Skip to content

Commit 2b6ec97

Browse files
leejetleejet
authored
sync: update ggml (leejet#134)
1 parent db38234 commit 2b6ec97

File tree

9 files changed

+21
-65
lines changed

9 files changed

+21
-65
lines changed

‎clip.hpp

Lines changed: 1 addition & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -443,16 +443,13 @@ struct ResidualAttentionBlock {
443443
struct ggml_tensor* ln2_w; // [hidden_size, ]
444444
struct ggml_tensor* ln2_b; // [hidden_size, ]
445445

446-
struct ggml_tensor* attn_scale; // [hidden_size, ]
447-
448446
size_t calculate_mem_size(ggml_type wtype) {
449447
double mem_size = 0;
450448
mem_size += 4 * hidden_size * hidden_size * ggml_type_sizef(wtype); // q_w/k_w/v_w/out_w
451449
mem_size += 8 * hidden_size * ggml_type_sizef(GGML_TYPE_F32); // q_b/k_b/v_b/out_b/ln1_w/ln1_b/ln2_w/ln2_b
452450
mem_size += 2 * hidden_size * intermediate_size * ggml_type_sizef(wtype); // fc1_w/fc2_w
453451
mem_size += intermediate_size * ggml_type_sizef(GGML_TYPE_F32); // fc1_b
454452
mem_size += hidden_size * ggml_type_sizef(GGML_TYPE_F32); // fc2_b
455-
mem_size += ggml_type_sizef(GGML_TYPE_F32); // attn_scale
456453
return static_cast<size_t>(mem_size);
457454
}
458455

@@ -479,10 +476,6 @@ struct ResidualAttentionBlock {
479476
ln2_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
480477
ln2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
481478

482-
attn_scale = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
483-
ggml_allocr_alloc(alloc, attn_scale);
484-
float scale = 1.0f / sqrt((float)d_model);
485-
ggml_backend_tensor_set(attn_scale, &scale, 0, sizeof(scale));
486479
}
487480

488481
void map_by_name(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
@@ -521,7 +514,7 @@ struct ResidualAttentionBlock {
521514
// self-attention
522515
{
523516
struct ggml_tensor* q = ggml_nn_linear(ctx, x, q_w, q_b);
524-
q = ggml_scale_inplace(ctx, q, attn_scale);
517+
q = ggml_scale_inplace(ctx, q, 1.0f / sqrt((float)d_model));
525518
q = ggml_reshape_4d(ctx, q, d_model, n_head, n_token, N); // [N, n_token, n_head, d_model]
526519
q = ggml_cont(ctx, ggml_permute(ctx, q, 0, 2, 1, 3)); // [N, n_head, n_token, d_model]
527520
q = ggml_reshape_3d(ctx, q, d_model, n_token, n_head * N); // [N * n_head, n_token, d_model]

‎esrgan.hpp

Lines changed: 5 additions & 10 deletions
Original file line numberDiff line numberDiff line change
@@ -91,7 +91,7 @@ struct ResidualDenseBlock {
9191
tensors[prefix + "conv5.bias"] = conv5_b;
9292
}
9393

94-
ggml_tensor* forward(ggml_context* ctx, ggml_tensor* out_scale, ggml_tensor* x /* feat */) {
94+
ggml_tensor* forward(ggml_context* ctx, float out_scale, ggml_tensor* x /* feat */) {
9595
// x1 = self.lrelu(self.conv1(x))
9696
ggml_tensor* x1 = ggml_nn_conv_2d(ctx, x, conv1_w, conv1_b, 1, 1, 1, 1);
9797
x1 = ggml_leaky_relu(ctx, x1, 0.2f, true);
@@ -161,7 +161,7 @@ struct EsrganBlock {
161161
}
162162
}
163163

164-
ggml_tensor* forward(ggml_context* ctx, ggml_tensor* out_scale, ggml_tensor* x) {
164+
ggml_tensor* forward(ggml_context* ctx, float out_scale, ggml_tensor* x) {
165165
ggml_tensor* out = x;
166166
for (int i = 0; i < num_residual_blocks; i++) {
167167
// out = self.rdb...(x)
@@ -325,7 +325,7 @@ struct ESRGAN : public GGMLModule {
325325
tensors["conv_last.bias"] = conv_last_b;
326326
}
327327

