From aa2341298924ac89778252015efcb792f2df1e20 Mon Sep 17 00:00:00 2001 From: John <78893154+cmp-nct@users.noreply.github.com> Date: Wed, 14 Feb 2024 08:38:35 +0100 Subject: [PATCH] llava : support v1.6 (#5267) * Create llava-survery-v2.py * Update convert-image-encoder-to-gguf.py * Update convert-image-encoder-to-gguf.py * Rename llava-survery-v2.py to llava-surgery-v2.py * Update convert-image-encoder-to-gguf.py will now search for projector * Update convert-image-encoder-to-gguf.py whoops * Update llava-surgery-v2.py * Clip: Bugfix for normalization (it did not loat the 3 std and mean values) Clip: bicubic resize function Clip: added save-to-bmp/pil for debugging and conversion from/to 32/8 images Clip: added normalization with FP16 precision simulation (image tensors match HF implementation, can be switched off, only used for llava-1.6) Clip: added newline tensor, mergetype kv, image-grid kv, new resize-pad function with resolution from gridpoints Clip: clip_image_preprocess now returns a float * vector instead of float, this way llava 1.5 and 1.6 is supported llava: added ggml cpu graph for embedding patching, added spatial_unpad preliminary support, added a lot of comments that need to be cleaned when all is final convert-image-encoder: fixed image-grid flattening * whitespace corrections * ws * Tensors are now properly permuted. Before the embeddings were inserted 1:1, now they are split into the 24x24 patches as in reference. * ws * added verbose_prompt support into cli added stopwords for llava-1.6 into cli * moved llava functions to llava.cpp, made clip.h C compatible API, replaced vector style functions with pointers, added a debug define to remove functions from compilation while not needed * ws * convert : skip unknown tensors (need for LLaVA) * llava : update readme * llava : fix compile warnings * llava : style * convert : add --skip-unknown CLI arg * server : remove clip structs * bugfix for non llava-1.6 It should now work with llava-1.5 as well * clip : minor code rearrange * llava : update readme a bit --------- Co-authored-by: John Co-authored-by: Georgi Gerganov --- convert.py | 37 +- examples/llava/README.md | 12 +- examples/llava/clip.cpp | 766 +++++++++++++++--- examples/llava/clip.h | 47 +- .../llava/convert-image-encoder-to-gguf.py | 66 +- examples/llava/llava-cli.cpp | 26 +- examples/llava/llava-surgery-v2.py | 167 ++++ examples/llava/llava.cpp | 296 ++++++- examples/llava/llava.h | 2 - examples/server/server.cpp | 15 +- 10 files changed, 1229 insertions(+), 205 deletions(-) create mode 100644 examples/llava/llava-surgery-v2.py diff --git a/convert.py b/convert.py index 323e8058d..63a0a5d78 100755 --- a/convert.py +++ b/convert.py @@ -1173,7 +1173,7 @@ def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyM for (name, tensor) in model.items()} -def convert_model_names(model: LazyModel, params: Params) -> LazyModel: +def convert_model_names(model: LazyModel, params: Params, skip_unknown: bool) -> LazyModel: tmap = gguf.TensorNameMap(ARCH, params.n_layer) should_skip: set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, [])) @@ -1199,7 +1199,11 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel: for name, lazy_tensor in model.items(): tensor_type, name_new = tmap.get_type_and_name(name, try_suffixes = (".weight", ".bias")) or (None, None) if name_new is None: - raise Exception(f"Unexpected tensor name: {name}") + if skip_unknown: + print(f"Unexpected tensor name: {name} - skipping") + continue + else: + raise Exception(f"Unexpected tensor name: {name}. Use --skip-unknown to ignore it (e.g. LLaVA)") if tensor_type in should_skip: print(f"skipping tensor {name_new}") @@ -1377,19 +1381,20 @@ def main(args_in: list[str] | None = None) -> None: output_choices.append("q8_0") vocab_types = ["spm", "bpe", "hfft"] parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file") - parser.add_argument("--awq-path", type=Path, help="Path to scale awq cache file", default=None) - parser.add_argument("--dump", action="store_true", help="don't convert, just show what's in the model") - parser.add_argument("--dump-single", action="store_true", help="don't convert, just show what's in a single model file") - parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") - parser.add_argument("--outtype", choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)") - parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file") - parser.add_argument("--vocab-type", choices=vocab_types, help="The vocabulary format used to define the tokenizer model (default: spm)", default="spm") - parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") - parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)") - parser.add_argument("--ctx", type=int, help="model training context (default: based on input)") - parser.add_argument("--concurrency", type=int, help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default=DEFAULT_CONCURRENCY) - parser.add_argument("--big-endian", action="store_true", help="model is executed on big endian machine") - parser.add_argument("--pad-vocab", action="store_true", help="add pad tokens when model vocab expects more than tokenizer metadata provides") + parser.add_argument("--awq-path", type=Path, help="Path to scale awq cache file", default=None) + parser.add_argument("--dump", action="store_true", help="don't convert, just show what's in the model") + parser.add_argument("--dump-single", action="store_true", help="don't convert, just show what's in a single model file") + parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") + parser.add_argument("--outtype", choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)") + parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file") + parser.add_argument("--vocab-type", choices=vocab_types, help="The vocabulary format used to define the tokenizer model (default: spm)", default="spm") + parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") + parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)") + parser.add_argument("--ctx", type=int, help="model training context (default: based on input)") + parser.add_argument("--concurrency", type=int, help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default=DEFAULT_CONCURRENCY) + parser.add_argument("--big-endian", action="store_true", help="model is executed on big endian machine") + parser.add_argument("--pad-vocab", action="store_true", help="add pad tokens when model vocab expects more than tokenizer metadata provides") + parser.add_argument("--skip-unknown", action="store_true", help="skip unknown tensor names instead of failing") args = parser.parse_args(args_in) if args.awq_path: @@ -1461,7 +1466,7 @@ def main(args_in: list[str] | None = None) -> None: print(f"Special vocab info: {special_vocab}") model = model_plus.model - model = convert_model_names(model, params) + model = convert_model_names(model, params, args.skip_unknown) ftype = pick_output_type(model, args.outtype) model = convert_to_output_type(model, ftype) outfile = args.outfile or default_outfile(model_plus.paths, ftype) diff --git a/examples/llava/README.md b/examples/llava/README.md index 19f1a50a2..e2ef0eff1 100644 --- a/examples/llava/README.md +++ b/examples/llava/README.md @@ -19,9 +19,9 @@ After building, run: `./llava-cli` to see the usage. For example: **note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so. -## Model conversion +## LLaVA 1.5 -- Clone `llava-v15-7b` and `clip-vit-large-patch14-336` locally: +- Clone a LLaVA and a CLIP model ([available options](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)). For example: ```sh git clone https://huggingface.co/liuhaotian/llava-v1.5-7b @@ -55,8 +55,14 @@ python ./convert.py ../llava-v1.5-7b Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` directory. +## LLaVA 1.6 + +- Use `llava-surgery-v2.py` + +- TODO: add detailed instructions + ## TODO -- [ ] Support non-CPU backend for the image encoding part. +- [x] Support non-CPU backend for the image encoding part. - [ ] Support different sampling methods. - [ ] Support more model variants. diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index ccd0d85ad..9c5091e61 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -1,7 +1,7 @@ // NOTE: This is modified from clip.cpp only for LLaVA, // so there might be still unnecessary artifacts hanging around // I'll gradually clean and extend it - +// Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch #include "clip.h" #include "ggml.h" #include "ggml-alloc.h" @@ -30,6 +30,26 @@ #include #include #include +#include + +//#define CLIP_DEBUG_FUNCTIONS + +// RGB uint8 image +struct clip_image_u8 { + int nx; + int ny; + + std::vector buf; +}; + +// RGB float32 image (NHWC) +// Memory layout: RGBRGBRGB... +struct clip_image_f32 { + int nx; + int ny; + + std::vector buf; +}; static std::string format(const char * fmt, ...) { va_list ap; @@ -50,50 +70,56 @@ static std::string format(const char * fmt, ...) { // key constants // -#define KEY_FTYPE "general.file_type" -#define KEY_NAME "general.name" -#define KEY_DESCRIPTION "general.description" -#define KEY_HAS_TEXT_ENC "clip.has_text_encoder" -#define KEY_HAS_VIS_ENC "clip.has_vision_encoder" +#define KEY_FTYPE "general.file_type" +#define KEY_NAME "general.name" +#define KEY_DESCRIPTION "general.description" +#define KEY_HAS_TEXT_ENC "clip.has_text_encoder" +#define KEY_HAS_VIS_ENC "clip.has_vision_encoder" #define KEY_HAS_LLAVA_PROJ "clip.has_llava_projector" -#define KEY_USE_GELU "clip.use_gelu" -#define KEY_N_EMBD "clip.