fix: improve GGUF metadata parsing and reader selection (#2062)

Co-authored-by: leejet <leejet714@gmail.com>

docs/wan.md

@@ -40,6 +40,9 @@ Wan models require `-M vid_gen`, including single-frame generation. `--video-fra
             - safetensors: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/tree/main/split_files/diffusion_models
             - gguf: https://huggingface.co/QuantStack/Wan2.2-S2V-14B-GGUF/tree/main
             - int8_convrot safetensors: https://huggingface.co/noctrex/Wan2.2-S2V-14B-int8_convrot
+        - Wan2.2 VACE-Fun A14B
+            - safetensors: https://huggingface.co/alibaba-pai/Wan2.2-VACE-Fun-A14B
+            - gguf: https://huggingface.co/QuantStack/Wan2.2-VACE-Fun-A14B-GGUF/tree/main
 - Download vae
     - wan_2.1_vae (for all the wan model except Wan2.2 TI2V 5B)
         - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/vae/wan_2.1_vae.safetensors
@@ -256,3 +259,25 @@ ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\fr
 ```
 
 <video src=../assets/wan/Wan2.1_14B_vace_v2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.2 VACE-Fun A14B
+
+VACE-Fun runs as a MoE pair: `--diffusion-model` takes the low-noise expert and
+`--high-noise-diffusion-model` the high-noise one. Reference-to-video uses `-i`
+for the reference image, same as Wan2.1 VACE.
+
+#### R2V
+
+```
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model ..\models\diffusion_models\Wan2.2-VACE-Fun-A14B-low-noise-Q8_0.gguf --high-noise-diffusion-model ..\models\diffusion_models\Wan2.2-VACE-Fun-A14B-high-noise-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
+```
+
+<video src=../assets/wan/Wan2.2_A14B_vace_r2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+#### T2V
+
+Same command without `-i` (VACE context is synthesized from an empty control
+video, like Wan2.1 VACE t2v).
+
+> On GPUs with ~12 GB VRAM, VACE also needs `--vae-tiling` — the control-video
+> encode can exceed the budget otherwise.

src/model_io/gguf_io.cpp

@@ -47,10 +47,14 @@ bool read_gguf_file(const std::string& file_path,
     gguf_context* ctx_gguf_ = nullptr;
     ggml_context* ctx_meta_ = nullptr;
 
-    ctx_gguf_ = gguf_init_from_file(file_path.c_str(), {true, &ctx_meta_});
-    if (!ctx_gguf_) {
     GGUFReader gguf_reader;
-        if (!gguf_reader.load(file_path)) {
+    bool probe_ok = gguf_reader.load(file_path);
+
+    if (!probe_ok || !gguf_reader.has_tensors_beyond_ggml_limits()) {
+        ctx_gguf_ = gguf_init_from_file(file_path.c_str(), {true, &ctx_meta_});
+    }
+    if (!ctx_gguf_) {
+        if (!probe_ok && !gguf_reader.load(file_path)) {
             set_error(error, "failed to open '" + file_path + "' with GGUFReader");
             return false;
         }

src/model_io/gguf_reader_ext.h

@@ -35,23 +35,72 @@ enum class GGUFMetadataType : uint32_t {
 class GGUFReader {
 private:
     std::vector<GGUFTensorInfo> tensors_;
+    bool has_wide_tensors_    = false;
+    uint64_t remaining_bytes_ = 0;
     size_t data_offset_;
     size_t alignment_ = 32;  // default alignment is 32
 
     template <typename T>
     bool safe_read(std::ifstream& fin, T& value) {
-        fin.read(reinterpret_cast<char*>(&value), sizeof(T));
-        return fin.good();
+        return safe_read(fin, reinterpret_cast<char*>(&value), sizeof(T));
     }
 
     bool safe_read(std::ifstream& fin, char* buffer, size_t size) {
+        if (size > remaining_bytes_)
+            return false;
         fin.read(buffer, size);
-        return fin.good();
+        if (!fin.good())
+            return false;
+        remaining_bytes_ -= size;
+        return true;
     }
 
-    bool safe_seek(std::ifstream& fin, std::streamoff offset, std::ios::seekdir dir) {
-        fin.seekg(offset, dir);
-        return fin.good();
+    bool safe_skip(std::ifstream& fin, uint64_t count, uint64_t element_size = 1) {
+        if (count > remaining_bytes_ / element_size)
+            return false;
+        uint64_t size = count * element_size;
+        fin.seekg(static_cast<std::streamoff>(size), std::ios::cur);
+        if (!fin.good())
+            return false;
+        remaining_bytes_ -= size;
+        return true;
+    }
+
+    bool skip_metadata_values(std::ifstream& fin, GGUFMetadataType type, uint64_t count) {
+        switch (type) {
+            case GGUFMetadataType::UINT8:
+            case GGUFMetadataType::INT8:
+            case GGUFMetadataType::BOOL:
