diff options
| -rw-r--r-- | .gitignore | 3 | ||||
| -rw-r--r-- | Makefile | 2 | ||||
| -rw-r--r-- | image.jpg | bin | 0 -> 15279 bytes | |||
| -rw-r--r-- | main.c | 89 |
4 files changed, 80 insertions, 14 deletions
@@ -1,6 +1,9 @@ app + *.o *.gguf deps/ + **/*.swp +**/*.core @@ -71,7 +71,7 @@ deps/lib/libllama.a deps/lib/libggml.a deps/lib/libggml-cpu.a: -DLLAMA_BUILD_APP=OFF \ -DLLAMA_BUILD_COMMON=ON \ -DLLAMA_BUILD_TESTS=OFF \ - -DLLAMA_BUILD_UI=OFF \ + -DLLAMA_BUILD_UI=OFF \ -DLLAMA_BUILD_EXAMPLES=OFF \ -DLLAMA_BUILD_SERVER=OFF \ -DLLAMA_BUILD_TOOLS=OFF diff --git a/image.jpg b/image.jpg Binary files differnew file mode 100644 index 0000000..408e34b --- /dev/null +++ b/image.jpg @@ -8,8 +8,10 @@ #include "mtmd-helper.h" #include "llama.h" -#define N_CTX 1024 +#define N_CTX 4096 +#define N_BATCH 2048 #define N_THREADS 4 +#define N_PREDICT 256 #define MODEL_PATH "model.gguf" #define MMPROJ_PATH "mmproj.gguf" @@ -55,6 +57,7 @@ static char * build_prompt(struct llama_model *model, int n_images) if (n < 0) { fprintf(stderr, "Chat template size check failed\n"); free(content); + return NULL; } buf_len = n + 1; @@ -63,6 +66,7 @@ static char * build_prompt(struct llama_model *model, int n_images) if (n < 0) { fprintf(stderr, "Failed to apply chat template\n"); free(content); + return NULL; } buf[n] = '\0'; @@ -71,7 +75,7 @@ static char * build_prompt(struct llama_model *model, int n_images) int main(int argc, char *argv[]) { - int n_images; + int i, n_images, rc; const char *prompt; const struct llama_vocab *vocab; @@ -82,18 +86,22 @@ int main(int argc, char *argv[]) struct llama_sampler *smpl = NULL; struct llama_sampler_chain_params sparams; - int i; - mtmd_input_chunks *chunks; + int n_bitmaps = 0; + mtmd_input_chunks *chunks = NULL; mtmd_bitmap **bitmaps = NULL; struct mtmd_context *mctx = NULL; struct mtmd_context_params mtparams; struct mtmd_helper_init_opt init_opt; struct mtmd_helper_bitmap_wrapper wrapper; + llama_pos n_past; + + rc = 1; + n_images = argc - 1; if (n_images < 1) { fprintf(stderr, "Usage: %s image1 [image2 ...]\n", argv[0]); - return 1; + return rc; } llama_backend_init(); @@ -157,16 +165,16 @@ int main(int argc, char *argv[]) init_opt = mtmd_helper_init_opt_default(); bitmaps = CALLOC(n_images, sizeof(*bitmaps)); - for (i = 0; i < n_images; i++) { + for (n_bitmaps = 0; n_bitmaps < n_images; n_bitmaps++) { wrapper = mtmd_helper_bitmap_init_from_file( mctx, - argv[i + 1], + argv[n_bitmaps + 1], false, init_opt); - bitmaps[i] = wrapper.bitmap; - if (!bitmaps[i]) { - fprintf(stderr, "Failed to load image: %s\n", argv[i + 1]); + bitmaps[n_bitmaps] = wrapper.bitmap; + if (!bitmaps[n_bitmaps]) { + fprintf(stderr, "Failed to load image: %s\n", argv[n_bitmaps + 1]); goto cleanup; } } @@ -177,21 +185,76 @@ int main(int argc, char *argv[]) goto cleanup; } - printf("Init success\n"); + { + mtmd_input_text text; + + memset(&text, 0, sizeof(text)); + text.text = prompt; + text.text_len = strlen(prompt); + text.add_special = true; /* use model's BOS/EOS tokens */ + text.parse_special = true; /* treat BOS/EOS tokens as special (Qwen/Llama) */ + + chunks = mtmd_input_chunks_init(); + if (mtmd_tokenize(mctx, chunks, &text, + (const mtmd_bitmap **)bitmaps, n_images) != 0) { + fprintf(stderr, "mtmd_tokenize error\n"); + goto cleanup; + } + } + + /* Encode images + decode prompt; logits for the last one */ + if (mtmd_helper_eval_chunks(mctx, ctx, chunks, 0, 0, N_BATCH, + true, &n_past) != 0) { + fprintf(stderr, "Failed to evaluate prompt/images\n"); + goto cleanup; + } + + /* Generation loop */ + { + int t, n; + llama_token token; + + rc = 0; + for (t = 0; t < N_PREDICT && n_past < N_CTX; t++) { + token = llama_sampler_sample(smpl, ctx, -1); + if (llama_vocab_is_eog(vocab, token)) + break; + + n = 0; + char buf[256]; + n = llama_token_to_piece(vocab, token, buf, sizeof(buf), 0, true); + if (n > 0) + fwrite(buf, 1, n, stdout); + fflush(stdout); + + if (llama_decode(ctx, llama_batch_get_one(&token, 1)) != 0) { + rc = 1; + fprintf(stderr, "llama_decode error\n"); + break; + } + + n_past++; + } + printf("\n"); + } cleanup: if (bitmaps) { - for (i = 0; bitmaps[i]; i++) + for (i = 0; i < n_bitmaps && bitmaps[i]; i++) mtmd_bitmap_free(bitmaps[i]); free(bitmaps); } + + if (chunks) + mtmd_input_chunks_free(chunks); if (smpl) llama_sampler_free(smpl); if (ctx) llama_free(ctx); if (model) llama_model_free(model); + llama_backend_free(); - return 0; + return rc; } |
