"gemma-3-12b-it-Q4_K_M":
    proxy: http://127.0.0.1:8999
    capabilities:
      context: 90000
    name: "gemma-3-12b-it-Q4_K_M"
    description: " -m /models/gemma-3-12b-it-Q4_K_M.gguf --mmproj /models/gemma-3-12b-mmproj-F16.gguf  --ctx-size 80000 --temp ${temp} --top-p 0.95 --top-k 20 --min-p 0.00  --reasoning-preserve --host 0.0.0.0 --fit off -ngl 40  --port 8999 --n-cpu-moe 35 -fa on --flash-attn on  --no-mmproj-offload  "
    cmd: /usr/local/bin/llama-server -m /models/gemma-3-12b-it-Q4_K_M.gguf --mmproj /models/gemma-3-12b-mmproj-F16.gguf   --ctx-size 80000 --temp ${temp} --top-p 0.95 --top-k 20 --min-p 0.00  --reasoning-preserve --host 0.0.0.0 --fit off -ngl 40  --port 8999 --n-cpu-moe 35 -fa on --flash-attn on  --no-mmproj-offload
    aliases:
      - "gemma-3-12b-it-Q4_K_M"
      - "llama-swap4"
    macros:
      "default_ctx": 90000
      "temp": 1.0
    metadata:
      temperature: ${temp}
      note: "The ${MODEL_ID} is running on port 8999 temp=${temp}, context=${default_ctx}"
    env:
      - "NVM_BIN=/root/.nvm/versions/node/v24.17.0/bin"
      - "NVM_DIR=/root/.nvm"
      - "NVM_INC=/root/.nvm/versions/node/v24.17.0/include/node"
      - "PATH=/root/.nvm/versions/node/v24.17.0/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin"
      - "CUDA_DEVICE_ORDER=PCI_BUS_ID"
    ttl: 60
    unloadTimeout: 30

Tuxedo17 a dit:

unknown model architecture: ‘mllama’

leafspark/Llama-3.2-11B-Vision-Instruct-GGUF · error loading model

llama_model_load: error loading model: error loading model architecture: unknown model architecture: 'mllama' llama_load_model_from_file: failed to load model

(huggingface.co)

Seulement avec ollama visiblement.

La principale différence entre ces deux versions réside dans l’écart temporel et le volume considérable de développement (plus d’un millier de commits) qui sépare le build 10030 du build 11102.

Voici les évolutions majeures introduites entre ces deux versions de llama.cpp :

Support des modèles multimodaux (mllama) : C'est le point critique pour votre cas. Le build 11102 intègre les parseurs et les couches nécessaires pour décoder l'architecture mllama (utilisée par Llama 3.2 Vision), alors que le build 10030 est trop ancien et renvoie l'erreur d'architecture inconnue.

Refonte et stabilité du serveur (llama-server) : De nombreux correctifs ont été apportés au serveur HTTP intégré, améliorant la gestion des connexions simultanées, des en-têtes CORS, et de la compatibilité avec l'API OpenAI.

Optimisations des backends matériels : Mises à jour des moteurs d'accélération (CUDA pour Nvidia, Metal pour Apple Silicon, Vulkan), améliorant l'efficacité du transfert des couches (-ngl) et la gestion de la mémoire VRAM.

Gestion de la mémoire et des contextes : Correction de bugs liés à l'allocation dynamique de la mémoire pour les contextes larges et les projecteurs de vision (mmproj).

# ./build/bin/llama-server -m /models/Llama-3.2-11B-Vision-Instruct.Q4_K_M.gguf --mmproj /models/Llama-3.2-11B-Vision-Instruct-mmproj.f16.gguf --host 0.0.0.0 --port 8999 -ngl 99
0.00.000.798 I srv  llama_server: initializing ...
0.03.603.511 I cmn  common_param: common_params_print_info: verbosity = 3 (adjust with the `-lv N` CLI arg)
0.03.604.094 W srv  llama_server: security: no API key is set and CORS allows all origins (see https://github.com/ggml-org/llama.cpp/pull/25655)
0.03.605.470 I srv    load_model: loading model '/models/Llama-3.2-11B-Vision-Instruct.Q4_K_M.gguf'
0.03.616.937 E mtmd_get_memory_usage: error: Failed to load CLIP model from /models/Llama-3.2-11B-Vision-Instruct-mmproj.f16.gguf

0.03.616.965 E srv    load_model: [mtmd] failed to get memory usage of mmproj
0.03.780.011 E llama_model_load: error loading model: unknown model architecture: 'mllama'
0.03.780.027 E llama_model_load_from_file_impl: failed to load model
0.03.780.131 E common_fit_params: encountered an error while trying to fit params to free device memory: failed to load model
0.03.920.968 E llama_model_load: error loading model: unknown model architecture: 'mllama'
0.03.921.005 E llama_model_load_from_file_impl: failed to load model
0.03.921.014 E cmn  common_init_: failed to load model '/models/Llama-3.2-11B-Vision-Instruct.Q4_K_M.gguf'
0.03.921.021 E srv    load_model: failed to load model, '/models/Llama-3.2-11B-Vision-Instruct.Q4_K_M.gguf'
0.03.921.023 I srv    operator(): operator(): cleaning up before exit...
0.03.922.987 E srv  llama_server: exiting due to model loading error

# git clone https://github.com/ggml-org/llama.cpp
# export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64
# export PATH=$PATH:$CUDA_HOME/bin
# cmake -B build -DGGML_CUDA=ON  -DCMAKE_CUDA_COMPILER=`which nvcc` -DLLAMA_CURL=ON -DCMAKE_CUDA_ARCHITECTURES=86
# cmake --build build -j$(nproc)
# systemctl stop llama-swap
# cmake --install build

Test :

#  /usr/local/bin/llama-server --version
0.00.000.502 I srv  llama_server: initializing ...
version: 0.4.1-dev (build 11102, commit bfd73a876)
built with GNU 13.3.0 for Linux x86_64

0.00.392.687 I srv    load_model: loading model '/models/Llama-3.2-11B-Vision-Instruct.Q4_K_M.gguf'
0.00.554.292 E llama_model_load: error loading model: unknown model architecture: 'mllama'
0.00.554.306 E llama_model_load_from_file_impl: failed to load model

Ma version :

# /usr/local/bin/llama-server --version
version: 10030 (c3d47e696)
built with GNU 13.3.0 for Linux x86_64

  "Best_Qwen3.6-27B-Q3_K_M.gguf_mmproj-BF16.gguf":
    proxy: http://127.0.0.1:8999
    capabilities:
      context: 90000
    name: "Best_Qwen3.6-27B-Q3_K_M"
    description: "-m /models/Qwen3.6-27B-Q3_K_M.gguf --mmproj /models/mmproj-BF16.gguf --ctx-size ${default_ctx} --temp ${temp} --top-p 0.95 --top-k 20 --min-p 0.00  --reasoning-preserve  --no-mmproj-offload --host 0.0.0.0 --fit off -ngl 40  --port 8999"
    cmd: /usr/local/bin/llama-server  -m /models/Qwen3.6-27B-Q3_K_M.gguf --mmproj /models/mmproj-BF16.gguf --ctx-size ${default_ctx} --temp ${temp} --top-p 0.95 --top-k 20 --min-p 0.00  --reasoning-preserve --host 0.0.0.0 --fit off -ngl 40  --port 8999
    aliases:
      - "best-Qwen3.6-27B-Q3_K_M"
      - "llama-swap2"
    macros:
      "default_ctx": 90000
      "temp": 1.0
    metadata:
      temperature: ${temp}
      note: "The ${MODEL_ID} is running on port 8999 temp=${temp}, context=${default_ctx}"
    env:
      - "NVM_BIN=/root/.nvm/versions/node/v24.17.0/bin"
      - "NVM_DIR=/root/.nvm"
      - "NVM_INC=/root/.nvm/versions/node/v24.17.0/include/node"
      - "PATH=/root/.nvm/versions/node/v24.17.0/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin"
      - "CUDA_DEVICE_ORDER=PCI_BUS_ID"
    ttl: 60
    unloadTimeout: 30

La nouvelle configuration :

  "Best_Qwen3.6-27B-Q3_K_M.gguf_mmproj-BF16.gguf":
    proxy: http://127.0.0.1:8999
    capabilities:
      context: 130000
    name: "Best_Qwen3.6-27B-Q3_K_M"
    description: "-m /models/Qwen3.6-27B-Q3_K_M.gguf --mmproj /models/mmproj-BF16.gguf --ctx-size ${default_ctx} --temp ${temp} --top-p 0.95 --top-k 20 --min-p 0.00  --reasoning-preserve  --no-mmproj-offload --host 0.0.0.0 --fit off -ngl 40  --port 8999"
    cmd: /usr/local/bin/llama-server  -m /models/Qwen3.6-27B-Q3_K_M.gguf --mmproj /models/mmproj-BF16.gguf --ctx-size ${default_ctx} --temp ${temp} --top-p 0.95 --top-k 20 --min-p 0.00  --reasoning-preserve --host 0.0.0.0 --fit off -ngl 40 --tensor-split 3,7  --port 8999
    aliases:
      - "best-Qwen3.6-27B-Q3_K_M"
      - "llama-swap2"
    macros:
      "default_ctx": 130000
      "temp": 1.0
    metadata:
      temperature: ${temp}
      note: "The ${MODEL_ID} is running on port 8999 temp=${temp}, context=${default_ctx}"
    env:
      - "NVM_BIN=/root/.nvm/versions/node/v24.17.0/bin"
      - "NVM_DIR=/root/.nvm"
      - "NVM_INC=/root/.nvm/versions/node/v24.17.0/include/node"
      - "PATH=/root/.nvm/versions/node/v24.17.0/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin"
      - "CUDA_DEVICE_ORDER=PCI_BUS_ID"
    ttl: 60
    unloadTimeout: 30

=> KO .

J’ai donc encore diminuer la taille du contexte à 100000 :

  "Best_Qwen3.6-27B-Q3_K_M.gguf_mmproj-BF16.gguf":
    proxy: http://127.0.0.1:8999
    capabilities:
      context: 100000
    name: "Best_Qwen3.6-27B-Q3_K_M"
    description: "-m /models/Qwen3.6-27B-Q3_K_M.gguf --mmproj /models/mmproj-BF16.gguf --ctx-size ${default_ctx} --temp ${temp} --top-p 0.95 --top-k 20 --min-p 0.00  --reasoning-preserve  --no-mmproj-offload --host 0.0.0.0 --fit off -ngl 40  --port 8999"
    cmd: /usr/local/bin/llama-server  -m /models/Qwen3.6-27B-Q3_K_M.gguf --mmproj /models/mmproj-BF16.gguf --ctx-size ${default_ctx} --temp ${temp} --top-p 0.95 --top-k 20 --min-p 0.00  --reasoning-preserve --host 0.0.0.0 --fit off -ngl 40 --tensor-split 3,7  --port 8999
    aliases:
      - "best-Qwen3.6-27B-Q3_K_M"
      - "llama-swap2"
    macros:
      "default_ctx": 100000
      "temp": 1.0
    metadata:
      temperature: ${temp}
      note: "The ${MODEL_ID} is running on port 8999 temp=${temp}, context=${default_ctx}"
    env:
      - "NVM_BIN=/root/.nvm/versions/node/v24.17.0/bin"
      - "NVM_DIR=/root/.nvm"
      - "NVM_INC=/root/.nvm/versions/node/v24.17.0/include/node"
      - "PATH=/root/.nvm/versions/node/v24.17.0/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin"
      - "CUDA_DEVICE_ORDER=PCI_BUS_ID"
    ttl: 60
    unloadTimeout: 30

=> OK.