%s.embedding_length" -#define KEY_N_FF "clip.%s.feed_forward_length" -#define KEY_N_BLOCK "clip.%s.block_count" -#define KEY_N_HEAD "clip.%s.attention.head_count" +#define KEY_USE_GELU "clip.use_gelu" +#define KEY_N_EMBD "clip.%s.embedding_length" +#define KEY_N_FF "clip.%s.feed_forward_length" +#define KEY_N_BLOCK "clip.%s.block_count" +#define KEY_N_HEAD "clip.%s.attention.head_count" #define KEY_LAYER_NORM_EPS "clip.%s.attention.layer_norm_epsilon" -#define KEY_PROJ_DIM "clip.%s.projection_dim" -#define KEY_TOKENS "tokenizer.ggml.tokens" -#define KEY_N_POSITIONS "clip.text.context_length" -#define KEY_IMAGE_SIZE "clip.vision.image_size" -#define KEY_PATCH_SIZE "clip.vision.patch_size" -#define KEY_IMAGE_MEAN "clip.vision.image_mean" -#define KEY_IMAGE_STD "clip.vision.image_std" -#define KEY_PROJ_TYPE "clip.projector_type" +#define KEY_PROJ_DIM "clip.%s.projection_dim" +#define KEY_TOKENS "tokenizer.ggml.tokens" +#define KEY_N_POSITIONS "clip.text.context_length" +#define KEY_IMAGE_SIZE "clip.vision.image_size" +#define KEY_PATCH_SIZE "clip.vision.patch_size" +#define KEY_IMAGE_MEAN "clip.vision.image_mean" +#define KEY_IMAGE_STD "clip.vision.image_std" +#define KEY_PROJ_TYPE "clip.projector_type" + +#define KEY_MM_PATCH_MERGE_TYPE "clip.vision.mm_patch_merge_type" +#define KEY_IMAGE_GRID_PINPOINTS "clip.vision.image_grid_pinpoints" +#define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution" + // // tensor name constants // -#define TN_TOKEN_EMBD "%s.token_embd.weight" -#define TN_POS_EMBD "%s.position_embd.weight" -#define TN_CLASS_EMBD "v.class_embd" -#define TN_PATCH_EMBD "v.patch_embd.weight" -#define TN_ATTN_K "%s.blk.%d.attn_k.%s" -#define TN_ATTN_Q "%s.blk.%d.attn_q.%s" -#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_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" -#define TN_LN_PRE "%s.pre_ln.%s" -#define TN_LN_POST "%s.post_ln.%s" -#define TN_TEXT_PROJ "text_projection.weight" -#define TN_VIS_PROJ "visual_projection.weight" -#define TN_LLAVA_PROJ "mm.%d.%s" -#define TN_MVLM_PROJ_MLP "mm.model.mlp.%d.%s" +#define TN_TOKEN_EMBD "%s.token_embd.weight" +#define TN_POS_EMBD "%s.position_embd.weight" +#define TN_CLASS_EMBD "v.class_embd" +#define TN_PATCH_EMBD "v.patch_embd.weight" +#define TN_ATTN_K "%s.blk.%d.attn_k.%s" +#define TN_ATTN_Q "%s.blk.%d.attn_q.%s" +#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_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" +#define TN_LN_PRE "%s.pre_ln.%s" +#define TN_LN_POST "%s.post_ln.%s" +#define TN_TEXT_PROJ "text_projection.weight" +#define TN_VIS_PROJ "visual_projection.weight" +#define TN_LLAVA_PROJ "mm.%d.%s" +#define TN_MVLM_PROJ_MLP "mm.model.mlp.%d.%s" #define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s" +#define TN_IMAGE_NEWLINE "model.image_newline" enum projector_type { @@ -104,8 +130,8 @@ enum projector_type { }; static std::map PROJECTOR_TYPE_NAMES = { - { PROJECTOR_TYPE_MLP, "mlp" }, - { PROJECTOR_TYPE_LDP, "ldp" }, + { PROJECTOR_TYPE_MLP, "mlp" }, + { PROJECTOR_TYPE_LDP, "ldp" }, }; @@ -165,7 +191,6 @@ static std::string gguf_data_to_str(enum gguf_type type, const void * data, int } } - static void replace_all(std::string & s, const std::string & search, const std::string & replace) { std::string result; for (size_t pos = 0; ; pos += search.length()) { @@ -217,7 +242,7 @@ static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) { } } -static void print_tensor_info(const ggml_tensor* tensor, const char* prefix = "") { +static void print_tensor_info(const ggml_tensor * tensor, const char * prefix = "") { size_t tensor_size = ggml_nbytes(tensor); printf("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%" PRId64 ", %" PRId64 ", %" PRId64 ", %" PRId64 "], type = %s\n", prefix, ggml_n_dims(tensor), tensor->name, tensor_size, @@ -233,31 +258,136 @@ static projector_type clip_projector_type_from_string(const std::string & name) return PROJECTOR_TYPE_UNKNOWN; } -// -// image data -// +#ifdef CLIP_DEBUG_FUNCTIONS +static void clip_image_write_image_to_ppm(const clip_image_u8& img, const std::string& filename) { + std::ofstream file(filename, std::ios::binary); + if (!file.is_open()) { + std::cerr << "Failed to open file for writing: " << filename << std::endl; + return; + } -// RGB uint8 image -struct clip_image_u8 { - int nx; - int ny; + // PPM header: P6 format, width, height, and max color value + file << "P6\n" << img.nx << " " << img.ny << "\n255\n"; - std::vector buf; -}; + // Write pixel data + for (size_t i = 0; i < img.buf.size(); i += 3) { + // PPM expects binary data in RGB format, which matches our image buffer + file.write(reinterpret_cast(&img.buf[i]), 3); + } -// RGB float32 image (NHWC) -// Memory layout: RGBRGBRGB... -struct clip_image_f32 { - int nx; - int ny; + file.close(); +} + +static void clip_image_save_to_bmp(const clip_image_u8& img, const std::string& filename) { + std::ofstream file(filename, std::ios::binary); + if (!file.is_open()) { + std::cerr << "Failed to open file for writing: " << filename << std::endl; + return; + } + + int fileSize = 54 + 3 * img.nx * img.ny; // File header + info header + pixel data + int bytesPerPixel = 3; + int widthInBytes = img.nx * bytesPerPixel; + int paddingAmount = (4 - (widthInBytes % 4)) % 4; + int stride = widthInBytes + paddingAmount; + + // Bitmap file header + unsigned char fileHeader[14] = { + 'B','M', // Signature + 0,0,0,0, // Image file size in bytes + 0,0,0,0, // Reserved + 54,0,0,0 // Start of pixel array + }; + + // Total file size + fileSize = 54 + (stride * img.ny); + fileHeader[2] = (unsigned char)(fileSize); + fileHeader[3] = (unsigned char)(fileSize >> 8); + fileHeader[4] = (unsigned char)(fileSize >> 16); + fileHeader[5] = (unsigned char)(fileSize >> 24); + + // Bitmap information header (BITMAPINFOHEADER) + unsigned char infoHeader[40] = { + 40,0,0,0, // Size of this header (40 bytes) + 0,0,0,0, // Image width + 0,0,0,0, // Image height + 1,0, // Number of color planes + 24,0, // Bits per pixel + 0,0,0,0, // No compression + 0,0,0,0, // Image size (can be 0 for no compression) + 0,0,0,0, // X pixels per meter (not specified) + 0,0,0,0, // Y pixels per meter (not specified) + 0,0,0,0, // Total colors (color table not used) + 0,0,0,0 // Important colors (all are important) + }; + + // Width and height in the information header + infoHeader[4] = (unsigned char)(img.nx); + infoHeader[5] = (unsigned char)(img.nx >> 8); + infoHeader[6] = (unsigned char)(img.nx >> 16); + infoHeader[7] = (unsigned char)(img.nx >> 24); + infoHeader[8] = (unsigned char)(img.ny); + infoHeader[9] = (unsigned char)(img.ny >> 8); + infoHeader[10] = (unsigned char)(img.ny >> 16); + infoHeader[11] = (unsigned char)(img.ny >> 24); + + // Write file headers + file.write(reinterpret_cast(fileHeader), sizeof(fileHeader)); + file.write(reinterpret_cast(infoHeader), sizeof(infoHeader)); + + // Pixel data + std::vector padding(3, 0); // Max padding size to be added to each row + for (int y = img.ny - 1; y >= 0; --y) { // BMP files are stored bottom-to-top + for (int x = 0; x < img.nx; ++x) { + // Each pixel + size_t pixelIndex = (y * img.nx + x) * 3; + unsigned char pixel[3] = { + img.buf[pixelIndex + 2], // BMP stores pixels in BGR format + img.buf[pixelIndex + 1], + img.buf[pixelIndex] + }; + file.write(reinterpret_cast(pixel), 3); + } + // Write padding for the row + file.write(reinterpret_cast(padding.data()), paddingAmount); + } + + file.close(); +} + +// debug function to convert f32 to u8 +static void clip_image_convert_f32_to_u8(const clip_image_f32& src, clip_image_u8& dst) { + dst.nx = src.nx; + dst.ny = src.ny; + dst.buf.resize(3 * src.nx * src.ny); + for (size_t i = 0; i < src.buf.size(); ++i) { + dst.buf[i] = static_cast(std::min(std::max(int(src.buf[i] * 255.0f), 0), 255)); + } +} +#endif - std::vector buf; -}; // // clip layers // +struct clip_hparams { + int32_t image_size; + int32_t patch_size; + int32_t hidden_size; + int32_t n_intermediate; + int32_t projection_dim; + int32_t n_head; + int32_t n_layer; + + float eps; + + char mm_patch_merge_type[32] = "flat"; // spatial_unpad or flat (default) + + int32_t image_grid_pinpoints[32]; + int32_t image_crop_resolution; +}; + struct clip_layer { // attention struct ggml_tensor * k_w; @@ -287,7 +417,7 @@ struct clip_layer { }; struct clip_vision_model { - struct clip_vision_hparams hparams; + struct clip_hparams hparams; // embeddings struct ggml_tensor * class_embedding; @@ -310,6 +440,8 @@ struct clip_vision_model { struct ggml_tensor * mm_2_w = NULL; struct ggml_tensor * mm_2_b = NULL; + struct ggml_tensor * image_newline = NULL; + // Yi type models with mlp+normalization projection struct ggml_tensor * mm_1_w = NULL; // Yi type models have 0, 1, 3, 4 struct ggml_tensor * mm_1_b = NULL; @@ -364,9 +496,10 @@ struct clip_ctx { std::vector buf_compute_meta; // memory buffers to evaluate the model - ggml_backend_buffer_t params_buffer = NULL; + ggml_backend_buffer_t params_buffer = NULL; ggml_backend_buffer_t compute_buffer = NULL; - ggml_backend_t backend = NULL; + + ggml_backend_t backend = NULL; ggml_gallocr_t compute_alloc = NULL; }; @@ -379,18 +512,19 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 const auto & model = ctx->vision_model; const auto & hparams = model.hparams; - const int image_size = hparams.image_size; - const int patch_size = hparams.patch_size; - const int num_patches = ((image_size / patch_size) * (image_size / patch_size)); - const int num_positions = num_patches + 1; - const int hidden_size = hparams.hidden_size; - const int n_head = hparams.n_head; - const int d_head = hidden_size / n_head; - const int n_layer = hparams.n_layer; - //const int n_intermediate = hparams.n_intermediate; - //const int projection_dim = hparams.projection_dim; - const float eps = hparams.eps; - int batch_size = imgs->size; + const int image_size = hparams.image_size; + const int patch_size = hparams.patch_size; + const int num_patches = ((image_size / patch_size) * (image_size / patch_size)); + const int num_patches_per_side = image_size / patch_size; GGML_UNUSED(num_patches_per_side); + const int num_positions = num_patches + 1; + const int hidden_size = hparams.hidden_size; + const int n_head = hparams.n_head; + const int d_head = hidden_size / n_head; + const int n_layer = hparams.n_layer; + const float eps = hparams.eps; + + const int batch_size = imgs->size; + if (ctx->has_llava_projector) { GGML_ASSERT(batch_size == 1); } @@ -540,7 +674,6 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 embeddings = ggml_add(ctx0, embeddings, model.mm_0_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); @@ -791,10 +924,10 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { if (idx != -1) { const std::string proj_type = gguf_get_val_str(ctx, idx); new_clip->proj_type = clip_projector_type_from_string(proj_type); - } - else { + } else { new_clip->proj_type = PROJECTOR_TYPE_MLP; } + if (new_clip->proj_type == PROJECTOR_TYPE_MLP) { if (gguf_find_tensor(ctx, format(TN_LLAVA_PROJ, 3, "weight").c_str()) != -1) { new_clip->proj_type = PROJECTOR_TYPE_MLP_NORM; @@ -920,11 +1053,41 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { hparams.projection_dim = get_u32(ctx, format(KEY_PROJ_DIM, "vision")); hparams.eps = get_f32(ctx, format(KEY_LAYER_NORM_EPS, "vision")); + try { + int idx = get_key_idx(ctx, KEY_IMAGE_GRID_PINPOINTS); + int n = gguf_get_arr_n(ctx, idx); + const int32_t * pinpoints = (const int32_t *)gguf_get_arr_data(ctx, idx); + for (int i = 0; i < 32 && i < n && pinpoints[i] != 0; ++i) { + hparams.image_grid_pinpoints[i] = pinpoints[i]; + } + if (n < 32) + hparams.image_grid_pinpoints[n] = 0; + } catch (std::runtime_error & e) { + hparams.image_grid_pinpoints[0]=0; + } + + try { + int idx = get_key_idx(ctx, KEY_MM_PATCH_MERGE_TYPE); + strcpy(hparams.mm_patch_merge_type, gguf_get_val_str(ctx, idx)); + } catch (std::runtime_error & e) { + strcpy(hparams.mm_patch_merge_type, "flat"); + } + + try { + hparams.image_crop_resolution = get_u32(ctx, KEY_IMAGE_CROP_RESOLUTION); // llava-1.6 + } catch(const std::exception& e) { + hparams.image_crop_resolution = hparams.image_size; + } + int idx_mean = get_key_idx(ctx, KEY_IMAGE_MEAN); int idx_std = get_key_idx(ctx, KEY_IMAGE_STD); + + const float * mean_data = (const float *)gguf_get_arr_data(ctx, idx_mean); + const float * std_data = (const float *)gguf_get_arr_data(ctx, idx_std); + for (int i = 0; i < 3; ++i) { - new_clip->image_mean[i] = *((const float *)gguf_get_arr_data(ctx, idx_mean)); - new_clip->image_std[i] = *((const float *)gguf_get_arr_data(ctx, idx_std)); + new_clip->image_mean[i] = mean_data[i]; + new_clip->image_std[i] = std_data[i]; } if (verbosity >= 2) { @@ -936,13 +1099,27 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { printf("v_projection_dim %d\n", hparams.projection_dim); printf("v_n_head %d\n", hparams.n_head); printf("v_n_layer %d\n", hparams.n_layer); + printf("v_eps %f\n", hparams.eps); + printf("v_image_mean %f %f %f\n", new_clip->image_mean[0], new_clip->image_mean[1], new_clip->image_mean[2]); + printf("v_image_std %f %f %f\n", new_clip->image_std[0], new_clip->image_std[1], new_clip->image_std[2]); + printf("v_image_grid_pinpoints: "); + for (int i = 0; i < 32 & hparams.image_grid_pinpoints[i]!=0; ++i) { + printf("%d ", hparams.image_grid_pinpoints[i]); + } + printf("\n"); + printf("v_mm_patch_merge_type: %s\n", hparams.mm_patch_merge_type); + } - vision_model.patch_embeddings = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD); - vision_model.class_embedding = get_tensor(new_clip->ctx_data, TN_CLASS_EMBD); - vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v")); - vision_model.pre_ln_w = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight")); - vision_model.pre_ln_b = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias")); + try { + vision_model.patch_embeddings = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD); + vision_model.class_embedding = get_tensor(new_clip->ctx_data, TN_CLASS_EMBD); + vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v")); + vision_model.pre_ln_w = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight")); + vision_model.pre_ln_b = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias")); + } catch(const std::exception& e) { + fprintf(stderr, "%s: failed to load vision model tensors\n", __func__); + } // LLaVA projection if (new_clip->proj_type == PROJECTOR_TYPE_MLP || new_clip->proj_type == PROJECTOR_TYPE_MLP_NORM) { @@ -968,40 +1145,43 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { vision_model.mm_4_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "weight")); vision_model.mm_4_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "bias")); } catch (std::runtime_error & e) { } - } - else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) { + try { + vision_model.image_newline = get_tensor(new_clip->ctx_data, TN_IMAGE_NEWLINE); + // fprintf(stderr, "%s: image_newline tensor (llava-1.6) found\n", __func__); + } catch (std::runtime_error & e) { } + } else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) { // MobileVLM projection - vision_model.mm_model_mlp_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "weight")); - vision_model.mm_model_mlp_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "bias")); - vision_model.mm_model_mlp_3_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "weight")); - vision_model.mm_model_mlp_3_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "bias")); - vision_model.mm_model_block_1_block_0_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "0.weight")); - vision_model.mm_model_block_1_block_0_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.weight")); - vision_model.mm_model_block_1_block_0_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.bias")); + vision_model.mm_model_mlp_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "weight")); + vision_model.mm_model_mlp_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "bias")); + vision_model.mm_model_mlp_3_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "weight")); + vision_model.mm_model_mlp_3_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "bias")); + vision_model.mm_model_block_1_block_0_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "0.weight")); + vision_model.mm_model_block_1_block_0_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.weight")); + vision_model.mm_model_block_1_block_0_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.bias")); vision_model.mm_model_block_1_block_1_fc1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.weight")); vision_model.mm_model_block_1_block_1_fc1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.bias")); vision_model.mm_model_block_1_block_1_fc2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.weight")); vision_model.mm_model_block_1_block_1_fc2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.bias")); - vision_model.mm_model_block_1_block_2_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "0.weight")); - vision_model.mm_model_block_1_block_2_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.weight")); - vision_model.mm_model_block_1_block_2_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.bias")); - vision_model.mm_model_block_2_block_0_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "0.weight")); - vision_model.mm_model_block_2_block_0_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.weight")); - vision_model.mm_model_block_2_block_0_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.bias")); + vision_model.mm_model_block_1_block_2_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "0.weight")); + vision_model.mm_model_block_1_block_2_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.weight")); + vision_model.mm_model_block_1_block_2_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.bias")); + vision_model.mm_model_block_2_block_0_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "0.weight")); + vision_model.mm_model_block_2_block_0_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.weight")); + vision_model.mm_model_block_2_block_0_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.bias")); vision_model.mm_model_block_2_block_1_fc1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.weight")); vision_model.mm_model_block_2_block_1_fc1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.bias")); vision_model.mm_model_block_2_block_1_fc2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.weight")); vision_model.mm_model_block_2_block_1_fc2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.bias")); - vision_model.mm_model_block_2_block_2_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "0.weight")); - vision_model.mm_model_block_2_block_2_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.weight")); - vision_model.mm_model_block_2_block_2_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.bias")); - } - else { + vision_model.mm_model_block_2_block_2_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "0.weight")); + vision_model.mm_model_block_2_block_2_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.weight")); + vision_model.mm_model_block_2_block_2_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.bias")); + } else { std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type]; throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str())); } vision_model.layers.resize(hparams.n_layer); + for (int il = 0; il < hparams.n_layer; ++il) { auto & layer = vision_model.layers[il]; layer.k_w = get_tensor(new_clip->ctx_data, format(TN_ATTN_K, "v", il, "weight")); @@ -1084,24 +1264,255 @@ bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length return true; } -// normalize: x = (x - mean) / std -// TODO: implement bicubic interpolation instead of linear. -bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32 * res, const bool pad2square) { +// Linear interpolation between two points +inline float lerp(float s, float e, float t) { + return s + (e - s) * t; +} +// Bilinear resize function +static void bilinear_resize(const clip_image_u8& src, clip_image_u8& dst, int target_width, int target_height) { + dst.nx = target_width; + dst.ny = target_height; + dst.buf.resize(3 * target_width * target_height); + + float x_ratio = static_cast(src.nx - 1) / target_width; + float y_ratio = static_cast(src.ny - 1) / target_height; + + for (int y = 0; y < target_height; y++) { + for (int x = 0; x < target_width; x++) { + float px = x_ratio * x; + float py = y_ratio * y; + int x_floor = static_cast(px); + int y_floor = static_cast(py); + float x_lerp = px - x_floor; + float y_lerp = py - y_floor; + + for (int c = 0; c < 3; c++) { + float top = lerp( + static_cast(src.buf[3 * (y_floor * src.nx + x_floor) + c]), + static_cast(src.buf[3 * (y_floor * src.nx + (x_floor + 1)) + c]), + x_lerp + ); + float bottom = lerp( + static_cast(src.buf[3 * ((y_floor + 1) * src.nx + x_floor) + c]), + static_cast(src.buf[3 * ((y_floor + 1) * src.nx + (x_floor + 1)) + c]), + x_lerp + ); + dst.buf[3 * (y * target_width + x) + c] = static_cast(lerp(top, bottom, y_lerp)); + } + } + } +} + +// Normalize image to float32 - careful with pytorch .to(model.device, dtype=torch.float16) - this sometimes reduces precision (32>16>32), sometimes not +static void normalize_image_u8_to_f32(const clip_image_u8* src, clip_image_f32* dst, const float mean[3], const float std[3]) { + dst->nx = src->nx; + dst->ny = src->ny; + dst->buf.resize(src->buf.size()); + + for (size_t i = 0; i < src->buf.size(); ++i) { + int c = i % 3; // rgb + dst->buf[i] = (static_cast(src->buf[i]) / 255.0f - mean[c]) / std[c]; + } +} + +inline float clip(float x, float lower, float upper) { + return std::max(lower, std::min(x, upper)); +} + +static bool bicubic_resize(const clip_image_u8 &img, clip_image_u8 &dst, int target_width, int target_height) { + const int nx = img.nx; + const int ny = img.ny; + + dst.nx = target_width; + dst.ny = target_height; + dst.buf.resize(3 * target_width * target_height); + + float Cc; + float C[5]; + float d0, d2, d3, a0, a1, a2, a3; + int i, j, k, jj; + int x, y; + float dx, dy; + float tx, ty; + + tx = (float)nx / (float)target_width; + ty = (float)ny / (float)target_height; + + // Bicubic interpolation; adapted from ViT.cpp, inspired from : + // -> https://github.com/yglukhov/bicubic-interpolation-image-processing/blob/master/libimage.c#L36 + // -> https://en.wikipedia.org/wiki/Bicubic_interpolation + + for (i = 0; i < target_height; i++) { + for (j = 0; j < target_width; j++) { + x = (int)(tx * j); + y = (int)(ty * i); + + dx = tx * j - x; + dy = ty * i - y; + + for (k = 0; k < 3; k++) { + for (jj = 0; jj <= 3; jj++) { + d0 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x - 1, 0, nx - 1)) * 3 + k] - img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k]; + d2 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x + 1, 0, nx - 1)) * 3 + k] - img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k]; + d3 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x + 2, 0, nx - 1)) * 3 + k] - img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k]; + a0 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k]; + + a1 = -1.0 / 3 * d0 + d2 - 1.0 / 6 * d3; + a2 = 1.0 / 2 * d0 + 1.0 / 2 * d2; + a3 = -1.0 / 6 * d0 - 1.0 / 2 * d2 + 1.0 / 6 * d3; + + C[jj] = a0 + a1 * dx + a2 * dx * dx + a3 * dx * dx * dx; + + d0 = C[0] - C[1]; + d2 = C[2] - C[1]; + d3 = C[3] - C[1]; + a0 = C[1]; + a1 = -1.0 / 3 * d0 + d2 - 1.0 / 6 * d3; + a2 = 1.0 / 2 * d0 + 1.0 / 2 * d2; + a3 = -1.0 / 6 * d0 - 1.0 / 2 * d2 + 1.0 / 6 * d3; + Cc = a0 + a1 * dy + a2 * dy * dy + a3 * dy * dy * dy; + + const uint8_t Cc2 = std::min(std::max(std::round(Cc), 0.0f), 255.0f); + dst.buf[(i * target_width + j) * 3 + k] = float(Cc2); + } + } + } + } + + return true; +} + +// llava-1.6 type of resize_and_pad (black) +static void resize_and_pad_image(const clip_image_u8& image, clip_image_u8 &image_output, const std::pair& target_resolution) { + int target_width = target_resolution.first; + int target_height = target_resolution.second; + + float scale_w = static_cast(target_width) / image.nx; + float scale_h = static_cast(target_height) / image.ny; + + int new_width, new_height; + + if (scale_w < scale_h) { + new_width = target_width; + new_height = std::min(static_cast(std::ceil(image.ny * scale_w)), target_height); + } else { + new_height = target_height; + new_width = std::min(static_cast(std::ceil(image.nx * scale_h)), target_width); + } + + clip_image_u8 resized_image; + // bilinear_resize(image, resized_image, new_width, new_height); + bicubic_resize(image, resized_image, new_width, new_height); + + clip_image_u8 padded_image; + padded_image.nx = target_width; + padded_image.ny = target_height; + padded_image.buf.resize(3 * target_width * target_height, 0); // Initialize with black + + // Calculate padding offsets + int pad_x = (target_width - new_width) / 2; + int pad_y = (target_height - new_height) / 2; + + // Copy the resized image into the center of the padded buffer + for (int y = 0; y < new_height; ++y) { + for (int x = 0; x < new_width; ++x) { + for (int c = 0; c < 3; ++c) { + padded_image.buf[3 * ((y + pad_y) * target_width + (x + pad_x)) + c] = resized_image.buf[3 * (y * new_width + x) + c]; + } + } + } + image_output = std::move(padded_image); +} + +/** + * Selects the best resolution from a list of possible resolutions based on the original size. + * + * @param original_size The original size of the image in the format (width, height). + * @param possible_resolutions A list of possible resolutions in the format [(width1, height1), (width2, height2), ...]. + * @return The best fit resolution in the format (width, height). + */ +static std::pair select_best_resolution(const std::pair & original_size, const std::vector> & possible_resolutions) { + int original_width = original_size.first; + int original_height = original_size.second; + std::pair best_fit; + int max_effective_resolution = 0; + int min_wasted_resolution = std::numeric_limits::max(); + + for (const auto& resolution : possible_resolutions) { + int width = resolution.first; + int height = resolution.second; + float scale = std::min(static_cast(width) / original_width, static_cast(height) / original_height); + int downscaled_width = static_cast(original_width * scale); + int downscaled_height = static_cast(original_height * scale); + int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height); + int wasted_resolution = (width * height) - effective_resolution; + // fprintf(stderr, "resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution); + if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) { + max_effective_resolution = effective_resolution; + min_wasted_resolution = wasted_resolution; + best_fit = resolution; + } + } + + return best_fit; +} + +static std::vector divide_to_patches_u8(const clip_image_u8 & image, int patch_size) { + std::vector patches; + int width = image.nx; + int height = image.ny; + for (int i = 0; i < height; i += patch_size) { + for (int j = 0; j < width; j += patch_size) { + clip_image_u8 *patch = clip_image_u8_init(); + patch->nx = std::min(patch_size, width - j); + patch->ny = std::min(patch_size, height - i); + patch->buf.resize(3 * patch->nx * patch->ny); + for (int y = 0; y < patch->ny; ++y) { + for (int x = 0; x < patch->nx; ++x) { + for (int c = 0; c < 3; ++c) { + patch->buf[3 * (y * patch->nx + x) + c] = image.buf[3 * ((i + y) * width + (j + x)) + c]; + } + } + } + patches.push_back(patch); + } + } + return patches; +} + +// returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector +// res_imgs memory is being allocated here, previous allocations will be freed if found +bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch & res_imgs) { + bool pad_to_square = true; if (!ctx->has_vision_encoder) { printf("This gguf file seems to have no vision encoder\n"); return false; } + auto & params = ctx->vision_model.hparams; + // The model config actually contains all we need to decide on how to preprocess, here we automatically switch to the new llava-1.6 preprocessing + if (strcmp(params.mm_patch_merge_type, "spatial_unpad") == 0) { + pad_to_square = false; + } + // free the previous res_imgs if any set + if (res_imgs.size > 0 && res_imgs.size < 100) { + for (size_t i = 0; i < res_imgs.size; i++) { + clip_image_f32_free(&(res_imgs.data[i])); + } + delete[] res_imgs.data; + } + res_imgs.data = nullptr; + res_imgs.size = 0; // the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104) // see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156 clip_image_u8 * temp = clip_image_u8_init(); // we will keep the input image data here temporarily - if (pad2square && img->nx != img->ny) { + if (pad_to_square && img->nx != img->ny) { int longer_side = std::max(img->nx, img->ny); temp->nx = longer_side; temp->ny = longer_side; temp->buf.resize(3 * longer_side * longer_side); - const uint8_t bc[3] = {122, 116, 104}; // background color in RGB from LLaVA + const uint8_t bc[3] = {122, 116, 104}; // background color in RGB from LLaVA (this is the mean rgb color * 255) // fill with background color for (size_t i = 0; i < temp->buf.size(); i++) { @@ -1119,18 +1530,63 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli } } } else { - temp->nx = img->nx; - temp->ny = img->ny; - temp->buf.resize(img->buf.size()); - memcpy(temp->buf.data(), img->buf.data(), temp->buf.size()); + if (params.image_grid_pinpoints[0] != 0) { + // "spatial_unpad" with "anyres" processing for llava-1.6 + std::vector> possible_resolutions; + for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) { + possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]}); + } + std::pair best_resolution = select_best_resolution({img->nx, img->ny}, possible_resolutions); + // clip_image_save_to_bmp(*img, "input.bmp"); + resize_and_pad_image(*img, *temp, best_resolution); // we do not pad with mean-bg color anymore in llava-1.6 + // clip_image_save_to_bmp(*temp, "resized.bmp"); + // visually verify normalized image: + // normalize_image_u8_to_f32(*temp, *res, ctx->image_mean, ctx->image_std); + // { + // clip_image_u8 * temp2 = clip_image_u8_init(); + // clip_image_convert_f32_to_u8(*res, *temp2); + // clip_image_save_to_bmp(*temp2, "resized_normalized_f32.bmp"); + // clip_image_u8_free(temp2); + // } + + std::vector patches = divide_to_patches_u8(*temp, params.image_size); // prepare spatial sorted main patches of image_size each (336 in llava-1.6) + + clip_image_u8 *image_original_resize = clip_image_u8_init(); + // bilinear_resize(*img, *image_original_resize, params.image_size, params.image_size); // in python this is "shortest_edge", but all CLIP are square + bicubic_resize(*img, *image_original_resize, params.image_size, params.image_size); // in python this is "shortest_edge", but all CLIP are square + patches.insert(patches.begin(), image_original_resize); + // clip_image_f32_batch_init(patches.size()); + res_imgs.size = patches.size(); + res_imgs.data = new clip_image_f32[res_imgs.size]; + int num=0; + for (auto& patch : patches) { + normalize_image_u8_to_f32(patch, &res_imgs.data[num], ctx->image_mean, ctx->image_std); + num++; + } + + for (size_t i = 0; i < patches.size(); i++) { + // printf("patch %d: %d %d\n", i, patches[i]->nx, patches[i]->ny); + clip_image_u8_free(patches[i]); + } + + clip_image_u8_free(temp); + + return true; + } else { + temp->nx = img->nx; + temp->ny = img->ny; + temp->buf.resize(img->buf.size()); + memcpy(temp->buf.data(), img->buf.data(), temp->buf.size()); + } } const int nx = temp->nx; const int ny = temp->ny; + // clip_image_save_to_bmp(*temp, "resized_vanilla.bmp"); const int nx2 = ctx->vision_model.hparams.image_size; const int ny2 = ctx->vision_model.hparams.image_size; - + clip_image_f32 * res = clip_image_f32_init(); res->nx = nx2; res->ny = ny2; res->buf.resize(3 * nx2 * ny2); @@ -1184,9 +1640,25 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli } clip_image_u8_free(temp); + // { + // clip_image_u8 * temp2 = clip_image_u8_init(); + // clip_image_convert_f32_to_u8(*res, *temp2); + // clip_image_save_to_bmp(*temp2, "resized_normalized_f32_vanilla.bmp"); + // clip_image_u8_free(temp2); + // } + // res_imgs.push_back(res); + + res_imgs.size = 1; + res_imgs.data = new clip_image_f32[res_imgs.size]; + res_imgs.data[0] = std::move(*res); + return true; } +ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx) { + return ctx->vision_model.image_newline; +} + void clip_free(clip_ctx * ctx) { ggml_free(ctx->ctx_data); gguf_free(ctx->ctx_gguf); @@ -1194,6 +1666,42 @@ void clip_free(clip_ctx * ctx) { delete ctx; } +size_t clip_embd_nbytes(const struct clip_ctx * ctx) { + return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float); +} + +int32_t clip_image_size(const struct clip_ctx * ctx) { + return ctx->vision_model.hparams.image_size; +} + +int32_t clip_patch_size(const struct clip_ctx * ctx) { + return ctx->vision_model.hparams.patch_size; +} + +int32_t clip_hidden_size(const struct clip_ctx * ctx) { + return ctx->vision_model.hparams.hidden_size; +} + +const char * clip_patch_merge_type(const struct clip_ctx * ctx) { + return ctx->vision_model.hparams.mm_patch_merge_type; +} + +const int32_t * clip_image_grid(const struct clip_ctx * ctx) { + return ctx->vision_model.hparams.image_grid_pinpoints; +} + +int clip_n_patches(const struct clip_ctx * ctx) { + const auto & params = ctx->vision_model.hparams; + + int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size); + + if (ctx->proj_type == PROJECTOR_TYPE_LDP) { + n_patches /= 4; + } + + return n_patches; +} + bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) { if (!ctx->has_vision_encoder) { printf("This gguf file seems to have no vision encoder\n"); @@ -1213,7 +1721,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima } int batch_size = imgs->size; - if(ctx->has_llava_projector) { + if (ctx->has_llava_projector) { GGML_ASSERT(batch_size == 1); // TODO: support multiple images } @@ -1224,9 +1732,10 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima // set inputs const auto & model = ctx->vision_model; const auto & hparams = model.hparams; - const int image_size = hparams.image_size; - const int patch_size = hparams.patch_size; - const int num_patches = ((image_size / patch_size) * (image_size / patch_size)); + + const int image_size = hparams.image_size; + const int patch_size = hparams.patch_size; + const int num_patches = ((image_size / patch_size) * (image_size / patch_size)); const int num_positions = num_patches + 1; { @@ -1301,11 +1810,11 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima // copy the embeddings to the location passed by the user ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings)); + return true; } bool clip_model_quantize(const char * fname_inp, const char * fname_out, const int itype) { - ggml_type type = GGML_TYPE_Q4_1; assert(itype < GGML_TYPE_COUNT); @@ -1494,26 +2003,13 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) { if (ctx->proj_type == PROJECTOR_TYPE_LDP) { return ctx->vision_model.mm_model_block_1_block_2_1_b->ne[0]; } - else if (ctx->proj_type == PROJECTOR_TYPE_MLP) { + if (ctx->proj_type == PROJECTOR_TYPE_MLP) { return ctx->vision_model.mm_2_b->ne[0]; - } else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) { + } + if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) { return ctx->vision_model.mm_3_b->ne[0]; } - else { - std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type]; - throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str())); - } -} -int clip_n_patches(const struct clip_ctx * ctx) { - auto & params = ctx->vision_model.hparams; - int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size); - if (ctx->proj_type == PROJECTOR_TYPE_LDP) { - n_patches /= 4; - } - return n_patches; -} - -size_t clip_embd_nbytes(const struct clip_ctx * ctx) { - return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float); + std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type]; + throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str())); } diff --git a/examples/llava/clip.h b/examples/llava/clip.h index 458a256a1..cd9a4022f 100644 --- a/examples/llava/clip.h +++ b/examples/llava/clip.h @@ -24,25 +24,7 @@ struct clip_ctx; extern "C" { #endif -struct clip_vision_hparams { - int32_t image_size; - int32_t patch_size; - int32_t hidden_size; - int32_t n_intermediate; - int32_t projection_dim; - int32_t n_head; - int32_t n_layer; - float eps; -}; - -CLIP_API struct clip_ctx * clip_model_load(const char * fname, int verbosity); - -CLIP_API void clip_free(struct clip_ctx * ctx); - -CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx); - -CLIP_API int clip_n_patches (const struct clip_ctx * ctx); -CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx); +struct clip_ctx; struct clip_image_u8_batch { struct clip_image_u8 * data; @@ -54,10 +36,29 @@ struct clip_image_f32_batch { size_t size; }; +CLIP_API struct clip_ctx * clip_model_load (const char * fname, int verbosity); +CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity); + +CLIP_API void clip_free(struct clip_ctx * ctx); + +CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx); + +CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx); +CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx); +CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx); + +// TODO: should be enum, not string +CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx); + +CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx); + +CLIP_API int clip_n_patches (const struct clip_ctx * ctx); +CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx); + CLIP_API struct clip_image_u8 * clip_image_u8_init (); CLIP_API struct clip_image_f32 * clip_image_f32_init(); -CLIP_API void clip_image_u8_free (struct clip_image_u8 * img); +CLIP_API void clip_image_u8_free (struct clip_image_u8 * img); CLIP_API void clip_image_f32_free(struct clip_image_f32 * img); CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img); @@ -65,7 +66,11 @@ CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 /** interpret bytes as an image file with length bytes_length, and use the result to populate img */ CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img); -CLIP_API bool clip_image_preprocess (struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32 * res, bool pad2square); +/** preprocess img and store the result in res_imgs, pad_to_square may be overriden to false depending on model configuration */ +CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch & res_imgs ); + +CLIP_API struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx); + CLIP_API bool clip_image_encode (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec); CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec); diff --git a/examples/llava/convert-image-encoder-to-gguf.py b/examples/llava/convert-image-encoder-to-gguf.py index e204b56be..c69f89ac2 100644 --- a/examples/llava/convert-image-encoder-to-gguf.py +++ b/examples/llava/convert-image-encoder-to-gguf.py @@ -78,18 +78,19 @@ ap.add_argument("--text-only", action="store_true", required=False, help="Save a text-only model. It can't be used to encode images") ap.add_argument("--vision-only", action="store_true", required=False, help="Save a vision-only model. It can't be used to encode texts") -ap.add_argument("--clip_model_is_vision", action="store_true", required=False, +ap.add_argument("--clip-model-is-vision", action="store_true", required=False, help="The clip model is a pure vision model (ShareGPT4V vision extract for example)") +ap.add_argument("--clip-model-is-openclip", action="store_true", required=False, + help="The clip model is from openclip (for ViT-SO400M type))") ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.") ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp", choices=["mlp", "ldp"], default="mlp") -ap.add_argument("--image-mean", nargs=3, type=float, required=False, help="Override image mean values") -ap.add_argument("--image-std", nargs=3, type=float, required=False, help="Override image std values") ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None) # Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711 +# Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5 default_image_mean = [0.48145466, 0.4578275, 0.40821073] default_image_std = [0.26862954, 0.26130258, 0.27577711] -ap.add_argument('--image_mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None) -ap.add_argument('--image_std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None) +ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None) +ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None) # with proper args = ap.parse_args() @@ -105,7 +106,7 @@ if args.use_f32: # output in the same directory as the model if output_dir is None dir_model = args.model_dir -if args.clip_model_is_vision: +if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip: vocab = None tokens = None else: @@ -133,7 +134,7 @@ ftype = 1 if args.use_f32: ftype = 0 -if args.clip_model_is_vision: +if args.clip_model_is_vision or args.clip_model_is_openclip: model = CLIPVisionModel.from_pretrained(dir_model) processor = None else: @@ -202,6 +203,57 @@ if has_vision_encoder: fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), v_hparams["layer_norm_eps"]) block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"] fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), block_count) + # /** + # "image_grid_pinpoints": [ + # [ + # 336, + # 672 + # ], + # [ + # 672, + # 336 + # ], + # [ + # 672, + # 672 + # ], + # [ + # 1008, + # 336 + # ], + # [ + # 336, + # 1008 + # ] + # ], + # Flattened: + # [ + # 336, 672, + # 672, 336, + # 672, 672, + # 1008, 336, + # 336, 1008 + # ] + # * + # */ + if "image_grid_pinpoints" in v_hparams: + # flatten it + image_grid_pinpoints = [] + for pinpoint in v_hparams["image_grid_pinpoints"]: + for p in pinpoint: + image_grid_pinpoints.append(p) + fout.add_array("clip.vision.image_grid_pinpoints", image_grid_pinpoints) + if "image_crop_resolution" in v_hparams: + fout.add_uint32("clip.vision.image_crop_resolution", v_hparams["image_crop_resolution"]) + if "image_aspect_ratio" in v_hparams: + fout.add_string("clip.vision.image_aspect_ratio", v_hparams["image_aspect_ratio"]) + if "image_split_resolution" in v_hparams: + fout.add_uint32("clip.vision.image_split_resolution", v_hparams["image_split_resolution"]) + if "mm_patch_merge_type" in v_hparams: + fout.add_string("clip.vision.mm_patch_merge_type", v_hparams["mm_patch_merge_type"]) + if "mm_projector_type" in v_hparams: + fout.add_string("clip.vision.mm_projector_type", v_hparams["mm_projector_type"]) + if processor is not None: image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean diff --git a/examples/llava/llava-cli.cpp b/examples/llava/llava-cli.cpp index 031e9806d..bef7f7c95 100644 --- a/examples/llava/llava-cli.cpp +++ b/examples/llava/llava-cli.cpp @@ -155,11 +155,29 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_ system_prompt = prompt.substr(0, image_pos); user_prompt = prompt.substr(image_pos + std::string("").length()); printf("system_prompt: %s\n", system_prompt.c_str()); + if (params->verbose_prompt) { + auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, system_prompt, true, true); + for (int i = 0; i < (int) tmp.size(); i++) { + printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str()); + } + } printf("user_prompt: %s\n", user_prompt.c_str()); + if (params->verbose_prompt) { + auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true); + for (int i = 0; i < (int) tmp.size(); i++) { + printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str()); + } + } } else { // llava-1.5 native mode system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:"; user_prompt = prompt + "\nASSISTANT:"; + if (params->verbose_prompt) { + auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true); + for (int i = 0; i < (int) tmp.size(); i++) { + printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str()); + } + } } eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, add_bos); @@ -171,13 +189,17 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_ fprintf(stderr, "\n"); struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams); - + std::string response = ""; for (int i = 0; i < max_tgt_len; i++) { const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past); + response += tmp; if (strcmp(tmp, "") == 0) break; if (strstr(tmp, "###")) break; // Yi-VL behavior - printf("%s", tmp); + if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works) + if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6 + if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6 + fflush(stdout); } diff --git a/examples/llava/llava-surgery-v2.py b/examples/llava/llava-surgery-v2.py new file mode 100644 index 000000000..5bc5bc513 --- /dev/null +++ b/examples/llava/llava-surgery-v2.py @@ -0,0 +1,167 @@ +import argparse +import glob +import os +import torch +from safetensors.torch import load as safe_load, save as safe_save, safe_open, save_file + +# Function to determine if file is a SafeTensor file +def is_safetensor_file(file_path): + return file_path.endswith('.safetensors') + + +# Unified loading function +def load_model(file_path): + if is_safetensor_file(file_path): + tensors = {} + with safe_open(file_path, framework="pt", device="cpu") as f: + for key in f.keys(): + tensors[key] = f.get_tensor(key).clone() + # output shape + print(f"{key} : {tensors[key].shape}") + return tensors, 'safetensor' + else: + return torch.load(file_path, map_location=torch.device('cpu')), 'pytorch' + + +# Unified saving function +def save_model(model, file_path, file_type): + if file_type == 'safetensor': + # safe_save(model, file_path) + save_file(model, file_path) + else: + torch.save(model, file_path) + + +# Adapted function to clean vision tower from checkpoint +def clean_vision_tower_from_checkpoint(checkpoint_path): + checkpoint, file_type = load_model(checkpoint_path) + # file_type = 'pytorch' + model_path = os.path.dirname(checkpoint_path) + print(f"Searching for vision tower tensors in {checkpoint_path}") + clip_tensors = [k for k, v in checkpoint.items() if (k.startswith("model.vision_tower") or k.startswith("vit."))] + + if len(clip_tensors) > 0: + print(f"Found {len(clip_tensors)} tensors to extract from {checkpoint_path}") + # Adapted for file type + clip_path = os.path.join(model_path, "llava.clip") + + if os.path.exists(clip_path): + print(f"Loading existing llava.clip from {clip_path}") + existing_clip, _ = load_model(clip_path) + else: + print(f"Creating new llava.clip at {clip_path}") + existing_clip = {} + # Update existing_clip with new tensors, avoid duplicates + for name in clip_tensors: + simple_name = name[name.index('vision_model.'):] if 'vision_model.' in name else name + print(f"Adding {simple_name} to llava.clip") + if simple_name not in existing_clip: + existing_clip[simple_name] = checkpoint[name] + + # Save the updated clip tensors back to llava.clip + save_model(existing_clip, clip_path, 'pytorch') + + # Remove the tensors from the original checkpoint + for name in clip_tensors: + del checkpoint[name] + + # Save the updated checkpoint + checkpoint_path = checkpoint_path + save_model(checkpoint, checkpoint_path, file_type) + return True + return False + +def find_relevant_checkpoints(checkpoint_paths, newline_criteria, projector): + newline_checkpoint_path = None + projector_checkpoint_path = None + + for path in checkpoint_paths: + checkpoint, _ = load_model(path) + if newline_criteria(checkpoint) and newline_checkpoint_path is None: + newline_checkpoint_path = path + if projector(checkpoint): + projector_checkpoint_path = path + + return newline_checkpoint_path, projector_checkpoint_path + +def newline_criteria(checkpoint): + return any(k.startswith("model.image_newline") for k in checkpoint.keys()) + +def proj_criteria(checkpoint): + return any(k.startswith("model.mm_projector") or k.startswith("vision_proj.") for k in checkpoint.keys()) + + +# Command-line interface setup +ap = argparse.ArgumentParser() +ap.add_argument("-m", "--model", required=True, help="Path to LLaVA v1.5+ model") +ap.add_argument("-C", "--clean-vision-tower", action="store_true", help="Remove any vision tower from the model files") +args = ap.parse_args() + +if args.clean_vision_tower: + # Generalized to handle both PyTorch and SafeTensors models + model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True) + # checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and path.startswith('pytorch')) or (path.endswith('.safetensors') and path.startswith('model'))] + checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])] + for projector_checkpoint_path in checkpoint_paths: + print(f"Cleaning {projector_checkpoint_path}") + if not clean_vision_tower_from_checkpoint(projector_checkpoint_path): + print(f"No vision tower found in {projector_checkpoint_path}") + # we break once none is found, so far all models append them at the end + # break + print("Done! All vision tower tensors are removed from the model files and stored in llava.clip file.") + +# Now we look for the projector in the last checkpoint +model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True) +checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])] +# last_checkpoint_path = checkpoint_paths[0] +# first_checkpoint_path = checkpoint_paths[-1] +newline_checkpoint_path, projector_checkpoint_path = find_relevant_checkpoints(checkpoint_paths, newline_criteria, proj_criteria) + +print(f"Taking projector from {projector_checkpoint_path}") +first_mm_tensors = [] +first_checkpoint = None +if newline_checkpoint_path is not None: + print(f"Taking newline from {newline_checkpoint_path}") + first_checkpoint, file_type = load_model(newline_checkpoint_path) + first_mm_tensors = [k for k, v in first_checkpoint.items() if k.startswith("model.image_newline")] + +# Load the checkpoint +mm_tensors = [] +last_checkpoint = None +if projector_checkpoint_path is not None: + last_checkpoint, file_type = load_model(projector_checkpoint_path) + mm_tensors = [k for k, v in last_checkpoint.items() if k.startswith("model.mm_projector") or k.startswith("vision_proj.")] + +if len(mm_tensors) == 0: + if last_checkpoint is not None: + for k, v in last_checkpoint.items(): + print(k) + print(f"Found {len(mm_tensors)} tensors to extract out of {len(last_checkpoint)} tensors.") + print("No tensors found. Is this a LLaVA model?") + exit() + +print(f"Found {len(mm_tensors)} tensors to extract.") +print(f"Found additional {len(first_mm_tensors)} tensors to extract.") +# projector = {name: checkpoint.[name].float() for name in mm_tensors} +projector = {} +for name in mm_tensors: + projector[name] = last_checkpoint[name].float() +for name in first_mm_tensors: + projector[name] = first_checkpoint[name].float() + +if len(projector) > 0: + save_model(projector, f"{args.model}/llava.projector", 'pytorch') + +for name in mm_tensors: + del last_checkpoint[name] +for name in first_mm_tensors: + del first_checkpoint[name] + +if len(mm_tensors) > 0: + save_model(last_checkpoint, projector_checkpoint_path, file_type) +if len(first_mm_tensors) > 0: + save_model(first_checkpoint, newline_checkpoint_path, file_type) + +print("Done!") +print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.") +print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.") diff --git a/examples/llava/llava.cpp b/examples/llava/llava.cpp index d42e7582e..22953417f 100644 --- a/examples/llava/llava.cpp +++ b/examples/llava/llava.cpp @@ -2,32 +2,296 @@ #include "common.h" #include "llama.h" #include "llava.h" +#include "base64.hpp" #include #include #include +#include + +// RGB uint8 image +struct clip_image_u8 { + int nx; + int ny; + + std::vector buf; +}; + +// RGB float32 image (NHWC) +// Memory layout: RGBRGBRGB... +struct clip_image_f32 { + int nx; + int ny; + + std::vector buf; +}; + +struct clip_image_grid_shape { + int first; + int second; +}; + +/** + * Selects the best resolution from a list of possible resolutions based on the original size. + * + * @param original_size The original size of the image in the format (width, height). + * @param possible_resolutions A list of possible resolutions in the format [(width1, height1), (width2, height2), ...]. + * @return The best fit resolution in the format (width, height). + */ +static std::pair select_best_resolution(const std::pair& original_size, const std::vector>& possible_resolutions) { + int original_width = original_size.first; + int original_height = original_size.second; + + std::pair best_fit; + int max_effective_resolution = 0; + int min_wasted_resolution = std::numeric_limits::max(); + + for (const auto& resolution : possible_resolutions) { + int width = resolution.first; + int height = resolution.second; + float scale = std::min(static_cast(width) / original_width, static_cast(height) / original_height); + int downscaled_width = static_cast(original_width * scale); + int downscaled_height = static_cast(original_height * scale); + int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height); + int wasted_resolution = (width * height) - effective_resolution; + // fprintf(stderr, "resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution); + if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) { + max_effective_resolution = effective_resolution; + min_wasted_resolution = wasted_resolution; + best_fit = resolution; + } + } + + return best_fit; +} + +/** + * @brief Get the anyres image grid shape object + * + * @param image_size + * @param grid_pinpoints + * @param image_patch_size + * @return + */ +static struct clip_image_grid_shape get_anyres_image_grid_shape(const std::pair & image_size, const std::vector> & grid_pinpoints, int image_patch_size) { + /** + Conversion from gguf flat array to vector: + std::vector> possible_resolutions; + for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) { + possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]}); + } + */ + auto best_resolution = select_best_resolution(image_size, grid_pinpoints); + return {best_resolution.first / image_patch_size, best_resolution.second / image_patch_size}; +} + +// Take the image segments in a grid configuration and return the embeddings and the number of embeddings into preallocated memory (image_embd_out) +static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out) { + struct { + struct ggml_tensor * newline; + struct ggml_context * ctx; + } model; + + const int32_t image_size = clip_image_size(ctx_clip); + const int32_t patch_size = clip_patch_size(ctx_clip); + + int32_t num_patches_per_side = image_size / patch_size; // 336 / 14 = 24 - used for embedding-patching boxes (24*24 = 576 patches) + + int num_patches_width = grid_shape.first; // grid 1-4 + int num_patches_height = grid_shape.second; // grid 1-4 + + const size_t num_images = num_patches_width + num_patches_height + 1; + + // TODO: size calculation is not calculated - it's only tens of MB + size_t ctx_size = 0; + + { + ctx_size += clip_embd_nbytes(ctx_clip) * num_images * 8; // image_features + ctx_size += 1024*1024 * ggml_type_size(GGML_TYPE_F32); + } + + struct ggml_init_params params { + /*.mem_size =*/ ctx_size, + /*.mem_buffer =*/ NULL, + /*.no_alloc =*/ false, // NOTE: this should be false when using the legacy API + }; + + // Python reference code for full unpad: + /* + base_image_feature = image_feature[0] + image_feature = image_feature[1:] + image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous() + image_feature = image_feature.flatten(1, 2).flatten(2, 3) + image_feature = unpad_image(image_feature, image_sizes[image_idx]) + image_feature = torch.cat(( + image_feature, + self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1) + ), dim=-1) + image_feature = image_feature.flatten(1, 2).transpose(0, 1) + image_feature = torch.cat((base_image_feature, image_feature), dim=0) + */ + // We now have two options: unpad or no unpad. Unpad removes tokens for faster llm eval. + // In terms of result quality it appears to make no difference, so we'll start with the easier approach given 5D tensors are not supported in ggml yet. + // Without unpad we have to split the sub-image embeddings into patches of 24 features each and permute them. + // Once all images are processed to prepended the base_image_features without any changes. + + // Pytorch reference simplified, modified for ggml compatibility - confirmed identical output in python (for a 2x2 grid image (676x676 scaling)) + /* + image_feature = image_feature.view(2, 2, 24, 24, 4096) + image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous() + image_feature = image_feature.view(2, 24, 2, 24, 4096) + image_feature = image_feature.flatten(0, 3) + + // Reshape to 4D tensor by merging the last two dimensions + image_feature = image_feature.view(2, 2, 24, 24*4096) + image_feature = image_feature.permute(0, 2, 1, 3).contiguous() + image_feature = image_feature.view(-1, 4096) + */ + + model.ctx = ggml_init(params); + + ggml_tensor * newline_tmp = clip_get_newline_tensor(ctx_clip); + model.newline = ggml_new_tensor_1d(model.ctx, GGML_TYPE_F32, newline_tmp->ne[0]); + if (newline_tmp->backend != GGML_BACKEND_CPU) { + if (newline_tmp->buffer == NULL) { + printf("newline_tmp tensor buffer is NULL\n"); + } + ggml_backend_tensor_get(newline_tmp, model.newline->data, 0, ggml_nbytes(newline_tmp)); + } else { + model.newline->data = newline_tmp->data; + if (model.newline->data == NULL) { + printf("newline_tmp tensor data is NULL\n"); + } + } + + struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_patches(ctx_clip), num_images - 1); // example: 4096 x 576 x 4 + // ggml_tensor_printf(image_features,"image_features",__LINE__,false,false); + // fill it with the image embeddings, ignoring the base + for (size_t i = 1; i < num_images; i++) { + size_t offset = (i-1) * clip_embd_nbytes(ctx_clip); + memcpy((uint8_t *)(image_features->data) + offset, image_embd_v[i], clip_embd_nbytes(ctx_clip)); + } + + struct ggml_cgraph * gf = ggml_new_graph(model.ctx); + size_t size_ele = ggml_type_size(GGML_TYPE_F32); + + struct ggml_tensor *image_features_patchview = ggml_view_4d(model.ctx, image_features, + num_patches_per_side * clip_n_mmproj_embd(ctx_clip), + num_patches_per_side, + num_patches_width, + num_patches_height, + size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip), + size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side, + size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side * num_patches_width, 0); + // ggml_tensor_printf(image_features_patchview,"image_features_patchview",__LINE__,false,false); + struct ggml_tensor *permuted_cont = ggml_cont(model.ctx, ggml_permute(model.ctx, image_features_patchview, 0, 2, 1, 3)); + /** + At the end of each row we have to add the row_end embeddings, which are the same as the newline embeddings + image_feature = torch.cat(( + image_feature, + self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device) + ), dim=-1) + * + */ + + // ggml_tensor_printf(permuted_cont,"permuted_cont",__LINE__,false,false); + struct ggml_tensor *flatten = ggml_view_2d(model.ctx, permuted_cont, clip_n_mmproj_embd(ctx_clip), num_patches_height * num_patches_width * num_patches_per_side * num_patches_per_side, size_ele * clip_n_mmproj_embd(ctx_clip), 0); + // ggml_tensor_printf(flatten,"flatten",__LINE__,false,false); + ggml_build_forward_expand(gf, flatten); + ggml_graph_compute_with_ctx(model.ctx, gf, 1); + struct ggml_tensor* result = gf->nodes[gf->n_nodes - 1]; + + memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context + // append without newline tokens (default behavior in llava_arch when not using unpad ): + memcpy(image_embd_out + clip_n_patches(ctx_clip) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches + *n_img_pos_out = static_cast(result->ne[1]+clip_n_patches(ctx_clip)); + + // Debug: Test single segments + // Current findings: sending base image, sending a segment embedding all works similar to python + // However, permuted embeddings do not work yet (stride issue?) + // memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as context + // memcpy(image_embd_out, (float*)prepared_cont->data, clip_embd_nbytes(ctx_clip)); // main image as context + // *n_img_pos_out=576; + + ggml_free(model.ctx); + return true; +} -#include "base64.hpp" static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) { - clip_image_f32 * img_res = clip_image_f32_init(); - if (!clip_image_preprocess(ctx_clip, img, img_res, /*pad2square =*/ true)) { + // std::vector img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336 + clip_image_f32_batch img_res_v; + img_res_v.size = 0; + img_res_v.data = nullptr; + if (!clip_image_preprocess(ctx_clip, img, img_res_v)) { fprintf(stderr, "%s: unable to preprocess image\n", __func__); - clip_image_f32_free(img_res); + delete[] img_res_v.data; return false; } - *n_img_pos = clip_n_patches(ctx_clip); - const int64_t t_img_enc_start_us = ggml_time_us(); - bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd); - clip_image_f32_free(img_res); - if (!encoded) { - fprintf(stderr, "Unable to encode image\n"); - return false; + const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip); + + if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) { + // flat / default llava-1.5 type embedding + *n_img_pos = clip_n_patches(ctx_clip); + bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096 + delete[] img_res_v.data; + if (!encoded) { + fprintf(stderr, "Unable to encode image\n"); + + return false; + } + } else { + // spatial_unpad llava-1.6 type embedding + // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working + std::vector image_embd_v; + image_embd_v.resize(img_res_v.size); + for (size_t i = 0; i < img_res_v.size; i++) { + image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184 + const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside + if (!encoded) { + fprintf(stderr, "Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size); + return false; + } + } + const int64_t t_img_enc_batch_us = ggml_time_us(); + printf("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0); + + const int32_t * image_grid = clip_image_grid(ctx_clip); + + std::vector> grid_pinpoints; + for (int i = 0; i < 32 && image_grid[i] != 0; i += 2) { + grid_pinpoints.push_back({image_grid[i], image_grid[i+1]}); + } + + // free all img_res_v - not needed anymore + delete[] img_res_v.data; + img_res_v.size = 0; + img_res_v.data = nullptr; + + const int32_t image_size = clip_image_size(ctx_clip); + + struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size); + + int n_img_pos_out; + clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out); + *n_img_pos = n_img_pos_out; + + for (size_t i = 0; i < image_embd_v.size(); i++) { + free(image_embd_v[i]); + } + image_embd_v.clear(); + + // debug image/segment/normalization content: + // clip_image_u8 * tmp = clip_image_u8_init(); + // clip_image_convert_f32_to_u8(*image_feature, *tmp); + // clip_image_save_to_bmp(*tmp, "image_feature.bmp"); } + printf("%s: image embedding created: %d tokens\n", __func__, *n_img_pos); + const int64_t t_img_enc_end_us = ggml_time_us(); float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0; @@ -48,7 +312,7 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * } static bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) { - float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)); + float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*6); // TODO: base on gridsize/llava model if (!image_embd) { fprintf(stderr, "Unable to allocate memory for image embeddings\n"); free(image_embd); @@ -85,7 +349,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_ return true; } -LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length) { +struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length) { clip_image_u8 * img = clip_image_u8_init(); if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) { clip_image_u8_free(img); @@ -142,7 +406,7 @@ static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long return true; } -LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) { +struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) { unsigned char* image_bytes; long image_bytes_length; auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length); @@ -151,13 +415,13 @@ LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct return NULL; } - auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length); + llava_image_embed *embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length); free(image_bytes); return embed; } -LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed) { +void llava_image_embed_free(struct llava_image_embed * embed) { free(embed->embed); free(embed); } diff --git a/examples/llava/llava.h b/examples/llava/llava.h index e08ce7883..9e9466a5d 100644 --- a/examples/llava/llava.h +++ b/examples/llava/llava.h @@ -3,7 +3,6 @@ #include "ggml.h" - #ifdef LLAMA_SHARED # if defined(_WIN32) && !defined(__MINGW32__) # ifdef LLAMA_BUILD @@ -42,7 +41,6 @@ LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed); /** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */ LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past); - #ifdef __cplusplus } #endif diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 1699eb76b..6e3434030 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -968,13 +968,20 @@ struct llama_server_context { continue; } - clip_image_f32 * img_res = clip_image_f32_init(); - if (!clip_image_preprocess(clp_ctx, img.img_data, img_res, /*pad2square =*/ true)) + clip_image_f32_batch img_res_v; + img_res_v.size = 0; + img_res_v.data = nullptr; + if (!clip_image_preprocess(clp_ctx, img.img_data, img_res_v)) { LOG_TEE("Error processing the given image"); clip_free(clp_ctx); + clip_image_f32_free(img_res_v.data); return false; } + + // note: assumes only one image was returned by clip_image_preprocess + clip_image_f32 * img_res = img_res_v.data; + img.image_tokens = clip_n_patches(clp_ctx); img.image_embedding = (float *)malloc(clip_embd_nbytes(clp_ctx)); if (!img.image_embedding) @@ -989,7 +996,9 @@ struct llama_server_context LOG_TEE("Unable to encode image\n"); return false; } - clip_image_f32_free(img_res); + + clip_image_f32_free(img_res_v.data); + img.request_encode_image = false; }