Skip to content

Commit d981f19

Browse files
sfallahsfallah
committed
minor editorconfig-check fixes
1 parent 1c88647 commit d981f19

File tree

2 files changed

+37
-35
lines changed

2 files changed

+37
-35
lines changed

‎tools/mtmd/clip-impl.h‎

Lines changed: 6 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -457,7 +457,7 @@ static std::string to_ne_string(const ggml_tensor * t) {
457457
static void print_tensor_info(ggml_tensor * t) {
458458
const struct ggml_tensor * src0 = t->src[0];
459459
const struct ggml_tensor * src1 = t->src[1];
460-
460+
461461
char src1_str[128] = {0};
462462
if (src1) {
463463
snprintf(src1_str, sizeof(src1_str), "%s{%s}", src1->name, to_ne_string(src1).c_str());
@@ -643,7 +643,7 @@ static void save_tensor_to_file(const struct ggml_tensor * tensor, const uint8_t
643643
append_str(" [");
644644
for (int64_t i0 = 0; i0 < ne[0]; i0++) {
645645
size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
646-
float v;
646+
float v;
647647
if (type == GGML_TYPE_F16) {
648648
v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]);
649649
} else if (type == GGML_TYPE_F32) {
@@ -659,13 +659,15 @@ static void save_tensor_to_file(const struct ggml_tensor * tensor, const uint8_t
659659
}
660660
int len = snprintf(num_buf, sizeof(num_buf), "%8.4f", v);
661661
append(num_buf, len);
662-
if (i0 < ne[0] - 1) append_str(", ");
662+
if (i0 < ne[0] - 1) {
663+
append_str(", ");
664+
}
663665
}
664666
append_str("],\n");
665667
}
666668
append_str(" ],\n");
667669
}
668-
append_str(" ]"); // End of batch
670+
append_str(" ]"); // End of batch
669671
if (i3 < ne[3] - 1) {
670672
append_str(",\n"); // Comma between batches
671673
} else {

‎tools/mtmd/clip.cpp‎

Lines changed: 31 additions & 31 deletions
Original file line numberDiff line numberDiff line change
@@ -662,19 +662,19 @@ struct clip_graph {
662662

663663
ggml_cgraph * build_deepseek_ocr() {
664664
//patch embedding
665-
ggml_tensor * inp_raw = build_inp_raw();
666-
ggml_tensor * sam_out = build_sam(inp_raw);
665+
ggml_tensor * inp_raw = build_inp_raw();
666+
ggml_tensor * sam_out = build_sam(inp_raw);
667667
ggml_tensor * clip_out = build_dsocr_clip(sam_out);
668-
668+
669669
int clip_n_patches = sam_out->ne[0] * sam_out->ne[1];
670-
671-
sam_out = ggml_cont(ctx0, ggml_permute(ctx0, sam_out, 1, 2, 0, 3));
672-
sam_out = ggml_reshape_2d(ctx0, sam_out, sam_out->ne[0], clip_n_patches);
670+
671+
sam_out = ggml_cont(ctx0, ggml_permute(ctx0, sam_out, 1, 2, 0, 3));
672+
sam_out = ggml_reshape_2d(ctx0, sam_out, sam_out->ne[0], clip_n_patches);
673673
clip_out = ggml_view_2d(ctx0, clip_out, n_embd, clip_n_patches, clip_out->nb[1], clip_out->nb[1]);
674-
674+
675675
ggml_tensor * cur;
676676
cur = ggml_concat(ctx0, clip_out, sam_out, 0);
677-
cur = ggml_reshape_2d(ctx0, cur, 2*n_embd,clip_n_patches);
677+
cur = ggml_reshape_2d(ctx0, cur, 2 * n_embd, clip_n_patches);
678678
cur = ggml_cont(ctx0, cur);
679679
cur = ggml_mul_mat(ctx0, model.fc_w, cur);
680680
cur = ggml_add(ctx0, cur, model.fc_b);
@@ -687,10 +687,10 @@ struct clip_graph {
687687
ggml_tensor * vs;
688688

