Merge remote-tracking branch 'origin/master' into out_of_range

Author: ngxson

cmake/build-info.cmake

@@ -41,14 +41,20 @@ endif()
 
 if(MSVC)
     set(BUILD_COMPILER "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
-    set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+    if (CMAKE_VS_PLATFORM_NAME)
+        set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+    else()
+        set(BUILD_TARGET "${CMAKE_SYSTEM_NAME} ${CMAKE_SYSTEM_PROCESSOR}")
+    endif()
 else()
     execute_process(
-        COMMAND sh -c "\"$@\" --version | head -1" _ ${CMAKE_C_COMPILER}
+        COMMAND ${CMAKE_C_COMPILER} --version
         OUTPUT_VARIABLE OUT
         OUTPUT_STRIP_TRAILING_WHITESPACE
     )
+    string(REGEX REPLACE " *\n.*" "" OUT "${OUT}")
     set(BUILD_COMPILER ${OUT})
+
     execute_process(
         COMMAND ${CMAKE_C_COMPILER} -dumpmachine
         OUTPUT_VARIABLE OUT

common/CMakeLists.txt

@@ -39,7 +39,9 @@ add_custom_command(
     COMMENT "Generating build details from Git"
     COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
             -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
-            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
+            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+            -DCMAKE_SYSTEM_NAME=${CMAKE_SYSTEM_NAME} -DCMAKE_SYSTEM_PROCESSOR=${CMAKE_SYSTEM_PROCESSOR}
+            -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
     WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.."
     DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
     VERBATIM

common/arg.cpp

@@ -217,13 +217,11 @@ struct curl_slist_ptr {
 #define CURL_MAX_RETRY 3
 #define CURL_RETRY_DELAY_SECONDS 2
 
-static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) {
+static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds, const char * method_name) {
     int remaining_attempts = max_attempts;
-    char * method = nullptr;
-    curl_easy_getinfo(curl, CURLINFO_EFFECTIVE_METHOD, &method);
 
     while (remaining_attempts > 0) {
-        LOG_INF("%s: %s %s (attempt %d of %d)...\n", __func__ , method, url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
+        LOG_INF("%s: %s %s (attempt %d of %d)...\n", __func__ , method_name, url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
 
         CURLcode res = curl_easy_perform(curl);
         if (res == CURLE_OK) {
@@ -287,24 +285,17 @@ static bool common_download_file_single(const std::string & url, const std::stri
             try {
                 metadata_in >> metadata;
                 LOG_DBG("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
-                if (metadata.contains("url") && metadata.at("url").is_string()) {
-                    auto previous_url = metadata.at("url").get<std::string>();
-                    if (previous_url != url) {
-                        LOG_ERR("%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str());
-                        return false;
-                    }
-                }
                 if (metadata.contains("etag") && metadata.at("etag").is_string()) {
                     etag = metadata.at("etag");
                 }
                 if (metadata.contains("lastModified") && metadata.at("lastModified").is_string()) {
                     last_modified = metadata.at("lastModified");
                 }
             } catch (const nlohmann::json::exception & e) {
-            LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
-                return false;
+                LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
             }
         }
+        // if we cannot open the metadata file, we assume that the downloaded file is not valid (etag and last-modified are left empty, so we will download it again)
     } else {
         LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
     }
@@ -350,7 +341,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
 
         // we only allow retrying once for HEAD requests
         // this is for the use case of using running offline (no internet), retrying can be annoying
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), 1, 0);
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), 1, 0, "HEAD");
         if (!was_perform_successful) {
             head_request_ok = false;
         }
@@ -432,7 +423,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
         // start the download
         LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
             llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS, "GET");
         if (!was_perform_successful) {
             return false;
         }

convert_hf_to_gguf.py

@@ -1899,7 +1899,10 @@ def prepare_tensors(self):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@ModelBase.register("LlavaForConditionalGeneration")
+@ModelBase.register(
+    "LlavaForConditionalGeneration", # pixtral
+    "Mistral3ForConditionalGeneration", # mistral small 3.1
+)
 class LlavaVisionModel(VisionModel):
     img_break_tok_id = -1
 
@@ -1908,17 +1911,38 @@ def __init__(self, *args, **kwargs):
         if self.hparams["model_type"] == "pixtral":
             # layer_norm_eps is not in config.json, it is hard-coded in modeling_pixtral.py
             self.hparams["layer_norm_eps"] = self.hparams.get("layer_norm_eps", 1e-5)
-            self.img_break_tok_id = 12 # see tokenizer_config.json
+            self.img_break_tok_id = self.get_token_id("[IMG_BREAK]")
+            logger.info(f"Image break token id: {self.img_break_tok_id}")
         else:
             raise ValueError(f"Unsupported model type: {self.hparams['model_type']}")
 
