model: Add PaddleOCR-VL model support by megemini · Pull Request #18825 · ggml-org/llama.cpp

megemini · 2026-01-14T06:15:03Z

Add PaddleOCR-VL model support.

Test with some images:

A receipt

with command:

./build/bin/llama-cli -m /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR-VL-GGUF.gguf \
  --mmproj /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR-VL-GGUF-mmproj.gguf \
  --color on\
  --image /home/shun/Pictures/1640.jpeg \
  --prompt "OCR:"

A table

with command:

./build/bin/llama-cli -m /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR-VL-GGUF.gguf \
  --mmproj /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR-VL-GGUF-mmproj.gguf \
  --color on\
  --image /home/shun/Pictures/paddleocr.jpg \
  --prompt "Table Recognition:"

can be formatted:

p.s. Thanks to @ngxson #16701

Update

The model converted by commands:

python3 convert_hf_to_gguf.py /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR-VL \
  --outfile /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR-VL-GGUF.gguf \
  --outtype bf16 \
  --verbose
python3 convert_hf_to_gguf.py /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR-VL \
  --mmproj \
  --outfile /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR-VL-GGUF-mmproj.gguf \
  --outtype bf16 \
  --verbose

or, can be downloaded from : https://modelscope.cn/models/megemini/PaddleOCR-VL-GGUF/summary

…warmup

…addleocr-vl

ngxson

this PR looks a bit suspicious, please explicitly specify if you're using AI to generate it

ngxson · 2026-01-14T16:30:25Z

gguf-py/gguf/constants.py

+        MAX_PIXELS          = "clip.vision.max_pixels"
+        MIN_PIXELS          = "clip.vision.min_pixels"


Please use the correct naming from #18719

ngxson · 2026-01-14T16:31:41Z

src/models/paddleocr.cpp

what's the different between this and qwen2vl.cpp? seems like 100% identical

convert_hf_to_gguf.py

megemini · 2026-01-15T05:02:34Z

this PR looks a bit suspicious, please explicitly specify if you're using AI to generate it

@ngxson

First, thanks for your review, but, NO! Absolutely not!

The initial commit was based on your PR: #16701.

However, as you mentioned, Model generate hallucinated text, likely because of the projector being incorrect.

Therefore, I tried to compare the code with the implementation from huggingface/transformers#42178, which adds PaddleOCR-VL support to transformers.

To be frank, I’ve only been working with the llama.cpp source code for about two weeks, so I DO USE AI to help me understand the code. However, since llama.cpp is a C++ project, the AI isn't very helpful, so I'm forced to work on it all by myself!

PaddleOCR-VL is almost identical to ERNIE4.5, but there are two main differences from your PR:

PaddleOCR-VL uses mrope instead of rope, which requires rope_sections.
The positions tensor is different from ERNIE4.5 or Qwen2VL.

llama.cpp/tools/mtmd/clip.cpp

Lines 3529 to 3550 in 9d5a701

    
           case PROJECTOR_TYPE_PADDLEOCR: 
        
               { 
        
                   const int merge_ratio = hparams.n_merge; 
        
                   const int pw = image_size_width  / patch_size; 
        
                   const int ph = image_size_height / patch_size; 
        
                   std::vector<int> positions(n_pos * 4); 
        
                   int ptr = 0; 
        
                   for (int y = 0; y < ph; y += merge_ratio) { 
        
                       for (int dy = 0; dy < 2; dy++) { 
        
                           for (int x = 0; x < pw; x += merge_ratio) { 
        
                               for (int dx = 0; dx < 2; dx++) { 
        
                                   positions[                  ptr] = y + dy; 
        
                                   positions[    num_patches + ptr] = x + dx; 
        
                                   positions[2 * num_patches + ptr] = y + dy; 
        
                                   positions[3 * num_patches + ptr] = x + dx; 
        
                                   ptr++; 
        
                               } 
        
                           } 
        
                       } 
        
                   } 
        
                   set_input_i32("positions", positions);

Additionally, ERNIE4.5 had a minor issue where add_prefix_space needs to be added.

llama.cpp/convert_hf_to_gguf.py

Lines 3728 to 3736 in 9d5a701

    
           def set_vocab(self): 
        
               self._set_vocab_sentencepiece() 
        
               tokenizer_config_file = self.dir_model / 'tokenizer_config.json' 
        
               if tokenizer_config_file.is_file(): 
        
                   with open(tokenizer_config_file, "r", encoding="utf-8") as f: 
        
                       tokenizer_config_json = json.load(f) 
        
                       if "add_prefix_space" in tokenizer_config_json: 
        
                           self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])

Or, PaddleOCR-VL prompt is changed from OCR: to _OCR:, which makes PaddleOCR-VL error-prone, because it is trained on OCR:, and the generative model is only 0.3B, which may not be so smart.

what's the different between this and qwen2vl.cpp? seems like 100% identical

The src/models/paddleocr.cpp is based on src/models/ernie4-5.cpp, but uses ggml_rope_multi instead of ggml_rope_ext.

