DarthAffe/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-02-04 10:53:34 +00:00
Code Issues Packages Projects Releases Wiki Activity

chore: eliminate compilation warnings under MSVC (#1170)

Browse Source
This commit is contained in:
leejet 2026-01-04 22:26:57 +08:00 committed by GitHub
parent 2cef4badb8
commit b90b1ee9cf
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
30 changed files with 420 additions and 461 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
CMakeLists.txt
View File

@ -8,6 +8,11 @@ if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo") set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
endif() endif()
if (MSVC)
add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
add_compile_definitions(_SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING)
endif()
set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin) set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin) set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)

2
cache_dit.hpp
View File

@ -117,7 +117,7 @@ struct TaylorSeerState {
continue; continue;
if (o > 0) if (o > 0)
factorial *= static_cast<float>(o); factorial *= static_cast<float>(o);
float coeff = std::pow(static_cast<float>(elapsed), o) / factorial; float coeff = ::powf(static_cast<float>(elapsed), static_cast<float>(o)) / factorial;
for (size_t i = 0; i < size; i++) { for (size_t i = 0; i < size; i++) {
output[i] += coeff * dY_prev[o][i]; output[i] += coeff * dY_prev[o][i];
} }

20
clip.hpp
View File

@ -296,7 +296,7 @@ public:
size_t max_length = 0, size_t max_length = 0,
bool padding = false) { bool padding = false) {
if (max_length > 0 && padding) { if (max_length > 0 && padding) {
size_t n = std::ceil(tokens.size() * 1.0 / (max_length - 2)); size_t n = static_cast<size_t>(std::ceil(tokens.size() * 1.0 / (max_length - 2)));
if (n == 0) { if (n == 0) {
n = 1; n = 1;
} }
@ -525,10 +525,10 @@ public:
struct CLIPEncoder : public GGMLBlock { struct CLIPEncoder : public GGMLBlock {
protected: protected:
int64_t n_layer; int n_layer;
public: public:
CLIPEncoder(int64_t n_layer, CLIPEncoder(int n_layer,
int64_t d_model, int64_t d_model,
int64_t n_head, int64_t n_head,
int64_t intermediate_size, int64_t intermediate_size,
@ -623,10 +623,10 @@ public:
class CLIPVisionEmbeddings : public GGMLBlock { class CLIPVisionEmbeddings : public GGMLBlock {
protected: protected:
int64_t embed_dim; int64_t embed_dim;
int64_t num_channels; int num_channels;
int64_t patch_size; int patch_size;
int64_t image_size; int image_size;
int64_t num_patches; int num_patches;
int64_t num_positions; int64_t num_positions;
void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override { void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
@ -641,9 +641,9 @@ protected:
public: public:
CLIPVisionEmbeddings(int64_t embed_dim, CLIPVisionEmbeddings(int64_t embed_dim,
int64_t num_channels = 3, int num_channels = 3,
int64_t patch_size = 14, int patch_size = 14,
int64_t image_size = 224) int image_size = 224)
: embed_dim(embed_dim), : embed_dim(embed_dim),
num_channels(num_channels), num_channels(num_channels),
patch_size(patch_size), patch_size(patch_size),

8
common.hpp
View File

@ -80,7 +80,7 @@ protected:
std::pair<int, int> padding) { std::pair<int, int> padding) {
GGML_ASSERT(dims == 2 || dims == 3); GGML_ASSERT(dims == 2 || dims == 3);
if (dims == 3) { if (dims == 3) {
return std::shared_ptr<GGMLBlock>(new Conv3dnx1x1(in_channels, out_channels, kernel_size.first, 1, padding.first)); return std::shared_ptr<GGMLBlock>(new Conv3d(in_channels, out_channels, {kernel_size.first, 1, 1}, {1, 1, 1}, {padding.first, 0, 0}));
} else { } else {
return std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, out_channels, kernel_size, {1, 1}, padding)); return std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, out_channels, kernel_size, {1, 1}, padding));
} }
@ -544,9 +544,9 @@ public:
class VideoResBlock : public ResBlock { class VideoResBlock : public ResBlock {
public: public:
VideoResBlock(int channels, VideoResBlock(int64_t channels,
int emb_channels, int64_t emb_channels,
int out_channels, int64_t out_channels,
std::pair<int, int> kernel_size = {3, 3}, std::pair<int, int> kernel_size = {3, 3},
int64_t video_kernel_size = 3, int64_t video_kernel_size = 3,
int dims = 2) // always 2 int dims = 2) // always 2