328-
ggml_tensor* forward(ggml_context* ctx0, ggml_tensor* out_scale, ggml_tensor* x /* feat */) {
328+
ggml_tensor* forward(ggml_context* ctx0, float out_scale, ggml_tensor* x /* feat */) {
329329
// feat = self.conv_first(feat)
330330
auto h = ggml_nn_conv_2d(ctx0, x, conv_first_w, conv_first_b, 1, 1, 1, 1);
331331

@@ -376,12 +376,7 @@ struct ESRGAN : public GGMLModule {
376376
struct ggml_cgraph* gf = ggml_new_graph(ctx0);
377377

378378
struct ggml_tensor* x_ = NULL;
379-
struct ggml_tensor* os = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
380-
ggml_allocr_alloc(compute_allocr, os);
381-
if (!ggml_allocr_is_measure(compute_allocr)) {
382-
float scale = 0.2f;
383-
ggml_backend_tensor_set(os, &scale, 0, sizeof(scale));
384-
}
379+
float out_scale = 0.2f;
385380

386381
// it's performing a compute, check if backend isn't cpu
387382
if (!ggml_backend_is_cpu(backend)) {
@@ -397,7 +392,7 @@ struct ESRGAN : public GGMLModule {
397392
x_ = x;
398393
}
399394

400-
struct ggml_tensor* out = forward(ctx0, os, x);
395+
struct ggml_tensor* out = forward(ctx0, out_scale, x);
401396

402397
ggml_build_forward_expand(gf, out);
403398
ggml_free(ctx0);

‎ggml

Submodule ggml updated 44 files

‎ggml_extend.hpp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -449,7 +449,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_group_norm(struct ggml_context* ct
449449
struct ggml_tensor* w,
450450
struct ggml_tensor* b,
451451
int num_groups = 32) {
452-
if (x->n_dims == 4) {
452+
if (ggml_n_dims(x) >= 3) {
453453
w = ggml_reshape_4d(ctx, w, 1, 1, w->ne[0], 1);
454454
b = ggml_reshape_4d(ctx, b, 1, 1, b->ne[0], 1);
455455
}

‎lora.hpp

Lines changed: 4 additions & 11 deletions
Original file line numberDiff line numberDiff line change
@@ -113,7 +113,7 @@ struct LoraModel : public GGMLModule {
113113
applied_lora_tensors.insert(scale_name);
114114

115115
// calc_cale
116-
int64_t dim = lora_down->ne[lora_down->n_dims - 1];
116+
int64_t dim = lora_down->ne[ggml_n_dims(lora_down) - 1];
117117
float scale_value = 1.0f;
118118
if (lora_tensors.find(scale_name) != lora_tensors.end()) {
119119
scale_value = ggml_backend_tensor_get_f32(lora_tensors[scale_name]);
@@ -123,17 +123,10 @@ struct LoraModel : public GGMLModule {
123123
}
124124
scale_value *= multiplier;
125125

126-
ggml_tensor* lora_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
127-
128-
ggml_allocr_alloc(compute_allocr, lora_scale);
129-
if (!ggml_allocr_is_measure(compute_allocr)) {
130-
ggml_backend_tensor_set(lora_scale, &scale_value, 0, ggml_nbytes(lora_scale));
131-
}
132-
133126
// flat lora tensors to multiply it
134-
int64_t lora_up_rows = lora_up->ne[lora_up->n_dims - 1];
127+
int64_t lora_up_rows = lora_up->ne[ggml_n_dims(lora_up) - 1];
135128
lora_up = ggml_reshape_2d(ctx0, lora_up, ggml_nelements(lora_up) / lora_up_rows, lora_up_rows);
136-
int64_t lora_down_rows = lora_down->ne[lora_down->n_dims - 1];
129+
int64_t lora_down_rows = lora_down->ne[ggml_n_dims(lora_down) - 1];
137130
lora_down = ggml_reshape_2d(ctx0, lora_down, ggml_nelements(lora_down) / lora_down_rows, lora_down_rows);
138131

139132
// ggml_mul_mat requires tensor b transposed
@@ -142,7 +135,7 @@ struct LoraModel : public GGMLModule {
142135
updown = ggml_cont(ctx0, ggml_transpose(ctx0, updown));
143136
updown = ggml_reshape(ctx0, updown, weight);
144137
GGML_ASSERT(ggml_nelements(updown) == ggml_nelements(weight));
145-
updown = ggml_scale_inplace(ctx0, updown, lora_scale);
138+
updown = ggml_scale_inplace(ctx0, updown, scale_value);
146139
ggml_tensor* final_weight;
147140
// if (weight->type != GGML_TYPE_F32 && weight->type != GGML_TYPE_F16) {
148141
// final_weight = ggml_new_tensor(ctx0, GGML_TYPE_F32, weight->n_dims, weight->ne);