While src/models/paddleocr.cpp, src/models/ernie4-5.cpp and src/models/qwen2vl.cpp are almost identical, but,

llama.cpp/src/models/paddleocr.cpp

Lines 73 to 75 in 9d5a701

    
           cur = build_attn(inp_attn, 
        
                   model.layers[il].wo, NULL, 
        
                   Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);

llama.cpp/src/models/qwen2vl.cpp

Lines 68 to 70 in 36f0132

    
           cur = build_attn(inp_attn, 
        
                   model.layers[il].wo, model.layers[il].bo, 
        
                   Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);

ERNIE4.5 and PaddleOCR-VL do not require model.layers[il].bo, although they behave the same if model.layers[il].bo is NULL.

I don't ge why you need to make this overcomplicated, what's wrong with my code on #16701 where I just reuse the same Ernie4_5Model arch?

As I mentioned above, PaddleOCR-VL requires mrope and rope_sections

llama.cpp/src/llama-model.cpp

Lines 8318 to 8320 in 9d5a701

    
           case LLM_ARCH_QWEN2VL: 
        
           case LLM_ARCH_PADDLEOCR: 
        
               return LLAMA_ROPE_TYPE_MROPE;

ngxson · 2026-01-15T05:25:05Z

Thanks for the explanation, that's useful. I think it should be put onto the PR description to make reviewing easier.

Indeed, it seems like the model is just qwen2vl under the hood, minus small difference in tokenizer. At least for the language model, it can just reuse the same qwen2vl arch if the bo tensor is not present in the weight.

I'll come back to this after #18719 being merged

…addleocr-vl

megemini · 2026-01-29T13:06:22Z

Update 20260129

PaddleOCR-VL-1.5 released.

Model exported: https://modelscope.cn/models/megemini/PaddleOCR-VL-1.5-GGUF/summary

Test with new type of image:

with command and result:

Compare with gt:

ngxson · 2026-02-07T17:20:49Z

the chat template is broken, should be fixed via #19419

IIIIIllllIIIIIlllll · 2026-02-08T08:21:16Z

I encountered an error after building with the latest code. Could you please tell me how to resolve it? Thank you very much.

2026-02-08 16:20:05.648 - warmup: flash attention is enabled
2026-02-08 16:20:05.648 - srv    load_model: loaded multimodal model, '/home/mark/Models/Test/PaddleOCR-VL-1.5-BF16/mmproj.gguf'
2026-02-08 16:20:05.648 - srv    load_model: initializing slots, n_slots = 1
2026-02-08 16:20:05.660 - no implementations specified for speculative decoding
2026-02-08 16:20:05.660 - slot   load_model: id  0 | task -1 | speculative decoding context not initialized
2026-02-08 16:20:05.660 - slot   load_model: id  0 | task -1 | new slot, n_ctx = 32768
2026-02-08 16:20:05.660 - srv    load_model: prompt cache is enabled, size limit: 8192 MiB
2026-02-08 16:20:05.660 - srv    load_model: use `--cache-ram 0` to disable the prompt cache
2026-02-08 16:20:05.660 - srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
2026-02-08 16:20:05.661 - srv          init: init: chat template parsing error: 
2026-02-08 16:20:05.661 - ------------
2026-02-08 16:20:05.661 - While executing For at line 34, column 13 in source:
2026-02-08 16:20:05.661 - ...age["role"] == "system" -%}↵        {%- for content in message["content"] -%}↵  ...
2026-02-08 16:20:05.661 -                                            ^
2026-02-08 16:20:05.661 - Error: Expected iterable or object type in for loop: got String
2026-02-08 16:20:05.661 - srv          init: init: please consider disabling jinja via --no-jinja, or use a custom chat template via --chat-template
2026-02-08 16:20:05.661 - srv          init: init: for example: --no-jinja --chat-template chatml
2026-02-08 16:20:05.661 - srv    operator(): operator(): cleaning up before exit...
2026-02-08 16:20:05.662 - main: exiting due to model loading error

megemini · 2026-02-08T10:28:09Z

I encountered an error after building with the latest code. Could you please tell me how to resolve it? Thank you very much.

wait #19419 merged, or use template below:

{%- if not add_generation_prompt is defined -%}
    {%- set add_generation_prompt = true -%}
{%- endif -%}
{%- if not cls_token is defined -%}
    {%- set cls_token = "<|begin_of_sentence|>" -%}
{%- endif -%}
{%- if not eos_token is defined -%}
    {%- set eos_token = "</s>" -%}
{%- endif -%}
{{- cls_token -}}
{%- for message in messages -%}
    {%- if message["role"] == "user" -%}
        {{- "User: " -}}

      {%- if message["content"] is string -%}
        {{- message["content"] }}
      {%- else -%}

        {%- for content in message["content"] -%}
            {%- if content["type"] == "image" -%}
                {{ "<|IMAGE_START|><|IMAGE_PLACEHOLDER|><|IMAGE_END|>" }}
            {%- endif -%}
        {%- endfor -%}
        {%- for content in message["content"] -%}
            {%- if content["type"] == "text" -%}
                {{ content["text"] }}
            {%- endif -%}
        {%- endfor -%}

      {%- endif -%}


        {{ "\n" -}}
    {%- elif message["role"] == "assistant" -%}
        {{- "Assistant: " -}}

      {%- if message["content"] is string -%}
        {{- message["content"] }}
      {%- else -%}

        {%- for content in message["content"] -%}
            {%- if content["type"] == "text" -%}
                {{ content["text"] }}
            {%- endif -%}
        {%- endfor -%}

      {%- endif -%}


        {{ eos_token -}}
    {%- elif message["role"] == "system" -%}

      {%- if message["content"] is string -%}
        {{- message["content"] }}
      {%- else -%}

        {%- for content in message["content"] -%}
            {%- if content["type"] == "text" -%}
                {{ content["text"] + "\n" }}
            {%- endif -%}
        {%- endfor -%}

      {%- endif -%}

    {%- endif -%}
{%- endfor -%}
{%- if add_generation_prompt -%}
    {{- "Assistant: " -}}
{%- endif -%}

with command:

./build/bin/llama-cli -m /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR-VL-GGUF.gguf \
  --mmproj /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR-VL-GGUF-mmproj.gguf \
  --color on\
  --image /home/shun/Pictures/1640.jpeg \
  --prompt "OCR:" \
  --chat-template-file /media/shun/bigdata/Models/PaddleOCR_VL_SFT/PaddleOCR-VL/chat_template_llama.jinja

zhang-prog · 2026-02-09T07:44:38Z

@ngxson Is there any documentation for specific models in llama.cpp? (like PaddleOCR-VL vLLM recipes)? If there is, we’ll document how to use the model there; otherwise, we’ll put it in the PaddleOCR-VL docs.

zhang-prog · 2026-02-10T08:13:45Z

@ngxson @megemini
We have added documentation for llama-server to the PaddleOCR-VL documentation.

lilydjwg · 2026-02-10T09:03:19Z

@zhang-prog I have a simple question: to parse an image from start to result, is CUDA still needed to run the whole process despite llama.cpp has a Vulkan backend?

ngxson · 2026-02-10T15:13:31Z

The chat template fix has just been merged, I'll give it a try soon. Hopefully things work this time and I'll be able to merge this

ro99 · 2026-02-19T11:11:49Z

hello @zhang-prog @ngxson, I can see that PP-DocLayoutV3 is an intermediate step in the inference pipeline.

Will PP-DocLayoutV3 work with this? Or it has to be executed separated (e.g. with transformers)?

convert_hf_to_gguf.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

megemini · 2026-02-19T12:07:46Z

hello @zhang-prog @ngxson, I can see that PP-DocLayoutV3 is an intermediate step in the inference pipeline.

Will PP-DocLayoutV3 work with this? Or it has to be executed separated (e.g. with transformers)?

PP-DocLayoutV3 not included in llama.cpp

jiabochao · 2026-02-20T05:53:20Z

the result:
can be formatted:

@megemini Hi, I have a question: how is the LLM Spotting format converted to this document? Are there any libraries or tools that can convert this format to DOCX or Markdown?

IIIIIllllIIIIIlllll · 2026-02-20T06:08:04Z

I'm encountering this error when using the latest version (B8115):
tokenize: error: number of bitmaps (1) does not match number of markers (0).
It seems that it only works with the previous chat template. Is this as expected?
the chat template:

{%- if not add_generation_prompt is defined -%}
    {%- set add_generation_prompt = true -%}
{%- endif -%}
{%- if not cls_token is defined -%}
    {%- set cls_token = "<|begin_of_sentence|>" -%}
{%- endif -%}
{%- if not eos_token is defined -%}
    {%- set eos_token = "</s>" -%}
{%- endif -%}
{{- cls_token -}}
{%- for message in messages -%}
    {%- if message["role"] == "user" -%}
        {{- "User: " -}}

      {%- if message["content"] is string -%}
        {{- message["content"] }}
      {%- else -%}

        {%- for content in message["content"] -%}
            {%- if content["type"] == "image" -%}
                {{ "<|IMAGE_START|><|IMAGE_PLACEHOLDER|><|IMAGE_END|>" }}
            {%- endif -%}
        {%- endfor -%}
        {%- for content in message["content"] -%}
            {%- if content["type"] == "text" -%}
                {{ content["text"] }}
            {%- endif -%}
        {%- endfor -%}

      {%- endif -%}


        {{ "\n" -}}
    {%- elif message["role"] == "assistant" -%}
        {{- "Assistant: " -}}

      {%- if message["content"] is string -%}
        {{- message["content"] }}
      {%- else -%}

        {%- for content in message["content"] -%}
            {%- if content["type"] == "text" -%}
                {{ content["text"] }}
            {%- endif -%}
        {%- endfor -%}

      {%- endif -%}


        {{ eos_token -}}
    {%- elif message["role"] == "system" -%}

      {%- if message["content"] is string -%}
        {{- message["content"] }}
      {%- else -%}

        {%- for content in message["content"] -%}
            {%- if content["type"] == "text" -%}
                {{ content["text"] + "\n" }}
            {%- endif -%}
        {%- endfor -%}

      {%- endif -%}

    {%- endif -%}
{%- endfor -%}
{%- if add_generation_prompt -%}
    {{- "Assistant: " -}}
{%- endif -%}

megemini · 2026-02-20T07:04:52Z

@megemini Hi, I have a question: how is the LLM Spotting format converted to this document? Are there any libraries or tools that can convert this format to DOCX or Markdown?

https://github.com/PaddlePaddle/PaddleOCR/blob/main/docs/version3.x/algorithm/PaddleOCR-VL/PaddleOCR-VL-1.5.en.md#core-features

text spotting (text-line localization and recognition) with prompt Spotting: gives result in text-line recognition

with prompt OCR: gives result reorganized

I think this is the main difference.

As for the conversion of formats, can refer https://github.com/PaddlePaddle/PaddleX/blob/release/3.4/paddlex/inference/pipelines/paddleocr_vl/pipeline.py

megemini · 2026-02-20T07:10:37Z

I'm encountering this error when using the latest version (B8115): tokenize: error: number of bitmaps (1) does not match number of markers (0). It seems that it only works with the previous chat template. Is this as expected? the chat template:

b8115 https://github.com/ggml-org/llama.cpp/releases/download/b8115/llama-b8115-bin-ubuntu-x64.tar.gz works fine without --chat-template-file

IIIIIllllIIIIIlllll · 2026-02-20T07:30:33Z

I'm encountering this error when using the latest version (B8115): tokenize: error: number of bitmaps (1) does not match number of markers (0). It seems that it only works with the previous chat template. Is this as expected? the chat template:

b8115 https://github.com/ggml-org/llama.cpp/releases/download/b8115/llama-b8115-bin-ubuntu-x64.tar.gz works fine without --chat-template-file

oh, sorry, I'm using llama-server.
C:\Users\Mark\App\llama-server-v0.4.2-b8112-windows-vulkan\llamacpp\win-vulkan\llama-server.exe -m C:\Users\Mark\Models\PaddleOCR-VL-1.5\PaddleOCR-VL-1.5.gguf --port 8083 --mmproj C:\Users\Mark\Models\PaddleOCR-VL-1.5\mmproj-PaddleOCR-VL-1.5.gguf --ctx-size 252144 --flash-attn on --no-mmap --temp 0 --top-p 0.95 --top-k 40 --min-p 0.05 --presence-penalty 0.0 --repeat-penalty 1.0 --frequency-penalty 0.0 --batch-size 2048 --ubatch-size 2048

megemini · 2026-02-20T13:15:11Z

I'm encountering this error when using the latest version (B8115): tokenize: error: number of bitmaps (1) does not match number of markers (0). It seems that it only works with the previous chat template. Is this as expected? the chat template:

b8115 https://github.com/ggml-org/llama.cpp/releases/download/b8115/llama-b8115-bin-ubuntu-x64.tar.gz works fine without --chat-template-file

oh, sorry, I'm using llama-server. C:\Users\Mark\App\llama-server-v0.4.2-b8112-windows-vulkan\llamacpp\win-vulkan\llama-server.exe -m C:\Users\Mark\Models\PaddleOCR-VL-1.5\PaddleOCR-VL-1.5.gguf --port 8083 --mmproj C:\Users\Mark\Models\PaddleOCR-VL-1.5\mmproj-PaddleOCR-VL-1.5.gguf --ctx-size 252144 --flash-attn on --no-mmap --temp 0 --top-p 0.95 --top-k 40 --min-p 0.05 --presence-penalty 0.0 --repeat-penalty 1.0 --frequency-penalty 0.0 --batch-size 2048 --ubatch-size 2048

The easiest way to fix this issue is add <__media__> before prompt, e.g. change OCR: to <__media__>OCR: when using llama-server , and do NOT use --chat-template-file, because there is a slight difference between PaddleOCR-VL and PaddleOCR-VL-1.5 templates.

@ngxson I found that, with llama-server command, the prompt of body_parsed after oaicompat_chat_params_parse should be something like "<|begin_of_sentence|>User: <__media__>Formula Recognition:\nAssistant:\n" other than "<|begin_of_sentence|>User: Formula Recognition:\nAssistant:\n", which means <__media__> not been added properly.

#19419 changed common_chat_msgs_to_json_oaicompat to render_message_to_json, should oaicompat_chat_params_parse in server-common also be updated?

0xFS0CIETY · 2026-02-20T14:38:19Z

why i got error in llama-server but cli works fine did i done something wrong

megemini · 2026-02-20T14:41:52Z

why i got error in llama-server but cli works fine did i done something wrong

use <__media__>OCR: instead of OCR: , same issue #18825 (comment)

0xFS0CIETY · 2026-02-20T14:50:05Z

thanks its works

IIIIIllllIIIIIlllll · 2026-02-20T15:02:48Z

Thank you very much! However, as shown in the image, I'm unsure if my usage is correct.
I used llama-server, and the prompt was: '<media>Spotting:'.
In any case, the identified mathematical formulas all contained errors.

megemini · 2026-02-20T15:29:58Z

Thank you very much! However, as shown in the image, I'm unsure if my usage is correct. I used llama-server, and the prompt was: '<media>Spotting:'. In any case, the identified mathematical formulas all contained errors.

I think the best way to process this test image is to first preprocess it using PP-DocLayoutV3, and then identify the text and formula parts separately. Otherwise, it should not be expected to obtain results that are 100% consistent with using PaddleOCR with PaddleOCR-VL model. Especially the formula in the middle of the picture.

p.s. @IIIIIllllIIIIIlllll @0xFS0CIETY the <__media__>XXX: is just a temporary solution, not a feature

* support PaddleOCR-VL * clip: update PaddleOCR model loader parameters to prevent OOM during warmup * [update] add paddleocr vl text model instead of ernie4.5 * [update] restore change of minicpmv * [update] format * [update] format * [update] positions and patch merge permute * [update] mtmd_decode_use_mrope for paddleocr * [update] image min/max pixels * [update] remove set_limit_image_tokens * upate: preprocess without padding * clean up * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> --------- Co-authored-by: Xuan Son Nguyen <son@huggingface.co> Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

megemini added 11 commits December 19, 2025 12:24

support PaddleOCR-VL

b4cde7c

clip: update PaddleOCR model loader parameters to prevent OOM during …

64f0a46

…warmup

[update] add paddleocr vl text model instead of ernie4.5

fbfa906

Merge branch 'master' of https://github.com/ggml-org/llama.cpp into p…

54de283

…addleocr-vl

[update] restore change of minicpmv

0995fbb

Merge branch 'master' of https://github.com/ggml-org/llama.cpp into p…

6634ff1

…addleocr-vl

[update] format

73573b7

[update] format

65e43e4

[update] positions and patch merge permute

d54c871

Merge branch 'master' of https://github.com/ggml-org/llama.cpp into p…

0ab3b43

…addleocr-vl

[update] mtmd_decode_use_mrope for paddleocr

9d5a701

megemini requested review from CISC, ggerganov and ngxson as code owners January 14, 2026 06:15

github-actions bot added model Model specific examples python python script changes labels Jan 14, 2026

loci-dev mentioned this pull request Jan 14, 2026

UPSTREAM PR #18825: model: Add PaddleOCR-VL model support auroralabs-loci/llama.cpp#914

Open

ngxson requested changes Jan 14, 2026

View reviewed changes

ngxson self-assigned this Jan 15, 2026

megemini added 3 commits January 15, 2026 13:33

[update] image min/max pixels

2f8c194

Merge branch 'master' of https://github.com/ggml-org/llama.cpp into p…

b9fd3c8

…addleocr-vl

[update] remove set_limit_image_tokens

81de7fd

wqerrewetw mentioned this pull request Jan 19, 2026

Feature Request: support PaddleOCR-VL #16627

Closed

4 tasks

rick-github mentioned this pull request Jan 19, 2026

PaddleOCR-VL ollama/ollama#12685

Open

ownia mentioned this pull request Jan 28, 2026

common : convert string contents to arrays if template requires typed content #19156

Closed

loci-dev mentioned this pull request Jan 28, 2026

UPSTREAM PR #19156: common : convert string contents to arrays if template requires typed content auroralabs-loci/llama.cpp#1061

Open

Merge branch 'master' into paddleocr-vl

df44105

ngxson approved these changes Feb 19, 2026

View reviewed changes

CISC reviewed Feb 19, 2026

View reviewed changes

convert_hf_to_gguf.py Outdated Show resolved Hide resolved

convert_hf_to_gguf.py Outdated Show resolved Hide resolved

megemini and others added 2 commits February 19, 2026 19:59

Update convert_hf_to_gguf.py

501080a

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

Update convert_hf_to_gguf.py

74ac94a

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

CISC approved these changes Feb 19, 2026

View reviewed changes

ngxson merged commit 237958d into ggml-org:master Feb 19, 2026
78 of 79 checks passed

This was referenced Feb 22, 2026

chat: fix llama-server image placeholder issue for PaddleOCR-VL #19799

Closed

docs: 添加 llama.cpp 关于 tokenize 错误的解决方案说明 PaddlePaddle/PaddleOCR#17729

Open

		MAX_PIXELS = "clip.vision.max_pixels"
		MIN_PIXELS = "clip.vision.min_pixels"

Comments

Conversation

megemini commented Jan 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Update

Uh oh!

ngxson left a comment

Choose a reason for hiding this comment

Uh oh!

ngxson Jan 14, 2026

Choose a reason for hiding this comment

Uh oh!

megemini Jan 15, 2026

Choose a reason for hiding this comment

Uh oh!

ngxson Jan 14, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

megemini commented Jan 15, 2026

Uh oh!

ngxson commented Jan 15, 2026

Uh oh!

megemini commented Jan 29, 2026

Update 20260129

Uh oh!

ngxson commented Feb 7, 2026

Uh oh!

IIIIIllllIIIIIlllll commented Feb 8, 2026

Uh oh!

megemini commented Feb 8, 2026

Uh oh!

zhang-prog commented Feb 9, 2026

Uh oh!

zhang-prog commented Feb 10, 2026

Uh oh!

lilydjwg commented Feb 10, 2026

Uh oh!

ngxson commented Feb 10, 2026

Uh oh!

ro99 commented Feb 19, 2026

Uh oh!

Uh oh!

Uh oh!

megemini commented Feb 19, 2026

Uh oh!

Uh oh!

jiabochao commented Feb 20, 2026

Uh oh!

IIIIIllllIIIIIlllll commented Feb 20, 2026

Uh oh!

megemini commented Feb 20, 2026

Uh oh!

megemini commented Feb 20, 2026

Uh oh!

IIIIIllllIIIIIlllll commented Feb 20, 2026

Uh oh!

megemini commented Feb 20, 2026

Uh oh!

0xFS0CIETY commented Feb 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

megemini commented Feb 20, 2026

Uh oh!

0xFS0CIETY commented Feb 20, 2026

Uh oh!

IIIIIllllIIIIIlllll commented Feb 20, 2026

Uh oh!

megemini commented Feb 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

11 participants

megemini commented Jan 14, 2026 •

edited

Loading

0xFS0CIETY commented Feb 20, 2026 •

edited

Loading