689689
imgnl = ggml_repeat_4d(ctx0, model.image_newline, n_dim, 1, h, 1);
690-
vs = ggml_reshape_2d(ctx0, model.view_seperator, n_dim, 1); // (n_dim, 1)
691-
cur = ggml_reshape_3d(ctx0, cur, n_dim, w, h);
692-
cur = ggml_reshape_2d(ctx0, ggml_concat(ctx0, cur, imgnl, 1), n_dim, (w+1)*h);
693-
cur = ggml_concat(ctx0, cur, vs, 1); // (n_dim, h*(w+1) + 1)
690+
vs = ggml_reshape_2d(ctx0, model.view_seperator, n_dim, 1); // (n_dim, 1)
691+
cur = ggml_reshape_3d(ctx0, cur, n_dim, w, h);
692+
cur = ggml_reshape_2d(ctx0, ggml_concat(ctx0, cur, imgnl, 1), n_dim, (w + 1) * h);
693+
cur = ggml_concat(ctx0, cur, vs, 1); // (n_dim, h*(w+1) + 1)
694694

695695
cb(cur, "dsocr_output", -1);
696696

@@ -2156,7 +2156,7 @@ struct clip_graph {
21562156
ggml_tensor * Qcur;
21572157
ggml_tensor * Kcur;
21582158
ggml_tensor * Vcur;
2159-
2159+
21602160
if (layer.qkv_w) {
21612161
ggml_tensor * QKV;
21622162

@@ -2181,12 +2181,12 @@ struct clip_graph {
21812181
if (layer.q_b) {
21822182
Qcur = ggml_add(ctx0, Qcur, layer.q_b);
21832183
}
2184-
2184+
21852185
Kcur = ggml_mul_mat(ctx0, layer.k_w, cur);
21862186
if (layer.k_b) {
21872187
Kcur = ggml_add(ctx0, Kcur, layer.k_b);
21882188
}
2189-
2189+
21902190
Vcur = ggml_mul_mat(ctx0, layer.v_w, cur);
21912191
if (layer.v_b) {
21922192
Vcur = ggml_add(ctx0, Vcur, layer.v_b);
@@ -2591,7 +2591,7 @@ struct clip_graph {
25912591
} else {
25922592
ggml_tensor * v = ggml_permute(ctx0, v_cur, 1, 2, 0, 3);
25932593
v = ggml_cont(ctx0, v);
2594-
2594+
25952595
ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
25962596
// F32 may not needed for vision encoders?
25972597
// ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
@@ -2727,11 +2727,11 @@ struct clip_graph {
27272727
const int d_heads = n_embd / n_heads;
27282728

27292729
ggml_tensor * inpL;
2730-
2730+
27312731
inpL = ggml_conv_2d_sk_p0(ctx0, model.patch_embed_proj_w, inp_raw);
27322732
inpL = ggml_add(ctx0, inpL, ggml_reshape_3d(ctx0, model.patch_embed_proj_b, 1, 1, n_embd));
27332733
inpL = ggml_cont(ctx0, ggml_permute(ctx0, inpL, 1, 2, 0, 3));
2734-
2734+
27352735
ggml_tensor * cur;
27362736
const auto tgt_size = inpL->ne[1];
27372737
const auto str_size = model.pos_embed->ne[1];
@@ -2776,7 +2776,7 @@ struct clip_graph {
27762776
// self-attention
27772777
{
27782778
const int B = cur->ne[3];
2779-
2779+
27802780
cur = ggml_mul_mat(ctx0, layer.qkv_w, cur);
27812781
cur = ggml_add(ctx0, cur, layer.qkv_b);
27822782
cur = ggml_cont(ctx0, cur); // Ensure tensor is contiguous before reshape
@@ -2853,7 +2853,7 @@ struct clip_graph {
28532853
cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 1, 2, 0, 3));
28542854
cur = build_norm(cur, model.neck_1_w, model.neck_1_b, NORM_TYPE_NORMAL, hparams.eps, -1);
28552855
cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 2, 0, 1, 3));
2856-
2856+
28572857
cur = ggml_conv_2d(ctx0, model.neck_2_w, cur, 1, 1, 1, 1, 1, 1);
28582858
cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 1, 2, 0, 3));
28592859
cur = build_norm(cur, model.neck_3_w, model.neck_3_b, NORM_TYPE_NORMAL, hparams.eps, -1);
@@ -2883,7 +2883,7 @@ struct clip_graph {
28832883
if (tgt_size != src_size) {
28842884
ggml_tensor * old_pos_embd;
28852885
ggml_tensor * cls_tok;
2886-
2886+
28872887
old_pos_embd = ggml_view_2d(
28882888
ctx0, new_pos_embd,
28892889
new_pos_embd->ne[0], src_size * src_size,
@@ -2912,7 +2912,7 @@ struct clip_graph {
29122912
ggml_tensor * positions = ggml_cast(ctx0, ggml_arange(ctx0, 0, n_pos, 1), GGML_TYPE_I32);
29132913
ggml_tensor * learned_pos_embd = ggml_get_rows(ctx0, new_pos_embd, positions);
29142914