+    def get_token_id(self, token: str) -> int:
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+            added_tokens_decoder = json.load(f)['added_tokens_decoder']
+            for id_, token_data in added_tokens_decoder.items():
+                if token_data["content"] == token:
+                    return int(id_)
+        raise ValueError(f"Token '{token}' not found in tokenizer config.")
+
     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)
             self.gguf_writer.add_vision_attention_layernorm_eps(hparams["layer_norm_eps"])
-            self.gguf_writer.add_vision_use_silu(True)
+
+            # hidden_act
+            if hparams["hidden_act"] == "silu":
+                self.gguf_writer.add_vision_use_silu(True)
+            elif hparams["hidden_act"] == "gelu":
+                self.gguf_writer.add_vision_use_gelu(True)
+            else:
+                raise ValueError(f"Unsupported hidden_act: {hparams['hidden_act']}")
+
+            # spatial_merge_size
+            if "spatial_merge_size" in self.global_config:
+                self.gguf_writer.add_vision_spatial_merge_size(self.global_config["spatial_merge_size"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused

examples/llava/README.md

@@ -34,6 +34,9 @@ llama-mtmd-cli -hf ggml-org/SmolVLM2-500M-Video-Instruct-GGUF
 
 # Pixtral 12B
 llama-mtmd-cli -hf ggml-org/pixtral-12b-GGUF
+
+# Mistral Small 3.1 24B (IQ2_M quantization)
+llama-mtmd-cli -hf ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF --chat-template mistral-v7
 ```
 
 ## How it works and what is `mmproj`?
@@ -73,3 +76,4 @@ For the following models, you can use `convert_hf_to_gguf.py`with `--mmproj` fla
 - 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
+- [Mistral Small 3.1 24B](https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503)

examples/llava/clip-impl.h

@@ -31,6 +31,7 @@
 #define KEY_FEATURE_LAYER       "clip.vision.feature_layer"
 #define KEY_PROJ_SCALE_FACTOR   "clip.vision.projector.scale_factor"
 #define KEY_PROJ_TYPE           "clip.projector_type"
+#define KEY_SPATIAL_MERGE_SIZE  "clip.vision.spatial_merge_size"
 
 #define KEY_USE_GLU_MLP         "clip.use_glu_mlp"  // for qwen2.5vl
 #define KEY_USE_RMS_NORM        "clip.use_rms_norm" // for qwen2.5vl
@@ -68,9 +69,11 @@
 #define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
 #define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
 #define TN_IMAGE_NEWLINE   "model.image_newline"
+#define TN_MM_INP_NORM     "mm.input_norm.weight"
 #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_MM_PATCH_MERGER "mm.patch_merger.weight"     // mistral small 3.1
 #define TN_TOK_IMG_BREAK   "v.token_embd.img_break"     // pixtral
 
 // mimicpmv

examples/llava/clip.cpp

@@ -172,6 +172,7 @@ struct clip_hparams {
     std::unordered_set<int32_t> vision_feature_layer;
     int32_t attn_window_size = 0;
     int32_t n_wa_pattern = 0;
+    int32_t spatial_merge_size = 0;
 };
 
 struct clip_layer {
@@ -232,6 +233,7 @@ struct clip_vision_model {
     struct ggml_tensor * projection;
 
     // LLaVA projection
+    struct ggml_tensor * mm_input_norm_w = nullptr;
     struct ggml_tensor * mm_0_w = nullptr;
     struct ggml_tensor * mm_0_b = nullptr;
     struct ggml_tensor * mm_2_w = nullptr;
@@ -311,6 +313,7 @@ struct clip_vision_model {
 
     // pixtral
     struct ggml_tensor * token_embd_img_break = nullptr;
+    struct ggml_tensor * mm_patch_merger_w = nullptr;
 };
 
 struct clip_ctx {
@@ -637,6 +640,7 @@ static ggml_cgraph * clip_image_build_graph_pixtral(clip_ctx * ctx, const clip_i
     const int d_head      = hidden_size / n_head;
     const int n_layer     = hparams.n_layer;
     const float eps       = hparams.eps;
+    const int n_merge     = hparams.spatial_merge_size;
 
     struct ggml_init_params params = {
         /*.mem_size   =*/ ctx->buf_compute_meta.size(),
@@ -721,7 +725,13 @@ static ggml_cgraph * clip_image_build_graph_pixtral(clip_ctx * ctx, const clip_i
         {
             ggml_tensor * gate_proj = ggml_mul_mat(ctx0, model.layers[il].ff_gate_w, cur);
             ggml_tensor * up_proj   = ggml_mul_mat(ctx0, model.layers[il].ff_up_w,   cur);
-            gate_proj = ggml_silu(ctx0, gate_proj); // pixtral uses silu
+            if (ctx->use_silu) {
+                gate_proj = ggml_silu(ctx0, gate_proj);
+            } else if (ctx->use_gelu) {
+                gate_proj = ggml_gelu(ctx0, gate_proj);