‎model.cpp

Lines changed: 4 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -673,7 +673,7 @@ bool ModelLoader::init_from_gguf_file(const std::string& file_path, const std::s
673673

674674
// LOG_DEBUG("%s", name.c_str());
675675

676-
TensorStorage tensor_storage(prefix + name, dummy->type, dummy->ne, dummy->n_dims, file_index, offset);
676+
TensorStorage tensor_storage(prefix + name, dummy->type, dummy->ne, ggml_n_dims(dummy), file_index, offset);
677677

678678
GGML_ASSERT(ggml_nbytes(dummy) == tensor_storage.nbytes());
679679

@@ -1417,6 +1417,9 @@ bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tenso
14171417
if (pair.first.find("cond_stage_model.transformer.text_model.encoder.layers.23") != std::string::npos) {
14181418
continue;
14191419
}
1420+
if (pair.first.find("alphas_cumprod") != std::string::npos) {
1421+
continue;
1422+
}
14201423

14211424
if (pair.first.find("alphas_cumprod") != std::string::npos) {
14221425
continue;

‎tae.hpp

Lines changed: 2 additions & 15 deletions
Original file line numberDiff line numberDiff line change
@@ -278,9 +278,6 @@ struct TinyDecoder {
278278
ggml_tensor* conv_final_w; // [output_channels, channels, 3, 3]
279279
ggml_tensor* conv_final_b; // [output_channels]
280280

281-
ggml_tensor* in_scale_1d3; // [1]
282-
ggml_tensor* in_scale_3; // [1]
283-
284281
TinyDecoder() {
285282
for (int i = 0; i < num_blocks; i++) {
286283
input_blocks[i].in_channels = channels;
@@ -351,16 +348,6 @@ struct TinyDecoder {
351348
}
352349

353350
final_block.init_params(ctx);
354-
355-
// initialize constants scales
356-
in_scale_1d3 = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
357-
in_scale_3 = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
358-
ggml_allocr_alloc(alloc, in_scale_1d3);
359-
float scale_1d3 = 1.0f / 3.0f;
360-
ggml_backend_tensor_set(in_scale_1d3, &scale_1d3, 0, sizeof(scale_1d3));
361-
ggml_allocr_alloc(alloc, in_scale_3);
362-
float scale_3 = 3.0f;
363-
ggml_backend_tensor_set(in_scale_3, &scale_3, 0, sizeof(scale_3));
364351
}
365352

366353
void map_by_name(std::map<std::string, ggml_tensor*>& tensors, std::string prefix) {
@@ -391,9 +378,9 @@ struct TinyDecoder {
391378

392379
ggml_tensor* forward(ggml_context* ctx, ggml_tensor* z) {
393380
// torch.tanh(x / 3) * 3
394-
auto h = ggml_scale(ctx, z, in_scale_1d3);
381+
auto h = ggml_scale(ctx, z, 1.0f / 3.0f);
395382
h = ggml_tanh_inplace(ctx, h);
396-
h = ggml_scale(ctx, h, in_scale_3);
383+
h = ggml_scale(ctx, h, 3.0f);
397384

398385
// conv(4, 64)
399386
h = ggml_nn_conv_2d(ctx, h, conv_input_w, conv_input_b, 1, 1, 1, 1);

‎unet.hpp

Lines changed: 2 additions & 10 deletions
Original file line numberDiff line numberDiff line change
@@ -182,8 +182,6 @@ struct SpatialTransformer {
182182

183183
std::vector<Transformer> transformers;
184184

185-
struct ggml_tensor* attn_scale;
186-
187185
// proj_out
188186
struct ggml_tensor* proj_out_w; // [in_channels, in_channels, 1, 1]
189187
struct ggml_tensor* proj_out_b; // [in_channels,]
@@ -202,7 +200,6 @@ struct SpatialTransformer {
202200
mem_size += 2 * in_channels * ggml_type_sizef(GGML_TYPE_F32); // norm_w/norm_b
203201
mem_size += 2 * in_channels * in_channels * 1 * 1 * ggml_type_sizef(GGML_TYPE_F16); // proj_in_w/proj_out_w
204202
mem_size += 2 * in_channels * ggml_type_sizef(GGML_TYPE_F32); // proj_in_b/proj_out_b
205-
mem_size += 1 * ggml_type_sizef(GGML_TYPE_F32); // attn_scale
206203

