mtmd : Support Pixtral 12B

convert_hf_to_gguf.py

@@ -776,6 +776,9 @@ def get_vocab_base_pre(self, tokenizer) -> str:
         if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2":
             # ref: https://huggingface.co/THUDM/glm-4-9b-hf
             res = "glm4"
+        if chkhsh == "0e9433cbbb161f89e264eb32e8e64bfe69e834973ffca5d41d3948a604a3e2a3":
+            # ref: https://huggingface.co/mistral-community/pixtral-12b
+            res = "pixtral"
 
         if res is None:
             logger.warning("\n")
@@ -1724,7 +1727,8 @@ def prepare_tensors(self):
     "MistralForCausalLM",
     "MixtralForCausalLM",
     "Idefics3ForConditionalGeneration",
-    "SmolVLMForConditionalGeneration")
+    "SmolVLMForConditionalGeneration",
+    "LlavaForConditionalGeneration")
 class LlamaModel(TextModel):
     model_arch = gguf.MODEL_ARCH.LLAMA
     undo_permute = True
@@ -1734,6 +1738,10 @@ def __init__(self, *args, **kwargs):
         # fix for SmolVLM2, missing `num_attention_heads` in config.json
         if self.hparams["architectures"][0] == "SmolVLMForConditionalGeneration":
             self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 32)
+        # fix for Pixtral, missing `num_attention_heads` in config.json
+        if self.hparams["architectures"][0] == "LlavaForConditionalGeneration" \
+                and self.hparams.get("model_type") == "mistral":
+            self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 32)
 
     def set_vocab(self):
         try:
@@ -1797,12 +1805,17 @@ def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         n_head = self.hparams["num_attention_heads"]
         n_kv_head = self.hparams.get("num_key_value_heads")
-        is_vision_tensor = "vision_tower" in name or "vision_model" in name or "model.connector" in name
+        is_vision_tensor = "vision_tower" in name \
+            or "vision_model" in name \
+            or "model.connector" in name \
+            or "multi_modal_projector" in name
 
         if is_vision_tensor:
             return [] # skip vision tensors
         elif name.startswith("model.text_model"):
             name = name.replace("text_model.", "") # for SmolVLM
+        elif name.startswith("language_model."):
+            name = name.replace("language_model.", "") # for the rest
 
         if self.undo_permute:
             if name.endswith(("q_proj.weight", "q_proj.bias")):
@@ -1885,6 +1898,55 @@ def prepare_tensors(self):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("LlavaForConditionalGeneration")
+class LlavaVisionModel(VisionModel):
+    img_break_tok_id = -1
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        if self.hparams["model_type"] == "pixtral":
+            # fix missing config.json values
+            self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 16)
+            self.hparams["num_hidden_layers"] = self.hparams.get("num_hidden_layers", 24)
+            self.hparams["intermediate_size"] = self.hparams.get("intermediate_size", 4096)
+            self.hparams["hidden_size"] = self.hparams.get("hidden_size", 1024)
+            self.hparams["layer_norm_eps"] = self.hparams.get("layer_norm_eps", 1e-5)
+            self.img_break_tok_id = 12 # see tokenizer_config.json
+        else:
+            raise ValueError(f"Unsupported model type: {self.hparams['model_type']}")
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        if hparams["model_type"] == "pixtral":
+            self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.PIXTRAL)
+            # default values below are taken from HF tranformers code
+            self.gguf_writer.add_vision_attention_layernorm_eps(hparams["layer_norm_eps"])
+            self.gguf_writer.add_vision_use_silu(True)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = n_head
+
+        if name.startswith("multi_modal_projector.") or name.startswith("vision_tower."):
+            # process vision tensors
+            if name.endswith(("q_proj.weight", "q_proj.bias")):
+                data_torch = LlamaModel.permute(data_torch, n_head, n_head)
+            if name.endswith(("k_proj.weight", "k_proj.bias")):
+                data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
+            return [(self.map_tensor_name(name), data_torch)]
+
+        if self.img_break_tok_id > 0 and "embed_tokens.weight" in name:
+            logger.info(f"Extracting [IMG_BREAK] token embedding from {name}")
+            # for pixtral model, we need to extract the [IMG_BREAK] token embedding
+            img_break_embd = data_torch[self.img_break_tok_id]
+            name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK]
+            return [(self.map_tensor_name(name), img_break_embd)]
+
+        return [] # skip other tensors
+
+
 @ModelBase.register("Idefics3ForConditionalGeneration", "SmolVLMForConditionalGeneration")
 class SmolVLMModel(VisionModel):
     def __init__(self, *args, **kwargs):

convert_hf_to_gguf_update.py

@@ -115,6 +115,7 @@ class TOKENIZER_TYPE(IntEnum):
     {"name": "bailingmoe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-lite", },
     {"name": "llama4",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", },
     {"name": "glm4",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", },
+    {"name": "pixtral",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistral-community/pixtral-12b", },
 ]
 