+            } else {
+                GGML_ABORT("Pixtral: Unsupported activation");
+            }
             cur = ggml_mul(ctx0, up_proj, gate_proj);
             cur = ggml_mul_mat(ctx0, model.layers[il].ff_down_w, cur);
         }
@@ -732,14 +742,42 @@ static ggml_cgraph * clip_image_build_graph_pixtral(clip_ctx * ctx, const clip_i
         embeddings = cur;
     }
 
-    // LlavaMultiModalProjector (with GELU activation)
+    // mistral small 3.1 patch merger
+    // ref: https://github.com/huggingface/transformers/blob/7a3e208892c06a5e278144eaf38c8599a42f53e7/src/transformers/models/mistral3/modeling_mistral3.py#L67
+    if (model.mm_patch_merger_w) {
+        GGML_ASSERT(hparams.spatial_merge_size > 0);
+
+        ggml_tensor * cur = embeddings;
+        cur = ggml_mul(ctx0, ggml_rms_norm(ctx0, cur, eps), model.mm_input_norm_w);
+
+        // reshape image tokens to 2D grid
+        cur = ggml_reshape_3d(ctx0, cur, hidden_size, n_patches_x, n_patches_y);
+        cur = ggml_permute(ctx0, cur, 2, 0, 1, 3); // [x, y, hidden_size]
+        cur = ggml_cont(ctx0, cur);
+
+        // torch.nn.functional.unfold is just an im2col under the hood
+        // we just need a dummy kernel to make it work
+        ggml_tensor * kernel = ggml_view_3d(ctx0, cur, n_merge, n_merge, cur->ne[2], 0, 0, 0);
+        cur = ggml_im2col(ctx0, kernel, cur, n_merge, n_merge, 0, 0, 1, 1, true, inp->type);
+
+        // project to hidden_size
+        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], cur->ne[1] * cur->ne[2]);
+        cur = ggml_mul_mat(ctx0, model.mm_patch_merger_w, cur);
+        embeddings = cur;
+    }
+
+    // LlavaMultiModalProjector (always using GELU activation)
     {
         embeddings = ggml_mul_mat(ctx0, model.mm_1_w, embeddings);
-        embeddings = ggml_add(ctx0, embeddings, model.mm_1_b);
+        if (model.mm_1_b) {
+            embeddings = ggml_add(ctx0, embeddings, model.mm_1_b);
+        }
 
         embeddings = ggml_gelu(ctx0, embeddings);
         embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
-        embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
+        if (model.mm_2_b) {
+            embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
+        }
     }
 
     // arrangement of the [IMG_BREAK] token
@@ -749,11 +787,14 @@ static ggml_cgraph * clip_image_build_graph_pixtral(clip_ctx * ctx, const clip_i
         // and then concatenate the [IMG_BREAK] token to the end of each row, aka n_patches_per_row dimension
         // after the concatenation, we have a tensor with shape [hidden_size, n_patches_per_row + 1, n_rows]
 
+        const int p_y             = n_merge > 0 ? n_patches_y / n_merge : n_patches_y;
+        const int p_x             = n_merge > 0 ? n_patches_x / n_merge : n_patches_x;
+        const int p_total         = p_x * p_y;
         const int n_embd_text     = embeddings->ne[0];
-        const int n_tokens_output = num_patches + n_patches_y - 1; // one [IMG_BREAK] per row, except the last row
+        const int n_tokens_output = p_total + p_y - 1; // one [IMG_BREAK] per row, except the last row
 
-        ggml_tensor * cur = ggml_reshape_3d(ctx0, embeddings, n_embd_text, n_patches_x, n_patches_y);
-        ggml_tensor * tok = ggml_new_tensor_3d(ctx0, embeddings->type, n_embd_text, 1, n_patches_y);
+        ggml_tensor * cur = ggml_reshape_3d(ctx0, embeddings, n_embd_text, p_x, p_y);
+        ggml_tensor * tok = ggml_new_tensor_3d(ctx0, embeddings->type, n_embd_text, 1, p_y);
         tok = ggml_scale(ctx0, tok, 0.0); // clear the tensor
         tok = ggml_add(ctx0, tok, model.token_embd_img_break);
         cur = ggml_concat(ctx0, cur, tok, 1);
@@ -1734,6 +1775,7 @@ struct clip_model_loader {
                 case PROJECTOR_TYPE_PIXTRAL:
                     {
                         hparams.rope_theta = 10000.0f;
+                        get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.spatial_merge_size, false);
                     } break;
                 case PROJECTOR_TYPE_QWEN25VL:
                     {
@@ -1957,11 +1999,14 @@ struct clip_model_loader {
             case PROJECTOR_TYPE_PIXTRAL:
                 {
                     vision_model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"));
-                    vision_model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"));