20
conditioner.hpp
View File

@ -303,11 +303,11 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
int class_token = clean_input_ids[class_token_index[0]]; int class_token = clean_input_ids[class_token_index[0]];
class_idx = tokens_acc + class_token_index[0]; class_idx = tokens_acc + class_token_index[0];
std::vector<int> clean_input_ids_tmp; std::vector<int> clean_input_ids_tmp;
for (uint32_t i = 0; i < class_token_index[0]; i++) for (int i = 0; i < class_token_index[0]; i++)
clean_input_ids_tmp.push_back(clean_input_ids[i]); clean_input_ids_tmp.push_back(clean_input_ids[i]);
for (uint32_t i = 0; i < (pm_version == PM_VERSION_2 ? 2 * num_input_imgs : num_input_imgs); i++) for (int i = 0; i < (pm_version == PM_VERSION_2 ? 2 * num_input_imgs : num_input_imgs); i++)
clean_input_ids_tmp.push_back(class_token); clean_input_ids_tmp.push_back(class_token);
for (uint32_t i = class_token_index[0] + 1; i < clean_input_ids.size(); i++) for (int i = class_token_index[0] + 1; i < clean_input_ids.size(); i++)
clean_input_ids_tmp.push_back(clean_input_ids[i]); clean_input_ids_tmp.push_back(clean_input_ids[i]);
clean_input_ids.clear(); clean_input_ids.clear();
clean_input_ids = clean_input_ids_tmp; clean_input_ids = clean_input_ids_tmp;
@ -322,7 +322,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
tokenizer.pad_tokens(tokens, weights, max_length, padding); tokenizer.pad_tokens(tokens, weights, max_length, padding);
int offset = pm_version == PM_VERSION_2 ? 2 * num_input_imgs : num_input_imgs; int offset = pm_version == PM_VERSION_2 ? 2 * num_input_imgs : num_input_imgs;
for (uint32_t i = 0; i < tokens.size(); i++) { for (int i = 0; i < tokens.size(); i++) {
// if (class_idx + 1 <= i && i < class_idx + 1 + 2*num_input_imgs) // photomaker V2 has num_tokens(=2)*num_input_imgs // if (class_idx + 1 <= i && i < class_idx + 1 + 2*num_input_imgs) // photomaker V2 has num_tokens(=2)*num_input_imgs
if (class_idx + 1 <= i && i < class_idx + 1 + offset) // photomaker V2 has num_tokens(=2)*num_input_imgs if (class_idx + 1 <= i && i < class_idx + 1 + offset) // photomaker V2 has num_tokens(=2)*num_input_imgs
// hardcode for now // hardcode for now
@ -1584,7 +1584,7 @@ struct T5CLIPEmbedder : public Conditioner {
chunk_hidden_states->ne[0], chunk_hidden_states->ne[0],
ggml_nelements(hidden_states) / chunk_hidden_states->ne[0]); ggml_nelements(hidden_states) / chunk_hidden_states->ne[0]);
modify_mask_to_attend_padding(t5_attn_mask, ggml_nelements(t5_attn_mask), mask_pad); modify_mask_to_attend_padding(t5_attn_mask, static_cast<int>(ggml_nelements(t5_attn_mask)), mask_pad);
return {hidden_states, t5_attn_mask, nullptr}; return {hidden_states, t5_attn_mask, nullptr};
} }
@ -1723,8 +1723,8 @@ struct LLMEmbedder : public Conditioner {
double factor = llm->params.vision.patch_size * llm->params.vision.spatial_merge_size; double factor = llm->params.vision.patch_size * llm->params.vision.spatial_merge_size;
int height = image.height; int height = image.height;
int width = image.width; int width = image.width;
int h_bar = static_cast<int>(std::round(height / factor)) * factor; int h_bar = static_cast<int>(std::round(height / factor) * factor);
int w_bar = static_cast<int>(std::round(width / factor)) * factor; int w_bar = static_cast<int>(std::round(width / factor) * factor);
if (static_cast<double>(h_bar) * w_bar > max_pixels) { if (static_cast<double>(h_bar) * w_bar > max_pixels) {
double beta = std::sqrt((height * width) / static_cast<double>(max_pixels)); double beta = std::sqrt((height * width) / static_cast<double>(max_pixels));
@ -1752,7 +1752,7 @@ struct LLMEmbedder : public Conditioner {
ggml_tensor* image_embed = nullptr; ggml_tensor* image_embed = nullptr;
llm->encode_image(n_threads, image_tensor, &image_embed, work_ctx); llm->encode_image(n_threads, image_tensor, &image_embed, work_ctx);
image_embeds.emplace_back(image_embed_idx, image_embed); image_embeds.emplace_back(image_embed_idx, image_embed);
image_embed_idx += 1 + image_embed->ne[1] + 6; image_embed_idx += 1 + static_cast<int>(image_embed->ne[1]) + 6;
img_prompt += "Picture " + std::to_string(i + 1) + ": <|vision_start|>"; // [24669, 220, index, 25, 220, 151652] img_prompt += "Picture " + std::to_string(i + 1) + ": <|vision_start|>"; // [24669, 220, index, 25, 220, 151652]
int64_t num_image_tokens = image_embed->ne[1]; int64_t num_image_tokens = image_embed->ne[1];
@ -1799,9 +1799,9 @@ struct LLMEmbedder : public Conditioner {
prompt = "[SYSTEM_PROMPT]You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object\nattribution and actions without speculation.[/SYSTEM_PROMPT][INST]"; prompt = "[SYSTEM_PROMPT]You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object\nattribution and actions without speculation.[/SYSTEM_PROMPT][INST]";
prompt_attn_range.first = prompt.size(); prompt_attn_range.first = static_cast<int>(prompt.size());
prompt += conditioner_params.text; prompt += conditioner_params.text;
prompt_attn_range.second = prompt.size(); prompt_attn_range.second = static_cast<int>(prompt.size());
prompt += "[/INST]"; prompt += "[/INST]";
} else if (version == VERSION_OVIS_IMAGE) { } else if (version == VERSION_OVIS_IMAGE) {

43
denoiser.hpp
View File

@ -245,7 +245,7 @@ struct SGMUniformScheduler : SigmaScheduler {
int t_max = TIMESTEPS - 1; int t_max = TIMESTEPS - 1;
int t_min = 0; int t_min = 0;
std::vector<float> timesteps = linear_space(static_cast<float>(t_max), static_cast<float>(t_min), n + 1); std::vector<float> timesteps = linear_space(static_cast<float>(t_max), static_cast<float>(t_min), n + 1);
for (int i = 0; i < n; i++) { for (uint32_t i = 0; i < n; i++) {
result.push_back(t_to_sigma_func(timesteps[i])); result.push_back(t_to_sigma_func(timesteps[i]));
} }
result.push_back(0.0f); result.push_back(0.0f);
@ -259,11 +259,11 @@ struct LCMScheduler : SigmaScheduler {
result.reserve(n + 1); result.reserve(n + 1);
const int original_steps = 50; const int original_steps = 50;
const int k = TIMESTEPS / original_steps; const int k = TIMESTEPS / original_steps;
for (int i = 0; i < n; i++) { for (uint32_t i = 0; i < n; i++) {
// the rounding ensures we match the training schedule of the LCM model // the rounding ensures we match the training schedule of the LCM model
int index = (i * original_steps) / n; int index = (i * original_steps) / n;
int timestep = (original_steps - index) * k - 1; int timestep = (original_steps - index) * k - 1;
result.push_back(t_to_sigma(timestep)); result.push_back(t_to_sigma(static_cast<float>(timestep)));
} }
result.push_back(0.0f); result.push_back(0.0f);
return result; return result;
@ -525,8 +525,8 @@ struct CompVisVDenoiser : public CompVisDenoiser {
}; };
struct EDMVDenoiser : public CompVisVDenoiser { struct EDMVDenoiser : public CompVisVDenoiser {
float min_sigma = 0.002; float min_sigma = 0.002f;
float max_sigma = 120.0; float max_sigma = 120.0f;
EDMVDenoiser(float min_sigma = 0.002, float max_sigma = 120.0) EDMVDenoiser(float min_sigma = 0.002, float max_sigma = 120.0)
: min_sigma(min_sigma), max_sigma(max_sigma) { : min_sigma(min_sigma), max_sigma(max_sigma) {
@ -537,7 +537,7 @@ struct EDMVDenoiser : public CompVisVDenoiser {
} }
float sigma_to_t(float s) override { float sigma_to_t(float s) override {
return 0.25 * std::log(s); return 0.25f * std::log(s);
} }
float sigma_min() override { float sigma_min() override {
@ -569,7 +569,7 @@ struct DiscreteFlowDenoiser : public Denoiser {
void set_parameters() { void set_parameters() {
for (int i = 1; i < TIMESTEPS + 1; i++) { for (int i = 1; i < TIMESTEPS + 1; i++) {
sigmas[i - 1] = t_to_sigma(i); sigmas[i - 1] = t_to_sigma(static_cast<float>(i));
} }
} }
@ -612,7 +612,7 @@ struct DiscreteFlowDenoiser : public Denoiser {
}; };
float flux_time_shift(float mu, float sigma, float t) { float flux_time_shift(float mu, float sigma, float t) {
return std::exp(mu) / (std::exp(mu) + std::pow((1.0 / t - 1.0), sigma)); return ::expf(mu) / (::expf(mu) + ::powf((1.0f / t - 1.0f), sigma));
} }
struct FluxFlowDenoiser : public Denoiser { struct FluxFlowDenoiser : public Denoiser {
@ -632,7 +632,7 @@ struct FluxFlowDenoiser : public Denoiser {
void set_parameters(float shift) { void set_parameters(float shift) {
set_shift(shift); set_shift(shift);
for (int i = 0; i < TIMESTEPS; i++) { for (int i = 0; i < TIMESTEPS; i++) {
sigmas[i] = t_to_sigma(i); sigmas[i] = t_to_sigma(static_cast<float>(i));
} }
} }
@ -1327,15 +1327,12 @@ static bool sample_k_diffusion(sample_method_t method,
// - pred_sample_direction -> "direction pointing to // - pred_sample_direction -> "direction pointing to
// x_t" // x_t"
// - pred_prev_sample -> "x_t-1" // - pred_prev_sample -> "x_t-1"
int timestep = int timestep = static_cast<int>(roundf(TIMESTEPS - i * ((float)TIMESTEPS / steps))) - 1;
roundf(TIMESTEPS -
i * ((float)TIMESTEPS / steps)) -
1;
// 1. get previous step value (=t-1) // 1. get previous step value (=t-1)
int prev_timestep = timestep - TIMESTEPS / steps; int prev_timestep = timestep - TIMESTEPS / static_cast<int>(steps);
// The sigma here is chosen to cause the // The sigma here is chosen to cause the
// CompVisDenoiser to produce t = timestep // CompVisDenoiser to produce t = timestep
float sigma = compvis_sigmas[timestep]; float sigma = static_cast<float>(compvis_sigmas[timestep]);
if (i == 0) { if (i == 0) {
// The function add_noise intializes x to // The function add_noise intializes x to
// Diffusers' latents * sigma (as in Diffusers' // Diffusers' latents * sigma (as in Diffusers'
@ -1392,10 +1389,10 @@ static bool sample_k_diffusion(sample_method_t method,
} }
} }
// 2. compute alphas, betas // 2. compute alphas, betas
float alpha_prod_t = alphas_cumprod[timestep]; float alpha_prod_t = static_cast<float>(alphas_cumprod[timestep]);
// Note final_alpha_cumprod = alphas_cumprod[0] due to // Note final_alpha_cumprod = alphas_cumprod[0] due to
// trailing timestep spacing // trailing timestep spacing
float alpha_prod_t_prev = prev_timestep >= 0 ? alphas_cumprod[prev_timestep] : alphas_cumprod[0]; float alpha_prod_t_prev = static_cast<float>(prev_timestep >= 0 ? alphas_cumprod[prev_timestep] : alphas_cumprod[0]);
float beta_prod_t = 1 - alpha_prod_t; float beta_prod_t = 1 - alpha_prod_t;
// 3. compute predicted original sample from predicted // 3. compute predicted original sample from predicted
// noise also called "predicted x_0" of formula (12) // noise also called "predicted x_0" of formula (12)
@ -1442,8 +1439,8 @@ static bool sample_k_diffusion(sample_method_t method,
// Two step inner loop without an explicit // Two step inner loop without an explicit
// tensor // tensor
float pred_sample_direction = float pred_sample_direction =
std::sqrt(1 - alpha_prod_t_prev - ::sqrtf(1 - alpha_prod_t_prev -
std::pow(std_dev_t, 2)) * ::powf(std_dev_t, 2)) *
vec_model_output[j]; vec_model_output[j];
vec_x[j] = std::sqrt(alpha_prod_t_prev) * vec_x[j] = std::sqrt(alpha_prod_t_prev) *
vec_pred_original_sample[j] + vec_pred_original_sample[j] +
@ -1518,7 +1515,7 @@ static bool sample_k_diffusion(sample_method_t method,
// Begin k-diffusion specific workaround for // Begin k-diffusion specific workaround for
// evaluating F_theta(x; ...) from D(x, sigma), same // evaluating F_theta(x; ...) from D(x, sigma), same
// as in DDIM (and see there for detailed comments) // as in DDIM (and see there for detailed comments)
float sigma = compvis_sigmas[timestep]; float sigma = static_cast<float>(compvis_sigmas[timestep]);
if (i == 0) { if (i == 0) {
float* vec_x = (float*)x->data; float* vec_x = (float*)x->data;
for (int j = 0; j < ggml_nelements(x); j++) { for (int j = 0; j < ggml_nelements(x); j++) {
@ -1557,14 +1554,14 @@ static bool sample_k_diffusion(sample_method_t method,
// is different from the notation alpha_t in // is different from the notation alpha_t in
// DPM-Solver. In fact, we have alpha_{t_n} = // DPM-Solver. In fact, we have alpha_{t_n} =
// \sqrt{\hat{alpha_n}}, [...]" // \sqrt{\hat{alpha_n}}, [...]"
float alpha_prod_t = alphas_cumprod[timestep]; float alpha_prod_t = static_cast<float>(alphas_cumprod[timestep]);
float beta_prod_t = 1 - alpha_prod_t; float beta_prod_t = 1 - alpha_prod_t;
// Note final_alpha_cumprod = alphas_cumprod[0] since // Note final_alpha_cumprod = alphas_cumprod[0] since
// TCD is always "trailing" // TCD is always "trailing"
float alpha_prod_t_prev = prev_timestep >= 0 ? alphas_cumprod[prev_timestep] : alphas_cumprod[0]; float alpha_prod_t_prev = static_cast<float>(prev_timestep >= 0 ? alphas_cumprod[prev_timestep] : alphas_cumprod[0]);
// The subscript _s are the only portion in this // The subscript _s are the only portion in this
// section (2) unique to TCD // section (2) unique to TCD
float alpha_prod_s = alphas_cumprod[timestep_s]; float alpha_prod_s = static_cast<float>(alphas_cumprod[timestep_s]);
float beta_prod_s = 1 - alpha_prod_s; float beta_prod_s = 1 - alpha_prod_s;
// 3. Compute the predicted noised sample x_s based on // 3. Compute the predicted noised sample x_s based on
// the model parameterization // the model parameterization

4
examples/cli/avi_writer.h
View File

@ -172,9 +172,9 @@ int create_mjpg_avi_from_sd_images(const char* filename, sd_image_t* images, int
// Write '00dc' chunk (video frame) // Write '00dc' chunk (video frame)
fwrite("00dc", 4, 1, f); fwrite("00dc", 4, 1, f);
write_u32_le(f, jpeg_data.size); write_u32_le(f, (uint32_t)jpeg_data.size);
index[i].offset = ftell(f) - 8; index[i].offset = ftell(f) - 8;
index[i].size = jpeg_data.size; index[i].size = (uint32_t)jpeg_data.size;
fwrite(jpeg_data.buf, 1, jpeg_data.size, f); fwrite(jpeg_data.buf, 1, jpeg_data.size, f);
// Align to even byte size // Align to even byte size

4
examples/common/common.hpp
View File

@ -1386,10 +1386,10 @@ struct SDGenerationParams {
if (!item.empty()) { if (!item.empty()) {
try { try {
custom_sigmas.push_back(std::stof(item)); custom_sigmas.push_back(std::stof(item));
} catch (const std::invalid_argument& e) { } catch (const std::invalid_argument&) {
LOG_ERROR("error: invalid float value '%s' in --sigmas", item.c_str()); LOG_ERROR("error: invalid float value '%s' in --sigmas", item.c_str());
return -1; return -1;
} catch (const std::out_of_range& e) { } catch (const std::out_of_range&) {
LOG_ERROR("error: float value '%s' out of range in --sigmas", item.c_str()); LOG_ERROR("error: float value '%s' out of range in --sigmas", item.c_str());
return -1; return -1;
} }

6
examples/server/main.cpp
View File

@ -44,7 +44,7 @@ inline bool is_base64(unsigned char c) {
} }
std::vector<uint8_t> base64_decode(const std::string& encoded_string) { std::vector<uint8_t> base64_decode(const std::string& encoded_string) {
int in_len = encoded_string.size(); int in_len = static_cast<int>(encoded_string.size());
int i = 0; int i = 0;
int j = 0; int j = 0;
int in_ = 0; int in_ = 0;
@ -617,7 +617,7 @@ int main(int argc, const char** argv) {
int img_h = height; int img_h = height;
uint8_t* raw_pixels = load_image_from_memory( uint8_t* raw_pixels = load_image_from_memory(
reinterpret_cast<const char*>(bytes.data()), reinterpret_cast<const char*>(bytes.data()),
bytes.size(), static_cast<int>(bytes.size()),
img_w, img_h, img_w, img_h,
width, height, 3); width, height, 3);
@ -635,7 +635,7 @@ int main(int argc, const char** argv) {
int mask_h = height; int mask_h = height;
uint8_t* mask_raw = load_image_from_memory( uint8_t* mask_raw = load_image_from_memory(
reinterpret_cast<const char*>(mask_bytes.data()), reinterpret_cast<const char*>(mask_bytes.data()),
mask_bytes.size(), static_cast<int>(mask_bytes.size()),
mask_w, mask_h, mask_w, mask_h,
width, height, 1); width, height, 1);
mask_image = {(uint32_t)mask_w, (uint32_t)mask_h, 1, mask_raw}; mask_image = {(uint32_t)mask_w, (uint32_t)mask_h, 1, mask_raw};

48
flux.hpp
View File

@ -263,7 +263,7 @@ namespace Flux {
bool use_yak_mlp = false, bool use_yak_mlp = false,
bool use_mlp_silu_act = false) bool use_mlp_silu_act = false)
: idx(idx), prune_mod(prune_mod) { : idx(idx), prune_mod(prune_mod) {
int64_t mlp_hidden_dim = hidden_size * mlp_ratio; int64_t mlp_hidden_dim = static_cast<int64_t>(hidden_size * mlp_ratio);
if (!prune_mod && !share_modulation) { if (!prune_mod && !share_modulation) {
blocks["img_mod"] = std::shared_ptr<GGMLBlock>(new Modulation(hidden_size, true)); blocks["img_mod"] = std::shared_ptr<GGMLBlock>(new Modulation(hidden_size, true));
@ -442,7 +442,7 @@ namespace Flux {
if (scale <= 0.f) { if (scale <= 0.f) {
scale = 1 / sqrt((float)head_dim); scale = 1 / sqrt((float)head_dim);
} }
mlp_hidden_dim = hidden_size * mlp_ratio; mlp_hidden_dim = static_cast<int64_t>(hidden_size * mlp_ratio);
mlp_mult_factor = 1; mlp_mult_factor = 1;
if (use_yak_mlp || use_mlp_silu_act) { if (use_yak_mlp || use_mlp_silu_act) {
mlp_mult_factor = 2; mlp_mult_factor = 2;
@ -744,9 +744,9 @@ namespace Flux {
struct ChromaRadianceParams { struct ChromaRadianceParams {
int64_t nerf_hidden_size = 64; int64_t nerf_hidden_size = 64;
int64_t nerf_mlp_ratio = 4; int nerf_mlp_ratio = 4;
int64_t nerf_depth = 4; int nerf_depth = 4;
int64_t nerf_max_freqs = 8; int nerf_max_freqs = 8;
bool use_x0 = false; bool use_x0 = false;
bool use_patch_size_32 = false; bool use_patch_size_32 = false;
}; };
@ -754,18 +754,18 @@ namespace Flux {
struct FluxParams { struct FluxParams {
SDVersion version = VERSION_FLUX; SDVersion version = VERSION_FLUX;
bool is_chroma = false; bool is_chroma = false;
int64_t patch_size = 2; int patch_size = 2;
int64_t in_channels = 64; int64_t in_channels = 64;
int64_t out_channels = 64; int64_t out_channels = 64;
int64_t vec_in_dim = 768; int64_t vec_in_dim = 768;
int64_t context_in_dim = 4096; int64_t context_in_dim = 4096;
int64_t hidden_size = 3072; int64_t hidden_size = 3072;
float mlp_ratio = 4.0f; float mlp_ratio = 4.0f;
int64_t num_heads = 24; int num_heads = 24;
int64_t depth = 19; int depth = 19;
int64_t depth_single_blocks = 38; int depth_single_blocks = 38;
std::vector<int> axes_dim = {16, 56, 56}; std::vector<int> axes_dim = {16, 56, 56};
int64_t axes_dim_sum = 128; int axes_dim_sum = 128;
int theta = 10000; int theta = 10000;
bool qkv_bias = true; bool qkv_bias = true;
bool guidance_embed = true; bool guidance_embed = true;
@ -969,7 +969,7 @@ namespace Flux {
vec = approx->forward(ctx, vec); // [344, N, hidden_size] vec = approx->forward(ctx, vec); // [344, N, hidden_size]
if (y != nullptr) { if (y != nullptr) {
txt_img_mask = ggml_pad(ctx->ggml_ctx, y, img->ne[1], 0, 0, 0); txt_img_mask = ggml_pad(ctx->ggml_ctx, y, static_cast<int>(img->ne[1]), 0, 0, 0);
} }
} else { } else {
auto time_in = std::dynamic_pointer_cast<MLPEmbedder>(blocks["time_in"]); auto time_in = std::dynamic_pointer_cast<MLPEmbedder>(blocks["time_in"]);
@ -1075,7 +1075,7 @@ namespace Flux {
int64_t W = x->ne[0]; int64_t W = x->ne[0];
int64_t H = x->ne[1]; int64_t H = x->ne[1];
int64_t C = x->ne[2]; int64_t C = x->ne[2];
int64_t patch_size = params.patch_size; int patch_size = params.patch_size;
int pad_h = (patch_size - H % patch_size) % patch_size; int pad_h = (patch_size - H % patch_size) % patch_size;
int pad_w = (patch_size - W % patch_size) % patch_size; int pad_w = (patch_size - W % patch_size) % patch_size;
@ -1149,12 +1149,12 @@ namespace Flux {
int64_t W = x->ne[0]; int64_t W = x->ne[0];
int64_t H = x->ne[1]; int64_t H = x->ne[1];
int64_t C = x->ne[2]; int64_t C = x->ne[2];
int64_t patch_size = params.patch_size; int patch_size = params.patch_size;
int pad_h = (patch_size - H % patch_size) % patch_size; int pad_h = (patch_size - H % patch_size) % patch_size;
int pad_w = (patch_size - W % patch_size) % patch_size; int pad_w = (patch_size - W % patch_size) % patch_size;
auto img = process_img(ctx, x); auto img = process_img(ctx, x);
uint64_t img_tokens = img->ne[1]; int64_t img_tokens = img->ne[1];
if (params.version == VERSION_FLUX_FILL) { if (params.version == VERSION_FLUX_FILL) {
GGML_ASSERT(c_concat != nullptr); GGML_ASSERT(c_concat != nullptr);
@ -1465,11 +1465,11 @@ namespace Flux {
txt_arange_dims = {1, 2}; txt_arange_dims = {1, 2};
} }
pe_vec = Rope::gen_flux_pe(x->ne[1], pe_vec = Rope::gen_flux_pe(static_cast<int>(x->ne[1]),
x->ne[0], static_cast<int>(x->ne[0]),
flux_params.patch_size, flux_params.patch_size,
x->ne[3], static_cast<int>(x->ne[3]),
context->ne[1], static_cast<int>(context->ne[1]),
txt_arange_dims, txt_arange_dims,
ref_latents, ref_latents,
increase_ref_index, increase_ref_index,
@ -1478,7 +1478,7 @@ namespace Flux {
circular_y_enabled, circular_y_enabled,
circular_x_enabled, circular_x_enabled,
flux_params.axes_dim); flux_params.axes_dim);
int pos_len = pe_vec.size() / flux_params.axes_dim_sum / 2; int pos_len = static_cast<int>(pe_vec.size() / flux_params.axes_dim_sum / 2);
// LOG_DEBUG("pos_len %d", pos_len); // LOG_DEBUG("pos_len %d", pos_len);
auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, flux_params.axes_dim_sum / 2, pos_len); auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, flux_params.axes_dim_sum / 2, pos_len);
// pe->data = pe_vec.data(); // pe->data = pe_vec.data();
@ -1487,8 +1487,8 @@ namespace Flux {
set_backend_tensor_data(pe, pe_vec.data()); set_backend_tensor_data(pe, pe_vec.data());
if (version == VERSION_CHROMA_RADIANCE) { if (version == VERSION_CHROMA_RADIANCE) {
int64_t patch_size = flux_params.patch_size; int patch_size = flux_params.patch_size;
int64_t nerf_max_freqs = flux_params.chroma_radiance_params.nerf_max_freqs; int nerf_max_freqs = flux_params.chroma_radiance_params.nerf_max_freqs;
dct_vec = fetch_dct_pos(patch_size, nerf_max_freqs); dct_vec = fetch_dct_pos(patch_size, nerf_max_freqs);
dct = ggml_new_tensor_2d(compute_ctx, GGML_TYPE_F32, nerf_max_freqs * nerf_max_freqs, patch_size * patch_size); dct = ggml_new_tensor_2d(compute_ctx, GGML_TYPE_F32, nerf_max_freqs * nerf_max_freqs, patch_size * patch_size);
// dct->data = dct_vec.data(); // dct->data = dct_vec.data();
@ -1577,12 +1577,12 @@ namespace Flux {
struct ggml_tensor* out = nullptr; struct ggml_tensor* out = nullptr;
int t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
compute(8, x, timesteps, context, nullptr, y, guidance, {}, false, &out, work_ctx); compute(8, x, timesteps, context, nullptr, y, guidance, {}, false, &out, work_ctx);
int t1 = ggml_time_ms(); int64_t t1 = ggml_time_ms();
print_ggml_tensor(out); print_ggml_tensor(out);
LOG_DEBUG("flux test done in %dms", t1 - t0); LOG_DEBUG("flux test done in %lldms", t1 - t0);
} }
} }

81
ggml_extend.hpp
View File

@ -99,9 +99,9 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_mul_n_mode(struct ggml_context* c
// reshape A // reshape A
// swap 0th and nth axis // swap 0th and nth axis
a = ggml_cont(ctx, ggml_permute(ctx, a, mode, mode != 1 ? 1 : 0, mode != 2 ? 2 : 0, mode != 3 ? 3 : 0)); a = ggml_cont(ctx, ggml_permute(ctx, a, mode, mode != 1 ? 1 : 0, mode != 2 ? 2 : 0, mode != 3 ? 3 : 0));
int ne1 = a->ne[1]; int64_t ne1 = a->ne[1];
int ne2 = a->ne[2]; int64_t ne2 = a->ne[2];
int ne3 = a->ne[3]; int64_t ne3 = a->ne[3];
// make 2D // make 2D
a = ggml_cont(ctx, ggml_reshape_2d(ctx, a, a->ne[0], (ne3 * ne2 * ne1))); a = ggml_cont(ctx, ggml_reshape_2d(ctx, a, a->ne[0], (ne3 * ne2 * ne1)));
@ -167,12 +167,12 @@ __STATIC_INLINE__ void ggml_ext_im_set_randn_f32(struct ggml_tensor* tensor, std
} }
} }
__STATIC_INLINE__ void ggml_ext_tensor_set_f32(struct ggml_tensor* tensor, float value, int i0, int i1 = 0, int i2 = 0, int i3 = 0) { __STATIC_INLINE__ void ggml_ext_tensor_set_f32(struct ggml_tensor* tensor, float value, int64_t i0, int64_t i1 = 0, int64_t i2 = 0, int64_t i3 = 0) {
GGML_ASSERT(tensor->nb[0] == sizeof(float)); GGML_ASSERT(tensor->nb[0] == sizeof(float));
*(float*)((char*)(tensor->data) + i3 * tensor->nb[3] + i2 * tensor->nb[2] + i1 * tensor->nb[1] + i0 * tensor->nb[0]) = value; *(float*)((char*)(tensor->data) + i3 * tensor->nb[3] + i2 * tensor->nb[2] + i1 * tensor->nb[1] + i0 * tensor->nb[0]) = value;
} }
__STATIC_INLINE__ float ggml_ext_tensor_get_f32(const ggml_tensor* tensor, int i0, int i1 = 0, int i2 = 0, int i3 = 0) { __STATIC_INLINE__ float ggml_ext_tensor_get_f32(const ggml_tensor* tensor, int64_t i0, int64_t i1 = 0, int64_t i2 = 0, int64_t i3 = 0) {
if (tensor->buffer != nullptr) { if (tensor->buffer != nullptr) {
float value; float value;
ggml_backend_tensor_get(tensor, &value, i3 * tensor->nb[3] + i2 * tensor->nb[2] + i1 * tensor->nb[1] + i0 * tensor->nb[0], sizeof(float)); ggml_backend_tensor_get(tensor, &value, i3 * tensor->nb[3] + i2 * tensor->nb[2] + i1 * tensor->nb[1] + i0 * tensor->nb[0], sizeof(float));
@ -182,9 +182,9 @@ __STATIC_INLINE__ float ggml_ext_tensor_get_f32(const ggml_tensor* tensor, int i
return *(float*)((char*)(tensor->data) + i3 * tensor->nb[3] + i2 * tensor->nb[2] + i1 * tensor->nb[1] + i0 * tensor->nb[0]); return *(float*)((char*)(tensor->data) + i3 * tensor->nb[3] + i2 * tensor->nb[2] + i1 * tensor->nb[1] + i0 * tensor->nb[0]);
} }
__STATIC_INLINE__ int ggml_ext_tensor_get_i32(const ggml_tensor* tensor, int i0, int i1 = 0, int i2 = 0, int i3 = 0) { __STATIC_INLINE__ int ggml_ext_tensor_get_i32(const ggml_tensor* tensor, int64_t i0, int64_t i1 = 0, int64_t i2 = 0, int64_t i3 = 0) {
if (tensor->buffer != nullptr) { if (tensor->buffer != nullptr) {
float value; int value;
ggml_backend_tensor_get(tensor, &value, i3 * tensor->nb[3] + i2 * tensor->nb[2] + i1 * tensor->nb[1] + i0 * tensor->nb[0], sizeof(int)); ggml_backend_tensor_get(tensor, &value, i3 * tensor->nb[3] + i2 * tensor->nb[2] + i1 * tensor->nb[1] + i0 * tensor->nb[0], sizeof(int));
return value; return value;
} }
@ -192,12 +192,12 @@ __STATIC_INLINE__ int ggml_ext_tensor_get_i32(const ggml_tensor* tensor, int i0,
return *(int*)((char*)(tensor->data) + i3 * tensor->nb[3] + i2 * tensor->nb[2] + i1 * tensor->nb[1] + i0 * tensor->nb[0]); return *(int*)((char*)(tensor->data) + i3 * tensor->nb[3] + i2 * tensor->nb[2] + i1 * tensor->nb[1] + i0 * tensor->nb[0]);
} }
__STATIC_INLINE__ ggml_fp16_t ggml_ext_tensor_get_f16(const ggml_tensor* tensor, int i0, int i1 = 0, int i2 = 0, int i3 = 0) { __STATIC_INLINE__ ggml_fp16_t ggml_ext_tensor_get_f16(const ggml_tensor* tensor, int64_t i0, int64_t i1 = 0, int64_t i2 = 0, int64_t i3 = 0) {
GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t)); GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
return *(ggml_fp16_t*)((char*)(tensor->data) + i3 * tensor->nb[3] + i2 * tensor->nb[2] + i1 * tensor->nb[1] + i0 * tensor->nb[0]); return *(ggml_fp16_t*)((char*)(tensor->data) + i3 * tensor->nb[3] + i2 * tensor->nb[2] + i1 * tensor->nb[1] + i0 * tensor->nb[0]);
} }
__STATIC_INLINE__ float sd_image_get_f32(sd_image_t image, int iw, int ih, int ic, bool scale = true) { __STATIC_INLINE__ float sd_image_get_f32(sd_image_t image, int64_t iw, int64_t ih, int64_t ic, bool scale = true) {
float value = *(image.data + ih * image.width * image.channel + iw * image.channel + ic); float value = *(image.data + ih * image.width * image.channel + iw * image.channel + ic);
if (scale) { if (scale) {
value /= 255.f; value /= 255.f;
@ -205,7 +205,7 @@ __STATIC_INLINE__ float sd_image_get_f32(sd_image_t image, int iw, int ih, int i
return value; return value;
} }
__STATIC_INLINE__ float sd_image_get_f32(sd_image_f32_t image, int iw, int ih, int ic, bool scale = true) { __STATIC_INLINE__ float sd_image_get_f32(sd_image_f32_t image, int64_t iw, int64_t ih, int64_t ic, bool scale = true) {
float value = *(image.data + ih * image.width * image.channel + iw * image.channel + ic); float value = *(image.data + ih * image.width * image.channel + iw * image.channel + ic);
if (scale) { if (scale) {
value /= 255.f; value /= 255.f;
@ -450,8 +450,8 @@ __STATIC_INLINE__ void ggml_ext_tensor_apply_mask(struct ggml_tensor* image_data
int64_t width = output->ne[0]; int64_t width = output->ne[0];
int64_t height = output->ne[1]; int64_t height = output->ne[1];
int64_t channels = output->ne[2]; int64_t channels = output->ne[2];
float rescale_mx = mask->ne[0] / output->ne[0]; float rescale_mx = 1.f * mask->ne[0] / output->ne[0];
float rescale_my = mask->ne[1] / output->ne[1]; float rescale_my = 1.f * mask->ne[1] / output->ne[1];
GGML_ASSERT(output->type == GGML_TYPE_F32); GGML_ASSERT(output->type == GGML_TYPE_F32);
for (int ix = 0; ix < width; ix++) { for (int ix = 0; ix < width; ix++) {
for (int iy = 0; iy < height; iy++) { for (int iy = 0; iy < height; iy++) {
@ -685,7 +685,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_torch_permute(struct ggml_context
__STATIC_INLINE__ struct ggml_tensor* ggml_ext_slice(struct ggml_context* ctx, __STATIC_INLINE__ struct ggml_tensor* ggml_ext_slice(struct ggml_context* ctx,
struct ggml_tensor* x, struct ggml_tensor* x,
int64_t dim, int dim,
int64_t start, int64_t start,
int64_t end) { int64_t end) {
GGML_ASSERT(dim >= 0 && dim < 4); GGML_ASSERT(dim >= 0 && dim < 4);
@ -785,7 +785,7 @@ __STATIC_INLINE__ void sd_tiling_calc_tiles(int& num_tiles_dim,
int small_dim, int small_dim,
int tile_size, int tile_size,
const float tile_overlap_factor) { const float tile_overlap_factor) {
int tile_overlap = (tile_size * tile_overlap_factor); int tile_overlap = static_cast<int>(tile_size * tile_overlap_factor);
int non_tile_overlap = tile_size - tile_overlap; int non_tile_overlap = tile_size - tile_overlap;
num_tiles_dim = (small_dim - tile_overlap) / non_tile_overlap; num_tiles_dim = (small_dim - tile_overlap) / non_tile_overlap;
@ -1346,7 +1346,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_attention_ext(struct ggml_context
// LOG_DEBUG("attention_ext L_q:%d L_k:%d n_head:%d C:%d d_head:%d N:%d", L_q, L_k, n_head, C, d_head, N); // LOG_DEBUG("attention_ext L_q:%d L_k:%d n_head:%d C:%d d_head:%d N:%d", L_q, L_k, n_head, C, d_head, N);
bool can_use_flash_attn = true; bool can_use_flash_attn = true;
if (can_use_flash_attn && L_k % 256 != 0) { if (can_use_flash_attn && L_k % 256 != 0) {
kv_pad = GGML_PAD(L_k, 256) - L_k; kv_pad = GGML_PAD(L_k, 256) - static_cast<int>(L_k);
} }
if (mask != nullptr) { if (mask != nullptr) {
@ -2361,53 +2361,6 @@ public:
} }
}; };
class Conv3dnx1x1 : public UnaryBlock {
protected:
int64_t in_channels;
int64_t out_channels;
int64_t kernel_size;
int64_t stride;
int64_t padding;
int64_t dilation;
bool bias;
void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map, const std::string prefix = "") override {
enum ggml_type wtype = GGML_TYPE_F16;
params["weight"] = ggml_new_tensor_4d(ctx, wtype, 1, kernel_size, in_channels, out_channels); // 5d => 4d
if (bias) {
enum ggml_type wtype = GGML_TYPE_F32;
params["bias"] = ggml_new_tensor_1d(ctx, wtype, out_channels);
}
}
public:
Conv3dnx1x1(int64_t in_channels,
int64_t out_channels,
int64_t kernel_size,
int64_t stride = 1,
int64_t padding = 0,
int64_t dilation = 1,
bool bias = true)
: in_channels(in_channels),
out_channels(out_channels),
kernel_size(kernel_size),
stride(stride),
padding(padding),
dilation(dilation),
bias(bias) {}
// x: [N, IC, ID, IH*IW]
// result: [N, OC, OD, OH*OW]
struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
struct ggml_tensor* w = params["weight"];
struct ggml_tensor* b = nullptr;
if (bias) {
b = params["bias"];
}
return ggml_ext_conv_3d_nx1x1(ctx->ggml_ctx, x, w, b, stride, padding, dilation);
}
};
class Conv3d : public UnaryBlock { class Conv3d : public UnaryBlock {
protected: protected:
int64_t in_channels; int64_t in_channels;
@ -2523,7 +2476,7 @@ public:
class GroupNorm : public GGMLBlock { class GroupNorm : public GGMLBlock {
protected: protected:
int64_t num_groups; int num_groups;
int64_t num_channels; int64_t num_channels;
float eps; float eps;
bool affine; bool affine;
@ -2540,7 +2493,7 @@ protected:
} }
public: public:
GroupNorm(int64_t num_groups, GroupNorm(int num_groups,
int64_t num_channels, int64_t num_channels,
float eps = 1e-05f, float eps = 1e-05f,
bool affine = true) bool affine = true)