 

docs/multimodal/gemma3.md

@@ -11,15 +11,15 @@ You can use pre-quantized model from [ggml-org](https://huggingface.co/ggml-org)
 ```bash
 # build
 cmake -B build
-cmake --build build --target llama-gemma3-cli
+cmake --build build --target llama-mtmd-cli
 
 # alternatively, install from brew (MacOS)
 brew install llama.cpp
 
 # run it
-llama-gemma3-cli -hf ggml-org/gemma-3-4b-it-GGUF
-llama-gemma3-cli -hf ggml-org/gemma-3-12b-it-GGUF
-llama-gemma3-cli -hf ggml-org/gemma-3-27b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-12b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-27b-it-GGUF
 
 # note: 1B model does not support vision
 ```
@@ -44,8 +44,8 @@ What you need:
 ```bash
 # build
 cmake -B build
-cmake --build build --target llama-gemma3-cli
+cmake --build build --target llama-mtmd-cli
 
 # run it
-./build/bin/llama-gemma3-cli -m {text_model}.gguf --mmproj mmproj.gguf --image your_image.jpg
+./build/bin/llama-mtmd-cli -m {text_model}.gguf --mmproj mmproj.gguf --image your_image.jpg
 ```

examples/llava/README.md

@@ -14,6 +14,28 @@ The naming and structure related to multimodal support have evolved, which might
 - [#12849](https://github.com/ggml-org/llama.cpp/pull/12849): `libmtmd` was introduced as a replacement for `llava.cpp`. Its goals include providing a single, unified command-line interface, improving the user/developer experience (UX/DX), and supporting both audio and image inputs.
 - [#13012](https://github.com/ggml-org/llama.cpp/pull/13012): `mtmd-cli` was added, consolidating the various model-specific CLIs into a single tool powered by `libmtmd`.
 
+## Pre-quantized models
+
+These are ready-to-use models, most of them come with `Q4_K_M` quantization by default:
+
+```sh
+# Gemma 3
+llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-12b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-27b-it-GGUF
+
+# SmolVLM
+llama-mtmd-cli -hf ggml-org/SmolVLM-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM-256M-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM-500M-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM2-2.2B-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM2-256M-Video-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM2-500M-Video-Instruct-GGUF
+
+# Pixtral 12B
+llama-mtmd-cli -hf ggml-org/pixtral-12b-GGUF
+```
+
 ## How it works and what is `mmproj`?
 
 Multimodal support in `llama.cpp` works by encoding images into embeddings using a separate model component, and then feeding these embeddings into the language model.
@@ -45,3 +67,9 @@ Multimodal projector (`mmproj`) files are specific to each model architecture. P
 - [MiniCPM-o 2.6](../../docs/multimodal/minicpmo2.6.md)
 - [IBM Granite Vision](../../docs/multimodal/granitevision.md)
 - [Google Gemma 3](../../docs/multimodal/gemma3.md)
+
+For the following models, you can use `convert_hf_to_gguf.py`with `--mmproj` flag to get the `mmproj` file:
+- [Gemma 3](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d) - Note: 1B variant does not have vision support
+- SmolVLM (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
+- SmolVLM2 (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
+- [Pixtral 12B](https://huggingface.co/mistral-community/pixtral-12b) - only works with `transformers`-compatible checkpoint

examples/llava/clip-impl.h

@@ -60,6 +60,7 @@
 #define TN_ATTN_V          "%s.blk.%d.attn_v.%s"
 #define TN_ATTN_OUTPUT     "%s.blk.%d.attn_out.%s"
 #define TN_FFN_DOWN        "%s.blk.%d.ffn_down.%s"
+#define TN_FFN_GATE        "%s.blk.%d.ffn_gate.%s"
 #define TN_FFN_UP          "%s.blk.%d.ffn_up.%s"
 #define TN_LN_1            "%s.blk.%d.ln1.%s"
 #define TN_LN_2            "%s.blk.%d.ln2.%s"
@@ -73,6 +74,7 @@
 #define TN_MM_INP_PROJ     "mm.input_projection.weight" // gemma3
 #define TN_MM_SOFT_EMB_N   "mm.soft_emb_norm.weight"    // gemma3
 #define TN_MM_PROJECTOR    "mm.model.fc.weight"         // idefics3
+#define TN_TOK_IMG_BREAK   "v.token_embd.img_break"     // pixtral
 
 // mimicpmv
 #define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
@@ -101,6 +103,7 @@ enum projector_type {
     PROJECTOR_TYPE_MERGER,
     PROJECTOR_TYPE_GEMMA3,
     PROJECTOR_TYPE_IDEFICS3,
+    PROJECTOR_TYPE_PIXTRAL,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -113,6 +116,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_MERGER,    "qwen2vl_merger"},
     { PROJECTOR_TYPE_GEMMA3,    "gemma3"},
     { PROJECTOR_TYPE_IDEFICS3,  "idefics3"},
+    { PROJECTOR_TYPE_PIXTRAL,   "pixtral"},
 };
 
 static projector_type clip_projector_type_from_string(const std::string & str) {