+                    vision_model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"), false);
                     vision_model.mm_2_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"));
-                    vision_model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"));
+                    vision_model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"), false);
                     // [IMG_BREAK] token embedding
                     vision_model.token_embd_img_break = get_tensor(TN_TOK_IMG_BREAK);
+                    // for mistral small 3.1
+                    vision_model.mm_input_norm_w   = get_tensor(TN_MM_INP_NORM,     false);
+                    vision_model.mm_patch_merger_w = get_tensor(TN_MM_PATCH_MERGER, false);
                 } break;
             default:
                 GGML_ASSERT(false && "unknown projector type");
@@ -2516,7 +2561,7 @@ struct llava_uhd {
 
         // no pinpoints, dynamically calculate the grid size (e.g. minicpmv)
 
-        auto best_size    = get_best_resize(original_size, slice_size, patch_size, has_slices);
+        auto best_size    = get_best_resize(original_size, slice_size, patch_size, !has_slices);
         res.overview_size = best_size;
 
         if (!has_slices) {
@@ -2926,8 +2971,9 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
     } else if (ctx->proj_type == PROJECTOR_TYPE_IDEFICS3) {
         n_patches /= ctx->vision_model.hparams.proj_scale_factor;
     } else if (ctx->proj_type == PROJECTOR_TYPE_PIXTRAL) {
-        int n_patches_x = img->nx / params.patch_size;
-        int n_patches_y = img->ny / params.patch_size;
+        int n_merge = ctx->vision_model.hparams.spatial_merge_size;
+        int n_patches_x = img->nx / params.patch_size / (n_merge > 0 ? n_merge : 1);
+        int n_patches_y = img->ny / params.patch_size / (n_merge > 0 ? n_merge : 1);
         n_patches = n_patches_y*n_patches_x + n_patches_y - 1; // + one [IMG_BREAK] per row, except the last row
     }
 
@@ -3484,7 +3530,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
             return ctx->vision_model.mm_model_peg_0_b->ne[0];
         case PROJECTOR_TYPE_MLP:
         case PROJECTOR_TYPE_PIXTRAL:
-            return ctx->vision_model.mm_2_b->ne[0];
+            return ctx->vision_model.mm_2_w->ne[1];
         case PROJECTOR_TYPE_MLP_NORM:
             return ctx->vision_model.mm_3_b->ne[0];
         case PROJECTOR_TYPE_MINICPMV:

examples/llava/mtmd-cli.cpp

@@ -96,6 +96,7 @@ struct mtmd_cli_context {
             LOG_ERR("Model does not have chat template.\n");
             LOG_ERR("  For old llava models, you may need to use '--chat-template vicuna'\n");
             LOG_ERR("  For MobileVLM models, use '--chat-template deepseek'\n");
+            LOG_ERR("  For Mistral Small 3.1, use '--chat-template mistral-v7'\n");
             exit(1);
         }
 

examples/llava/tests.sh

@@ -59,6 +59,7 @@ add_test "llama-mtmd-cli"  "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M"
 
 # to test the big models, run: ./tests.sh big
 add_test_big "llama-mtmd-cli" "ggml-org/pixtral-12b-GGUF:Q4_K_M"
+add_test_big "llama-mtmd-cli" "ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF" "mistral-v7"
 
 # these models always give the wrong answer, not sure why
 # add_test "llama-mtmd-cli"  "ggml-org/SmolVLM-Instruct-GGUF:Q4_K_M"

ggml/CMakeLists.txt

@@ -360,3 +360,27 @@ write_basic_package_version_file(
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake
               ${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/ggml)
+
+if (MSVC)
+    set(MSVC_WARNING_FLAGS
+        /wd4005  # Macro redefinition
+        /wd4244  # Conversion from one type to another type, possible loss of data
+        /wd4267  # Conversion from 'size_t' to a smaller type, possible loss of data
+    )
+    function(disable_msvc_warnings target_name)
+        if(TARGET ${target_name})
+            target_compile_options(${target_name} PRIVATE ${MSVC_WARNING_FLAGS})
+        endif()
+    endfunction()
+
+    disable_msvc_warnings(ggml-base)
+    disable_msvc_warnings(ggml)
+    disable_msvc_warnings(ggml-cpu)
+    disable_msvc_warnings(ggml-cpu-x64)
+    disable_msvc_warnings(ggml-cpu-sse42)
+    disable_msvc_warnings(ggml-cpu-sandybridge)
+    disable_msvc_warnings(ggml-cpu-haswell)
+    disable_msvc_warnings(ggml-cpu-skylakex)
+    disable_msvc_warnings(ggml-cpu-icelake)
+    disable_msvc_warnings(ggml-cpu-alderlake)
+endif()