2915-
ggml_tensor * cur = build_vit(inp, n_pos, NORM_TYPE_NORMAL, ffn_op_type::FFN_GELU_QUICK,
2915+
ggml_tensor * cur = build_vit(inp, n_pos, NORM_TYPE_NORMAL, ffn_op_type::FFN_GELU_QUICK,
29162916
learned_pos_embd, nullptr); // shape [1024, 16, 16]
29172917

29182918
ggml_build_forward_expand(gf, cur);
@@ -5193,11 +5193,11 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
51935193
const int orig_h = original_size.height;
51945194
const int orig_area = orig_h * orig_w;
51955195
std::array<uint8_t, 3u> color;
5196-
5196+
51975197
for (int i = 0; i < 3; i++) {
51985198
color[i] = (int)(255 * params.image_mean[i]);
51995199
}
5200-
5200+
52015201
int mode_i = 0;
52025202
int min_diff = orig_area;
52035203

@@ -5212,7 +5212,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
52125212
if (mode_i < 2) {
52135213
/* Native Resolution (Tiny/Small) */
52145214
const int image_size = native_resolutions[mode_i];
5215-
5215+
52165216
// Just resize the image to image_size × image_size
52175217
clip_image_u8_ptr resized_img(clip_image_u8_init());
52185218
img_tool::resize(*img, *resized_img,
@@ -5229,7 +5229,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
52295229
else if (mode_i < 4) {
52305230
/* Native Resolution (Base/Large) */
52315231
const int image_size = native_resolutions[mode_i];
5232-
5232+
52335233
// Resize maintaining aspect ratio, then pad to square
52345234
float scale = std::min(
52355235
static_cast<float>(image_size) / orig_w,
@@ -5286,7 +5286,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
52865286
else {
52875287
GGML_ABORT("DeepSeek-OCR hasn't supported Gundam/Gundam-Master yet");
52885288
/* Dynamic Resolution (Gundam/Gundam-Master) */
5289-
5289+
52905290
// configurable, or read from params
52915291
const int min_num = 2;
52925292
const int max_num = 9;
@@ -5295,18 +5295,18 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
52955295
// original image size
52965296
const int orig_w = original_size.width;
52975297
const int orig_h = original_size.height;
5298-
5298+
52995299
// create overview image (thumbnail)
53005300
clip_image_u8_ptr overview_img(clip_image_u8_init());
5301-
img_tool::resize(*img, *overview_img, { image_size, image_size },
5301+
img_tool::resize(*img, *overview_img, { image_size, image_size },
53025302
img_tool::RESIZE_ALGO_BICUBIC_PILLOW, true, color);
53035303
clip_image_f32_ptr overview_f32(clip_image_f32_init());
53045304
normalize_image_u8_to_f32(*overview_img, *overview_f32, params.image_mean, params.image_std);
53055305
res_imgs->entries.push_back(std::move(overview_f32));
53065306

53075307
// build candidate grids (cols, rows)
53085308
auto target_ratios = ds_build_target_ratios(min_num, max_num);
5309-
5309+
53105310
// pick the grid that best matches the original aspect ratio
53115311
const float aspect_ratio = static_cast<float>(orig_w) / static_cast<float>(orig_h);
53125312
auto best = ds_find_closest_ratio(aspect_ratio, target_ratios, orig_w, orig_h, image_size);
@@ -5315,7 +5315,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
53155315

53165316
// resize to refined size (no padding, direct resize)
53175317
clip_image_u8_ptr refined_img(clip_image_u8_init());
5318-
img_tool::resize(*img, *refined_img, { image_size * grid_cols, image_size * grid_rows },
5318+
img_tool::resize(*img, *refined_img, { image_size * grid_cols, image_size * grid_rows },
53195319
img_tool::RESIZE_ALGO_BICUBIC_PILLOW, false);
53205320

53215321
// crop slices from the refined image

0 commit comments

Comments
 (0)