Open
Conversation
EXAONE 4.5 uses <vision> and </vision> for image boundaries; Qwen keeps <|vision_start|> and <|vision_end|>. Route EXAONE 4.5 through the Qwen2.5-VL-style encode path (window attention pattern, optional mmproj input norm). Update exaone4_5 projector weights and convert_hf_to_gguf for mmproj export.
Align EXAONE4 tensor registration with EXAONE_MOE for NextN/MTP slots and avoid skip-flag propagation on duplicated rope_freqs so model loading succeeds for EXAONE 4.5 GGUF.
Contributor
|
please rebase before requesting a review |
github-actions
bot
added
model
…llama-cpp/exaone4-5
ngxson
requested changes
Apr 11, 2026
Comment on lines
+122
to
+123
| Kcur = clip_repeat_kv_heads(ctx0, Kcur, d_head, n_kv_head, n_head, n_patches); | ||
| Vcur = clip_repeat_kv_heads(ctx0, Vcur, d_head, n_kv_head, n_head, n_patches); |
Contributor
There was a problem hiding this comment.
this should be redundant, ggml support broadcasting automatically
Comment on lines
+177
to
+180
| window_idx = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos / 4); | ||
| ggml_set_name(window_idx, "window_idx"); | ||
| ggml_set_input(window_idx); | ||
|
|
Contributor
There was a problem hiding this comment.
move this to top. any inputs should be defined on top of cgraph
| hparams.n_merge = 2; | ||
| get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false); | ||
| get_u32(KEY_WIN_ATTN_PATTERN, hparams.n_wa_pattern, false); | ||
| hparams.set_limit_image_tokens(8, 4096); |
Contributor
There was a problem hiding this comment.
set_limit_image_tokens is only used by old model. if you already know the min/max pixels supported, write it to GGUF metadata instead
Comment on lines
+1473
to
+1476
| get_u32(KEY_N_HEAD_KV, hparams.n_kv_head, false); | ||
| if (hparams.n_kv_head <= 0) { | ||
| hparams.n_kv_head = 8; | ||
| } |
Contributor
There was a problem hiding this comment.
I see now reasons for this being an optional param. GGUF must have it, otherwise it's a faulty file
Suggested change
| get_u32(KEY_N_HEAD_KV, hparams.n_kv_head, false); | |
| if (hparams.n_kv_head <= 0) { | |
| hparams.n_kv_head = 8; | |
| } | |
| get_u32(KEY_N_HEAD_KV, hparams.n_kv_head); |
Contributor
There was a problem hiding this comment.
please specify exactly how this is different from qwen2, so that we can merge 2 models into one file in the future
Comment on lines
+10481
to
+10485
| remapper = { | ||
| "mtp.fc": "model.layers.{bid}.eh_proj", | ||
| "mtp.pre_fc_norm_embedding": "model.layers.{bid}.enorm", | ||
| "mtp.pre_fc_norm_hidden": "model.layers.{bid}.hnorm", | ||
| "mtp.norm": "model.layers.{bid}.shared_head.norm", |
Contributor
There was a problem hiding this comment.
please use proper tensor_mapping
Comment on lines
+10530
to
+10551
| if name.startswith("visual.") and ".qkv." in name: | ||
| assert self.hparams_vision is not None | ||
| hv = self.hparams_vision | ||
| n_heads = hv["num_heads"] | ||
| n_kv = int(hv.get("num_key_value_heads", n_heads)) | ||
| hidden = hv["hidden_size"] | ||
| head_dim = hidden // n_heads | ||
| q_dim = n_heads * head_dim | ||
| kv_dim = n_kv * head_dim | ||
| total_out = q_dim + 2 * kv_dim | ||
| out_dim = data_torch.shape[0] | ||
| if out_dim != total_out: | ||
| raise ValueError(f"EXAONE 4.5 vision qkv out dim mismatch: got {out_dim}, expected {total_out} ({name})") | ||
| wq = data_torch[:q_dim] | ||
| wk = data_torch[q_dim : q_dim + kv_dim] | ||
| wv = data_torch[q_dim + kv_dim :] | ||
| nq = name.replace("qkv", "q", 1) | ||
| nk = name.replace("qkv", "k", 1) | ||
| nv = name.replace("qkv", "v", 1) | ||
| yield from ModelBase.modify_tensors(self, wq, nq, bid) | ||
| yield from ModelBase.modify_tensors(self, wk, nk, bid) | ||
| yield from ModelBase.modify_tensors(self, wv, nv, bid) |
Contributor
There was a problem hiding this comment.
we should not split qkv, instead, use ggml_view to split the result (not the weight)
see build_vit in clip.cpp for an example
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
3 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.
Overview
Add support for the EXAONE 4.5 architecture for the EXAONE 4.5 model released by LG AI Research.
Additional information
This PR adds the modeling code for EXAONE 4.5, which uses the same LLM architecture as EXAONE 4.
It also adds
n_kv_headsto the CLIP model to make the ViT compatible with the GQA structure.Requirements
YES. The modeling code was implemented with the help of an AI assistant.