LG AI Research renamed the architecture from ExaoneMoEForCausalLM to
ExaoneMoeForCausalLM when exaone_moe was upstreamed into Transformers
v5, and updated the configs of already-released models accordingly
(K-EXAONE-236B-A23B config.json, commit f7db87a, "Update README.md and
config.json for Transformers v5").

Freshly downloaded K-EXAONE-236B-A23B and K-EXAONE-2.0-750B-A37B
checkpoints therefore fail conversion with

  ERROR:hf-to-gguf:Model ExaoneMoeForCausalLM is not supported

even though the architecture is implemented. The spelling comes from
the checkpoint's own config.json, so it is unaffected by the installed
Transformers version.

Register both spellings so that older local copies keep working, as is
already done for e.g. BitNet.

Note this fixes the architecture lookup only. K-EXAONE-236B-A23B
converts correctly with this change alone; K-EXAONE-2.0-750B-A37B
additionally needs support for its two leading dense layers (declared
via mlp_layer_types rather than first_k_dense_replace), its per-layer
sliding windows of 0/128/4096, and clamped SwiGLU on layers 62-77.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@ggml-gh-bot

This comment was marked as low quality.

@CISC CISC added the merge ready A maintainer can use this label to indicate that they consider the changes final and ready to merge. label Aug 6, 2026
@taronaeo
taronaeo merged commit e700bfb into ggml-org:master Aug 6, 2026
5 checks passed
miltos22 pushed a commit to miltos22/llama.cpp-wackMall-merge-request that referenced this pull request Aug 10, 2026
satindergrewal pushed a commit to satindergrewal/llama.cpp that referenced this pull request Aug 11, 2026
satindergrewal pushed a commit to satindergrewal/llama.cpp that referenced this pull request Aug 12, 2026
huaxel pushed a commit to huaxel/CachyLLama that referenced this pull request Aug 12, 2026
brittlewis12 pushed a commit to brittlewis12/llama.cpp that referenced this pull request Aug 17, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

conversion merge ready A maintainer can use this label to indicate that they consider the changes final and ready to merge.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants