vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-10 09:04:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

readme : model : mtdm : lfm2 improvements (#15476)

Browse source 
* Support untied embeddings

* Increase number of image tokens to 1024

* Add LFM2-VL to readme

* Actually use untied embeddings
This commit is contained in:
Tarek Dakhran 2025-08-22 09:29:08 +02:00 • committed by GitHub
parent a0f98dd604
commit e288693669
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
5 changed files with 11 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

1
README.md
View file

@ -151,6 +151,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
- [x] [Bunny](https://github.com/BAAI-DCAI/Bunny) - [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
- [x] [GLM-EDGE](https://huggingface.co/models?search=glm-edge) - [x] [GLM-EDGE](https://huggingface.co/models?search=glm-edge)
- [x] [Qwen2-VL](https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d) - [x] [Qwen2-VL](https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d)
- [x] [LFM2-VL](https://huggingface.co/collections/LiquidAI/lfm2-vl-68963bbc84a610f7638d5ffa)
</details> </details>

1
gguf-py/gguf/constants.py
View file

@ -2590,6 +2590,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.ATTN_K, MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V, MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT, MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.OUTPUT,
], ],
MODEL_ARCH.SMALLTHINKER: [ MODEL_ARCH.SMALLTHINKER: [
MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.TOKEN_EMBD,

1
src/llama-arch.cpp
View file

@ -2010,6 +2010,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" }, { LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" },
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" }, { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" }, { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
} }
}, },
{ {

10
src/llama-model.cpp
View file

@ -5474,8 +5474,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
} break; } break;
case LLM_ARCH_LFM2: case LLM_ARCH_LFM2:
{ {
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0); tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
if (output == NULL) {
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
}
for (int i = 0; i < n_layer; ++i) { for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i]; auto & layer = layers[i];
@ -17787,8 +17792,7 @@ struct llm_build_lfm2 : public llm_graph_context {
cb(cur, "model.embedding_norm", -1); cb(cur, "model.embedding_norm", -1);
res->t_embd = cur; res->t_embd = cur;
// lm_head is tied with embeddings cur = build_lora_mm(model.output, cur);
cur = build_lora_mm(model.tok_embd, cur);
cb(cur, "lm_head", -1); cb(cur, "lm_head", -1);
res->t_logits = cur; res->t_logits = cur;

2
tools/mtmd/clip.cpp
View file

@ -3513,7 +3513,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
const int height = img->ny; const int height = img->ny;
const int total_factor = params.patch_size * params.proj_scale_factor; const int total_factor = params.patch_size * params.proj_scale_factor;
constexpr int min_image_tokens = 64; constexpr int min_image_tokens = 64;
constexpr int max_image_tokens = 256; constexpr int max_image_tokens = 1024;
const float min_pixels = min_image_tokens * total_factor * total_factor; const float min_pixels = min_image_tokens * total_factor * total_factor;
const float max_pixels = max_image_tokens * total_factor * total_factor; const float max_pixels = max_image_tokens * total_factor * total_factor;