happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

mtmd: remove --dsocr-mode argument

This commit is contained in:
bluebread 2025-12-03 14:54:16 +00:00
parent 43dfc0c8d6
commit b696c54756
7 changed files with 8 additions and 82 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
common/arg.cpp
View File

@ -1829,21 +1829,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.image_max_tokens = value;
}
).set_examples(mmproj_examples).set_env("LLAMA_ARG_IMAGE_MAX_TOKENS"));
add_opt(common_arg(
{"--dsocr-mode"}, "MODE",
"DeepSeek-OCR resolution mode, one of:\n"
"- auto (default): automatically select resolution\n"
"- tiny, small, base, large: native resolution\n"
"- gundam, gundam-master: dynamic resolution",
[](common_params & params, const std::string & value) {
if (value == "auto" || value == "tiny" || value == "small" || value == "base" ||
value == "large" || value == "gundam" || value == "gundam-master") {
params.dsocr_mode = value;
} else {
throw std::invalid_argument("invalid value");
}
}
).set_examples(mmproj_examples).set_env("LLAMA_ARG_DSOCR_MODE"));
if (llama_supports_rpc()) {
add_opt(common_arg(
{"--rpc"}, "SERVERS",

1
common/common.h
View File

@ -432,7 +432,6 @@ struct common_params {
std::vector<std::string> image; // path to image file(s)
int image_min_tokens = -1;
int image_max_tokens = -1;
std::string dsocr_mode = "auto"; // DeepSeek-OCR resolution mode: auto, tiny, small, base, large, gundam, gundam-master
// finetune
struct lr_opt lr;

37
tools/mtmd/clip.cpp
View File

@ -206,9 +206,6 @@ struct clip_hparams {
int32_t custom_image_min_tokens = -1;
int32_t custom_image_max_tokens = -1;
// DeepSeek-OCR resolution mode
enum clip_dsocr_mode dsocr_mode = clip_dsocr_mode::CLIP_DSOCR_MODE_AUTO;
void set_limit_image_tokens(int n_tokens_min, int n_tokens_max) {
const int cur_merge = n_merge == 0 ? 1 : n_merge;
const int patch_area = patch_size * patch_size * cur_merge * cur_merge;
@ -513,7 +510,6 @@ struct clip_ctx {
if (ctx_params.image_max_tokens > 0) {
model.hparams.custom_image_max_tokens = ctx_params.image_max_tokens;
}
model.hparams.dsocr_mode = ctx_params.dsocr_mode;
backend_ptrs.push_back(backend_cpu);
backend_buft.push_back(ggml_backend_get_default_buffer_type(backend_cpu));
@ -5291,30 +5287,13 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
}
int mode_i = 0;
if (params.dsocr_mode == clip_dsocr_mode::CLIP_DSOCR_MODE_TINY) {
mode_i = 0;
} else if (params.dsocr_mode == clip_dsocr_mode::CLIP_DSOCR_MODE_SMALL) {
mode_i = 1;
} else if (params.dsocr_mode == clip_dsocr_mode::CLIP_DSOCR_MODE_BASE) {
mode_i = 2;
} else if (params.dsocr_mode == clip_dsocr_mode::CLIP_DSOCR_MODE_LARGE) {
mode_i = 3;
} else if (params.dsocr_mode == clip_dsocr_mode::CLIP_DSOCR_MODE_GUNDAM) {
mode_i = 4;
} else if (params.dsocr_mode == clip_dsocr_mode::CLIP_DSOCR_MODE_GUNDAM_MASTER) {
mode_i = 5;
} else {
if (params.dsocr_mode != clip_dsocr_mode::CLIP_DSOCR_MODE_AUTO) {
LOG_WRN("%s: unknown dsocr_mode, using auto mode\n", __func__);
}
int min_diff = orig_area;
for (int i = 0; i < 4; i++) {
int r = native_resolutions[i];
if (std::abs(orig_area - r*r) < min_diff) {
mode_i = i;
min_diff = std::abs(orig_area - r*r);
}
int min_diff = orig_area;
for (int i = 0; i < 4; i++) {
int r = native_resolutions[i];
if (std::abs(orig_area - r*r) < min_diff) {
mode_i = i;
min_diff = std::abs(orig_area - r*r);
}
}
@ -5393,7 +5372,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
res_imgs->grid_y = 1;
}
else {
GGML_ABORT("DeepSeek-OCR: Gundam/Gundam-Master haven't been tested yet.\n");
GGML_ABORT("DeepSeek-OCR hasn't supported Gundam/Gundam-Master yet");
/* Dynamic Resolution (Gundam/Gundam-Master) */
// configurable, or read from params

