// 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 "log.h"
#include "ggml.h"
#include "ggml-alloc.h"
#include "ggml-backend.h"
#include <cstdlib>
#include <cstring>
#include <fstream>
-#include <iostream>
#include <map>
#include <regex>
#include <stdexcept>
static int get_key_idx(const gguf_context * ctx, const char * key) {
int i = gguf_find_key(ctx, key);
if (i == -1) {
- fprintf(stderr, "key %s not found in file\n", key);
+ LOG_TEE("key %s not found in file\n", key);
throw std::runtime_error(format("Missing required key: %s", key));
}
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",
+ LOG_TEE("%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,
tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], ggml_type_name(tensor->type));
}
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;
+ LOG_TEE("Failed to open file for writing: %s\n", filename.c_str());
return;
}
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;
+ LOG_TEE("Failed to open file for writing: %s\n", filename.c_str());
return;
}
static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs) {
if (!ctx->has_vision_encoder) {
- printf("This gguf file seems to have no vision encoder\n");
+ LOG_TEE("This gguf file seems to have no vision encoder\n");
return nullptr;
}
const int idx_name = gguf_find_key(ctx, KEY_NAME);
if (idx_name != -1) { // make name optional temporarily as some of the uploaded models missing it due to a bug
const std::string name = gguf_get_val_str(ctx, idx_name);
- printf("%s: model name: %s\n", __func__, name.c_str());
+ LOG_TEE("%s: model name: %s\n", __func__, name.c_str());
}
- printf("%s: description: %s\n", __func__, description.c_str());
- printf("%s: GGUF version: %d\n", __func__, gguf_get_version(ctx));
- printf("%s: alignment: %zu\n", __func__, gguf_get_alignment(ctx));
- printf("%s: n_tensors: %d\n", __func__, n_tensors);
- printf("%s: n_kv: %d\n", __func__, n_kv);
- printf("%s: ftype: %s\n", __func__, ftype_str.c_str());
- printf("\n");
+ LOG_TEE("%s: description: %s\n", __func__, description.c_str());
+ LOG_TEE("%s: GGUF version: %d\n", __func__, gguf_get_version(ctx));
+ LOG_TEE("%s: alignment: %zu\n", __func__, gguf_get_alignment(ctx));
+ LOG_TEE("%s: n_tensors: %d\n", __func__, n_tensors);
+ LOG_TEE("%s: n_kv: %d\n", __func__, n_kv);
+ LOG_TEE("%s: ftype: %s\n", __func__, ftype_str.c_str());
+ LOG_TEE("\n");
}
const int n_tensors = gguf_get_n_tensors(ctx);
// kv
const int n_kv = gguf_get_n_kv(ctx);
- printf("%s: loaded meta data with %d key-value pairs and %d tensors from %s\n",
+ LOG_TEE("%s: loaded meta data with %d key-value pairs and %d tensors from %s\n",
__func__, n_kv, n_tensors, fname);
{
std::map<enum ggml_type, uint32_t> n_type;
n_type[type]++;
}
- printf("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__);
+ LOG_TEE("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__);
for (int i = 0; i < n_kv; i++) {
const char * name = gguf_get_key(ctx, i);
const enum gguf_type type = gguf_get_kv_type(ctx, i);
}
replace_all(value, "\n", "\\n");
- printf("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str());
+ LOG_TEE("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str());
}
// print type counts
continue;
}
- printf("%s: - type %4s: %4d tensors\n", __func__, ggml_type_name(kv.first), kv.second);
+ LOG_TEE("%s: - type %4s: %4d tensors\n", __func__, ggml_type_name(kv.first), kv.second);
}
}
size_t tensor_size = ggml_nbytes(cur);
model_size += tensor_size;
if (verbosity >= 3) {
- printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n",
+ LOG_TEE("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n",
__func__, i, ggml_n_dims(cur), cur->name, tensor_size, offset, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3], ggml_type_name(type));
}
}
#ifdef GGML_USE_CUDA
new_clip->backend = ggml_backend_cuda_init(0);
- printf("%s: CLIP using CUDA backend\n", __func__);
+ LOG_TEE("%s: CLIP using CUDA backend\n", __func__);
#endif
#ifdef GGML_USE_METAL
new_clip->backend = ggml_backend_metal_init();
- printf("%s: CLIP using Metal backend\n", __func__);
+ LOG_TEE("%s: CLIP using Metal backend\n", __func__);
#endif
if (!new_clip->backend) {
new_clip->backend = ggml_backend_cpu_init();
- printf("%s: CLIP using CPU backend\n", __func__);
+ LOG_TEE("%s: CLIP using CPU backend\n", __func__);
}
// model size and capabilities
new_clip->use_gelu = gguf_get_val_bool(ctx, idx);
if (verbosity >= 1) {
- printf("%s: text_encoder: %d\n", __func__, new_clip->has_text_encoder);
- printf("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
- printf("%s: llava_projector: %d\n", __func__, new_clip->has_llava_projector);
- printf("%s: model size: %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
- printf("%s: metadata size: %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
+ LOG_TEE("%s: text_encoder: %d\n", __func__, new_clip->has_text_encoder);
+ LOG_TEE("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
+ LOG_TEE("%s: llava_projector: %d\n", __func__, new_clip->has_llava_projector);
+ LOG_TEE("%s: model size: %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
+ LOG_TEE("%s: metadata size: %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
}
}
- printf("%s: params backend buffer size = % 6.2f MB (%i tensors)\n", __func__, model_size / (1024.0 * 1024.0), n_tensors);
+ LOG_TEE("%s: params backend buffer size = % 6.2f MB (%i tensors)\n", __func__, model_size / (1024.0 * 1024.0), n_tensors);
// load tensors
{
new_clip->ctx_data = ggml_init(params);
if (!new_clip->ctx_data) {
- fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+ LOG_TEE("%s: ggml_init() failed\n", __func__);
clip_free(new_clip);
gguf_free(ctx);
return nullptr;
auto fin = std::ifstream(fname, std::ios::binary);
if (!fin) {
- printf("cannot open model file for loading tensors\n");
+ LOG_TEE("cannot open model file for loading tensors\n");
clip_free(new_clip);
gguf_free(ctx);
return nullptr;
const size_t offset = gguf_get_data_offset(ctx) + gguf_get_tensor_offset(ctx, i);
fin.seekg(offset, std::ios::beg);
if (!fin) {
- printf("%s: failed to seek for tensor %s\n", __func__, name);
+ LOG_TEE("%s: failed to seek for tensor %s\n", __func__, name);
clip_free(new_clip);
gguf_free(ctx);
return nullptr;
}
if (verbosity >= 2) {
- printf("\n%s: vision model hparams\n", __func__);
- printf("image_size %d\n", hparams.image_size);
- printf("patch_size %d\n", hparams.patch_size);
- printf("v_hidden_size %d\n", hparams.hidden_size);
- printf("v_n_intermediate %d\n", hparams.n_intermediate);
- 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: ");
+ LOG_TEE("\n%s: vision model hparams\n", __func__);
+ LOG_TEE("image_size %d\n", hparams.image_size);
+ LOG_TEE("patch_size %d\n", hparams.patch_size);
+ LOG_TEE("v_hidden_size %d\n", hparams.hidden_size);
+ LOG_TEE("v_n_intermediate %d\n", hparams.n_intermediate);
+ LOG_TEE("v_projection_dim %d\n", hparams.projection_dim);
+ LOG_TEE("v_n_head %d\n", hparams.n_head);
+ LOG_TEE("v_n_layer %d\n", hparams.n_layer);
+ LOG_TEE("v_eps %f\n", hparams.eps);
+ LOG_TEE("v_image_mean %f %f %f\n", new_clip->image_mean[0], new_clip->image_mean[1], new_clip->image_mean[2]);
+ LOG_TEE("v_image_std %f %f %f\n", new_clip->image_std[0], new_clip->image_std[1], new_clip->image_std[2]);
+ LOG_TEE("v_image_grid_pinpoints: ");
for (int i = 0; i < 32 && (hparams.image_grid_pinpoints[i] != 0); ++i) {
- printf("%d ", hparams.image_grid_pinpoints[i]);
+ LOG_TEE("%d ", hparams.image_grid_pinpoints[i]);
}
- printf("\n");
- printf("v_mm_patch_merge_type: %s\n", hparams.mm_patch_merge_type);
+ LOG_TEE("\n");
+ LOG_TEE("v_mm_patch_merge_type: %s\n", hparams.mm_patch_merge_type);
}
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__);
+ LOG_TEE("%s: failed to load vision model tensors\n", __func__);
}
// LLaVA projection
} catch (std::runtime_error & e) { }
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__);
+ // LOG_TEE("%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
ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch);
ggml_gallocr_reserve(new_clip->compute_alloc, gf);
size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0);
- printf("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
+ LOG_TEE("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
}
return new_clip;
int nx, ny, nc;
auto * data = stbi_load(fname, &nx, &ny, &nc, 3);
if (!data) {
- fprintf(stderr, "%s: failed to load image '%s'\n", __func__, fname);
+ LOG_TEE("%s: failed to load image '%s'\n", __func__, fname);
return false;
}
build_clip_img_from_data(data, nx, ny, img);
int nx, ny, nc;
auto * data = stbi_load_from_memory(bytes, bytes_length, &nx, &ny, &nc, 3);
if (!data) {
- fprintf(stderr, "%s: failed to decode image bytes\n", __func__);
+ LOG_TEE("%s: failed to decode image bytes\n", __func__);
return false;
}
build_clip_img_from_data(data, nx, ny, img);
int downscaled_height = static_cast<int>(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);
+ // LOG_TEE("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;
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");
+ LOG_TEE("This gguf file seems to have no vision encoder\n");
return false;
}
auto & params = ctx->vision_model.hparams;
}
for (size_t i = 0; i < patches.size(); i++) {
- // printf("patch %d: %d %d\n", i, patches[i]->nx, patches[i]->ny);
+ // LOG_TEE("patch %d: %d %d\n", i, patches[i]->nx, patches[i]->ny);
clip_image_u8_free(patches[i]);
}
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");
+ LOG_TEE("This gguf file seems to have no vision encoder\n");
return false;
}
bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs, float * vec) {
if (!ctx->has_vision_encoder) {
- printf("This gguf file seems to have no vision encoder\n");
+ LOG_TEE("This gguf file seems to have no vision encoder\n");
return false;
}
new_type = type;
if (new_type >= GGML_TYPE_Q2_K && name.find("embd") != std::string::npos) {
new_type = GGML_TYPE_Q8_0; // ggml_get_rows needs non K type
- // fprintf(stderr, "%s: quantizing %s to %s\n", __func__, name.c_str(), ggml_type_name(new_type));
+ // LOG_TEE("%s: quantizing %s to %s\n", __func__, name.c_str(), ggml_type_name(new_type));
}
const size_t n_elms = ggml_nelements(cur);
float * f32_data;
f32_data = (float *)conv_buf.data();
break;
default:
- printf("Please use an input file in f32 or f16\n");
+ LOG_TEE("Please use an input file in f32 or f16\n");
gguf_free(ctx_out);
return false;
}
fout.put(0);
}
- printf("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), ggml_n_dims(cur), quantize,
+ LOG_TEE("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), ggml_n_dims(cur), quantize,
orig_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
}
gguf_free(ctx_out);
{
- printf("%s: original size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
- printf("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
+ LOG_TEE("%s: original size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
+ LOG_TEE("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
}
return true;
#include "ggml.h"
+#include "log.h"
#include "common.h"
#include "clip.h"
#include "llava.h"
n_eval = n_batch;
}
if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
- fprintf(stderr, "%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
+ LOG_TEE("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
return false;
}
*n_past += n_eval;
size_t img_base64_str_start, img_base64_str_end;
find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end);
if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) {
- fprintf(stderr, "%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
+ LOG_TEE("%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
return NULL;
}
auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size());
if (!embed) {
- fprintf(stderr, "%s: could not load image from base64 string.\n", __func__);
+ LOG_TEE("%s: could not load image from base64 string.\n", __func__);
return NULL;
}
};
static void show_additional_info(int /*argc*/, char ** argv) {
-
fprintf(stderr, "\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
- fprintf(stderr, " note: a lower temperature value like 0.1 is recommended for better quality.\n");
+
LOG_TEE("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+ LOG_TEE(" note: a lower temperature value like 0.1 is recommended for better quality.\n");
}
static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_params * params) {
auto prompt = params->prompt;
if (prompt_contains_image(prompt)) {
if (!params->image.empty()) {
- fprintf(stderr, "using base64 encoded image instead of command line image path\n");
+ LOG_TEE("using base64 encoded image instead of command line image path\n");
}
embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->n_threads, prompt);
if (!embed) {
- fprintf(stderr, "%s: can't load image from prompt\n", __func__);
+ LOG_TEE("%s: can't load image from prompt\n", __func__);
return NULL;
}
params->prompt = remove_image_from_prompt(prompt);
} else {
embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->n_threads, params->image.c_str());
if (!embed) {
- fprintf(stderr, "%s: is %s really an image file?\n", __func__, params->image.c_str());
+ LOG_TEE("%s: is %s really an image file?\n", __func__, params->image.c_str());
return NULL;
}
}
// new templating mode: Provide the full prompt including system message and use <image> as a placeholder for the image
system_prompt = prompt.substr(0, image_pos);
user_prompt = prompt.substr(image_pos + std::string("<image>").length());
- printf("system_prompt: %s\n", system_prompt.c_str());
+ LOG_TEE("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());
+ LOG_TEE("%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());
+ LOG_TEE("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());
+ LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
}
}
} else {
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());
+ LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
}
}
}
// generate the response
- fprintf(stderr, "\n");
+ LOG_TEE("\n");
struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
std::string response = "";
llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
if (model == NULL) {
- fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+ LOG_TEE("%s: error: unable to load model\n" , __func__);
return NULL;
}
llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