207204
// transformer
208205
for (auto& transformer : transformers) {
@@ -226,11 +223,6 @@ struct SpatialTransformer {
226223
proj_out_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 1, 1, in_channels, in_channels);
227224
proj_out_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, in_channels);
228225

229-
attn_scale = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
230-
ggml_allocr_alloc(alloc, attn_scale);
231-
float scale = 1.0f / sqrt((float)d_head);
232-
ggml_backend_tensor_set(attn_scale, &scale, 0, sizeof(scale));
233-
234226
// transformer
235227
for (auto& transformer : transformers) {
236228
transformer.norm1_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, in_channels);
@@ -332,7 +324,7 @@ struct SpatialTransformer {
332324
x = ggml_reshape_2d(ctx, x, c, h * w * n); // [N * h * w, in_channels]
333325
struct ggml_tensor* q = ggml_mul_mat(ctx, transformer.attn1_q_w, x); // [N * h * w, in_channels]
334326
#if !defined(SD_USE_FLASH_ATTENTION) || defined(SD_USE_CUBLAS) || defined(SD_USE_METAL)
335-
q = ggml_scale_inplace(ctx, q, attn_scale);
327+
q = ggml_scale_inplace(ctx, q, 1.0f / sqrt((float)d_head));
336328
#endif
337329
q = ggml_reshape_4d(ctx, q, d_head, n_head, h * w, n); // [N, h * w, n_head, d_head]
338330
q = ggml_cont(ctx, ggml_permute(ctx, q, 0, 2, 1, 3)); // [N, n_head, h * w, d_head]
@@ -380,7 +372,7 @@ struct SpatialTransformer {
380372
context = ggml_reshape_2d(ctx, context, context->ne[0], context->ne[1] * context->ne[2]); // [N * max_position, hidden_size]
381373
struct ggml_tensor* q = ggml_mul_mat(ctx, transformer.attn2_q_w, x); // [N * h * w, in_channels]
382374
#if !defined(SD_USE_FLASH_ATTENTION) || defined(SD_USE_CUBLAS) || defined(SD_USE_METAL)
383-
q = ggml_scale_inplace(ctx, q, attn_scale);
375+
q = ggml_scale_inplace(ctx, q, 1.0f / sqrt((float)d_head));
384376
#endif
385377
q = ggml_reshape_4d(ctx, q, d_head, n_head, h * w, n); // [N, h * w, n_head, d_head]
386378
q = ggml_cont(ctx, ggml_permute(ctx, q, 0, 2, 1, 3)); // [N, n_head, h * w, d_head]

‎vae.hpp

Lines changed: 1 addition & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -118,8 +118,6 @@ struct AttnBlock {
118118
struct ggml_tensor* proj_out_w; // [in_channels, in_channels, 1, 1]
119119
struct ggml_tensor* proj_out_b; // [in_channels,]
120120

121-
struct ggml_tensor* attn_scale;
122-
123121
size_t calculate_mem_size(ggml_type wtype) {
124122
double mem_size = 0;
125123
mem_size += 6 * in_channels * ggml_type_sizef(GGML_TYPE_F32); // norm_w/norm_b/q_b/k_v/v_b/proj_out_b
@@ -140,11 +138,6 @@ struct AttnBlock {
140138

141139
proj_out_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 1, 1, in_channels, in_channels);
142140
proj_out_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, in_channels);
143-
144-
attn_scale = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
145-
ggml_allocr_alloc(alloc, attn_scale);
146-
float scale = 1.0f / sqrt((float)in_channels);
147-
ggml_backend_tensor_set(attn_scale, &scale, 0, sizeof(scale));
148141
}
149142

150143
void map_by_name(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
@@ -181,7 +174,7 @@ struct AttnBlock {
181174
k = ggml_reshape_3d(ctx, k, c, h * w, n); // [N, h * w, in_channels]
182175

183176
auto w_ = ggml_mul_mat(ctx, k, q); // [N, h * w, h * w]
184-
w_ = ggml_scale_inplace(ctx, w_, attn_scale);
177+
w_ = ggml_scale_inplace(ctx, w_, 1.0f / sqrt((float)in_channels));
185178
w_ = ggml_soft_max_inplace(ctx, w_);
186179

187180
v = ggml_reshape_3d(ctx, v, h * w, c, n); // [N, in_channels, h * w]

0 commit comments

Comments
 (0)