+                return safe_skip(fin, count);
+
+            case GGUFMetadataType::UINT16:
+            case GGUFMetadataType::INT16:
+                return safe_skip(fin, count, 2);
+
+            case GGUFMetadataType::UINT32:
+            case GGUFMetadataType::INT32:
+            case GGUFMetadataType::FLOAT32:
+                return safe_skip(fin, count, 4);
+
+            case GGUFMetadataType::UINT64:
+            case GGUFMetadataType::INT64:
+            case GGUFMetadataType::FLOAT64:
+                return safe_skip(fin, count, 8);
+
+            case GGUFMetadataType::STRING:
+                if (count > remaining_bytes_ / sizeof(uint64_t))
+                    return false;
+                for (uint64_t i = 0; i < count; i++) {
+                    uint64_t len = 0;
+                    if (!safe_read(fin, len) || !safe_skip(fin, len))
+                        return false;
+                }
+                return true;
+
+            default:
+                LOG_ERROR("Unknown metadata type=%u", static_cast<uint32_t>(type));
+                return false;
+        }
     }
 
     bool read_metadata(std::ifstream& fin) {
@@ -84,52 +133,12 @@ private:
             return true;
         }
 
-        switch (static_cast<GGUFMetadataType>(type)) {
-            case GGUFMetadataType::UINT8:
-            case GGUFMetadataType::INT8:
-            case GGUFMetadataType::BOOL:
-                return safe_seek(fin, 1, std::ios::cur);
-
-            case GGUFMetadataType::UINT16:
-            case GGUFMetadataType::INT16:
-                return safe_seek(fin, 2, std::ios::cur);
-
-            case GGUFMetadataType::UINT32:
-            case GGUFMetadataType::INT32:
-            case GGUFMetadataType::FLOAT32:
-                return safe_seek(fin, 4, std::ios::cur);
-
-            case GGUFMetadataType::UINT64:
-            case GGUFMetadataType::INT64:
-            case GGUFMetadataType::FLOAT64:
-                return safe_seek(fin, 8, std::ios::cur);
-
-            case GGUFMetadataType::STRING: {
-                uint64_t len = 0;
-                if (!safe_read(fin, len))
-                    return false;
-                return safe_seek(fin, len, std::ios::cur);
-            }
-
-            case GGUFMetadataType::ARRAY: {
-                uint32_t elem_type = 0;
-                uint64_t len       = 0;
-                if (!safe_read(fin, elem_type))
-                    return false;
-                if (!safe_read(fin, len))
-                    return false;
-
-                for (uint64_t i = 0; i < len; i++) {
-                    if (!read_metadata(fin))
-                        return false;
-                }
-                return true;
-            }
-
-            default:
-                LOG_ERROR("Unknown metadata type=%u", type);
+        uint64_t count = 1;
+        if (type == static_cast<uint32_t>(GGUFMetadataType::ARRAY)) {
+            if (!safe_read(fin, type) || !safe_read(fin, count))
                 return false;
         }
+        return skip_metadata_values(fin, static_cast<GGUFMetadataType>(type), count);
     }
 
     GGUFTensorInfo read_tensor_info(std::ifstream& fin) {
@@ -154,6 +163,7 @@ private:
         }
 
         if (n_dims > GGML_MAX_DIMS) {
+            has_wide_tensors_ = true;
             for (uint32_t i = GGML_MAX_DIMS; i < n_dims; i++) {
                 info.shape[GGML_MAX_DIMS - 1] *= info.shape[i];  // stack to last dim;
             }
@@ -174,12 +184,20 @@ private:
 
 public:
     bool load(const std::string& file_path) {
-        std::ifstream fin(file_path, std::ios::binary);
+        std::ifstream fin(file_path, std::ios::binary | std::ios::ate);
         if (!fin) {
             LOG_ERROR("failed to open '%s'", file_path.c_str());
             return false;
         }
 
+        std::streamoff file_size = fin.tellg();
+        if (file_size < 0)
+            return false;
+        remaining_bytes_ = static_cast<uint64_t>(file_size);
+        fin.seekg(0, std::ios::beg);
+        if (!fin.good())
+            return false;
+
         // --- Header ---
         char magic[4];
         if (!safe_read(fin, magic, 4) || strncmp(magic, "GGUF", 4) != 0) {
@@ -228,6 +246,8 @@ public:
     }
 
     const std::vector<GGUFTensorInfo>& tensors() const { return tensors_; }
+
+    bool has_tensors_beyond_ggml_limits() const { return has_wide_tensors_; }
     size_t data_offset() const { return data_offset_; }
 };