11
tools/mtmd/clip.h
View File

@ -29,23 +29,12 @@ enum clip_flash_attn_type {
CLIP_FLASH_ATTN_TYPE_ENABLED = 1,
};
enum clip_dsocr_mode {
CLIP_DSOCR_MODE_AUTO,
CLIP_DSOCR_MODE_TINY,
CLIP_DSOCR_MODE_SMALL,
CLIP_DSOCR_MODE_BASE,
CLIP_DSOCR_MODE_LARGE,
CLIP_DSOCR_MODE_GUNDAM,
CLIP_DSOCR_MODE_GUNDAM_MASTER,
};
struct clip_context_params {
bool use_gpu;
enum clip_flash_attn_type flash_attn_type;
int image_min_tokens;
int image_max_tokens;
bool warmup;
enum clip_dsocr_mode dsocr_mode;
};
struct clip_init_result {

1
tools/mtmd/mtmd-cli.cpp
View File

@ -139,7 +139,6 @@ struct mtmd_cli_context {
mparams.warmup = params.warmup;
mparams.image_min_tokens = params.image_min_tokens;
mparams.image_max_tokens = params.image_max_tokens;
mparams.dsocr_mode = params.dsocr_mode.c_str();
ctx_vision.reset(mtmd_init_from_file(clip_path, model, mparams));
if (!ctx_vision.get()) {
LOG_ERR("Failed to load vision model from %s\n", clip_path);

22
tools/mtmd/mtmd.cpp
View File

@ -111,7 +111,6 @@ mtmd_context_params mtmd_context_params_default() {
/* warmup */ true,
/* image_min_tokens */ -1,
/* image_max_tokens */ -1,
/* dsocr_mode */ "auto",
};
return params;
}
@ -174,33 +173,12 @@ struct mtmd_context {
throw std::runtime_error("media_marker must not be empty");
}
enum clip_dsocr_mode dsocr_mode;
if (std::string(ctx_params.dsocr_mode) == "auto") {
dsocr_mode = clip_dsocr_mode::CLIP_DSOCR_MODE_AUTO;
} else if (std::string(ctx_params.dsocr_mode) == "tiny") {
dsocr_mode = clip_dsocr_mode::CLIP_DSOCR_MODE_TINY;
} else if (std::string(ctx_params.dsocr_mode) == "small") {
dsocr_mode = clip_dsocr_mode::CLIP_DSOCR_MODE_SMALL;
} else if (std::string(ctx_params.dsocr_mode) == "base") {
dsocr_mode = clip_dsocr_mode::CLIP_DSOCR_MODE_BASE;
} else if (std::string(ctx_params.dsocr_mode) == "large") {
dsocr_mode = clip_dsocr_mode::CLIP_DSOCR_MODE_LARGE;
} else if (std::string(ctx_params.dsocr_mode) == "gundam") {
dsocr_mode = clip_dsocr_mode::CLIP_DSOCR_MODE_GUNDAM;
} else if (std::string(ctx_params.dsocr_mode) == "gundam-master") {
dsocr_mode = clip_dsocr_mode::CLIP_DSOCR_MODE_GUNDAM_MASTER;
} else {
throw std::invalid_argument("invalid value");
}
clip_context_params ctx_clip_params {
/* use_gpu */ ctx_params.use_gpu,
/* flash_attn_type */ CLIP_FLASH_ATTN_TYPE_AUTO,
/* image_min_tokens */ ctx_params.image_min_tokens,
/* image_max_tokens */ ctx_params.image_max_tokens,
/* warmup */ ctx_params.warmup,
/* dsocr_mode */ dsocr_mode,
};
auto res = clip_init(mmproj_fname, ctx_clip_params);

3
tools/mtmd/mtmd.h
View File

@ -87,9 +87,6 @@ struct mtmd_context_params {
// limit number of image tokens, only for vision models with dynamic resolution
int image_min_tokens; // minimum number of tokens for image input (default: read from metadata)
int image_max_tokens; // maximum number of tokens for image input (default: read from metadata)
// DeepSeek-OCR resolution mode
const char * dsocr_mode; // one of: auto, tiny, small, base, large, gundam, gundam-master
};
MTMD_API const char * mtmd_default_marker(void);