if (ctx_llama == NULL) {
- fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+ LOG_TEE("%s: error: failed to create the llama_context\n" , __func__);
return NULL;
}
llama_backend_free();
}
+static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
+ (void) level;
+ (void) user_data;
+ LOG_TEE("%s", text);
+}
+
int main(int argc, char ** argv) {
ggml_time_init();
show_additional_info(argc, argv);
return 1;
}
+
+#ifndef LOG_DISABLE_LOGS
+ log_set_target(log_filename_generator("llava", "log"));
+ LOG_TEE("Log start\n");
+ log_dump_cmdline(argc, argv);
+ llama_log_set(llama_log_callback_logTee, nullptr);
+#endif // LOG_DISABLE_LOGS
+
if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
gpt_print_usage(argc, argv, params);
show_additional_info(argc, argv);
auto ctx_llava = llava_init(¶ms);
if (ctx_llava == NULL) {
- fprintf(stderr, "%s: error: failed to init llava\n", __func__);
+ LOG_TEE("%s: error: failed to init llava\n", __func__);
return 1;
}
int downscaled_height = static_cast<int>(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);
+ // LOG_TEE("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;
model.newline = ggml_new_tensor_1d(model.ctx, GGML_TYPE_F32, newline_tmp->ne[0]);
if (newline_tmp->backend != GGML_BACKEND_TYPE_CPU) {
if (newline_tmp->buffer == NULL) {
- printf("newline_tmp tensor buffer is NULL\n");
+ LOG_TEE("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");
+ LOG_TEE("newline_tmp tensor data is NULL\n");
}
}
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__);
+ LOG_TEE("%s: unable to preprocess image\n", __func__);
delete[] img_res_v.data;
return false;
}
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");
+ LOG_TEE("Unable to encode image\n");
return false;
}
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);
+ LOG_TEE("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);
+ LOG_TEE("%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);
// clip_image_save_to_bmp(*tmp, "image_feature.bmp");
}
- printf("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
+ LOG_TEE("%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;
- printf("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos);
+ LOG_TEE("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos);
return true;
}
int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama));
auto n_image_embd = clip_n_mmproj_embd(ctx_clip);
if (n_image_embd != n_llama_embd) {
- printf("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
+ LOG_TEE("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
return false;
}
return true;
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)*6); // TODO: base on gridsize/llava model
if (!image_embd) {
- fprintf(stderr, "Unable to allocate memory for image embeddings\n");
+ LOG_TEE("Unable to allocate memory for image embeddings\n");
return false;
}
int n_img_pos;
if (!encode_image_with_clip(ctx_clip, n_threads, img, image_embd, &n_img_pos)) {
- fprintf(stderr, "%s: cannot encode image, aborting\n", __func__);
+ LOG_TEE("%s: cannot encode image, aborting\n", __func__);
free(image_embd);
return false;
}
}
llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
if (llama_decode(ctx_llama, batch)) {
- fprintf(stderr, "%s : failed to eval\n", __func__);
+ LOG_TEE("%s : failed to eval\n", __func__);
return false;
}
*n_past += n_eval;
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);
- fprintf(stderr, "%s: can't load image from bytes, is it a valid image?", __func__);
+ LOG_TEE("%s: can't load image from bytes, is it a valid image?", __func__);
return NULL;
}
bool image_embed_result = llava_image_embed_make_with_clip_img(ctx_clip, n_threads, img, &image_embed, &n_image_pos);
if (!image_embed_result) {
clip_image_u8_free(img);
- fprintf(stderr, "%s: coulnd't embed the image\n", __func__);
+ LOG_TEE("%s: coulnd't embed the image\n", __func__);
return NULL;
}
static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long *sizeOut) {
auto file = fopen(path, "rb");
if (file == NULL) {
- fprintf(stderr, "%s: can't read file %s\n", __func__, path);
+ LOG_TEE("%s: can't read file %s\n", __func__, path);
return false;
}
auto buffer = (unsigned char *)malloc(fileSize); // Allocate memory to hold the file data
if (buffer == NULL) {
- fprintf(stderr, "%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path);
+ LOG_TEE("%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path);
perror("Memory allocation error");
fclose(file);
return false;
long image_bytes_length;
auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
if (!loaded) {
- fprintf(stderr, "%s: failed to load %s\n", __func__, image_path);
+ LOG_TEE("%s: failed to load %s\n", __func__, image_path);
return NULL;
}
overview / pkg / ggml / sources / llama.cpp / commitdiff