2
gguf_reader.hpp
View File

@ -151,7 +151,7 @@ private:
} }
if (n_dims > GGML_MAX_DIMS) { if (n_dims > GGML_MAX_DIMS) {
for (int i = GGML_MAX_DIMS; i < n_dims; i++) { for (uint32_t i = GGML_MAX_DIMS; i < n_dims; i++) {
info.shape[GGML_MAX_DIMS - 1] *= info.shape[i]; // stack to last dim; info.shape[GGML_MAX_DIMS - 1] *= info.shape[i]; // stack to last dim;
} }
info.shape.resize(GGML_MAX_DIMS); info.shape.resize(GGML_MAX_DIMS);

16
latent-preview.h
View File

@ -166,12 +166,12 @@ float sd_latent_rgb_bias[3] = {-0.017478f, -0.055834f, -0.105825f};
void preview_latent_video(uint8_t* buffer, struct ggml_tensor* latents, const float (*latent_rgb_proj)[3], const float latent_rgb_bias[3], int patch_size) { void preview_latent_video(uint8_t* buffer, struct ggml_tensor* latents, const float (*latent_rgb_proj)[3], const float latent_rgb_bias[3], int patch_size) {
size_t buffer_head = 0; size_t buffer_head = 0;
uint32_t latent_width = latents->ne[0]; uint32_t latent_width = static_cast<uint32_t>(latents->ne[0]);
uint32_t latent_height = latents->ne[1]; uint32_t latent_height = static_cast<uint32_t>(latents->ne[1]);
uint32_t dim = latents->ne[ggml_n_dims(latents) - 1]; uint32_t dim = static_cast<uint32_t>(latents->ne[ggml_n_dims(latents) - 1]);
uint32_t frames = 1; uint32_t frames = 1;
if (ggml_n_dims(latents) == 4) { if (ggml_n_dims(latents) == 4) {
frames = latents->ne[2]; frames = static_cast<uint32_t>(latents->ne[2]);
} }
uint32_t rgb_width = latent_width * patch_size; uint32_t rgb_width = latent_width * patch_size;
@ -179,9 +179,9 @@ void preview_latent_video(uint8_t* buffer, struct ggml_tensor* latents, const fl
uint32_t unpatched_dim = dim / (patch_size * patch_size); uint32_t unpatched_dim = dim / (patch_size * patch_size);
for (int k = 0; k < frames; k++) { for (uint32_t k = 0; k < frames; k++) {
for (int rgb_x = 0; rgb_x < rgb_width; rgb_x++) { for (uint32_t rgb_x = 0; rgb_x < rgb_width; rgb_x++) {
for (int rgb_y = 0; rgb_y < rgb_height; rgb_y++) { for (uint32_t rgb_y = 0; rgb_y < rgb_height; rgb_y++) {
int latent_x = rgb_x / patch_size; int latent_x = rgb_x / patch_size;
int latent_y = rgb_y / patch_size; int latent_y = rgb_y / patch_size;
@ -197,7 +197,7 @@ void preview_latent_video(uint8_t* buffer, struct ggml_tensor* latents, const fl
float r = 0, g = 0, b = 0; float r = 0, g = 0, b = 0;
if (latent_rgb_proj != nullptr) { if (latent_rgb_proj != nullptr) {
for (int d = 0; d < unpatched_dim; d++) { for (uint32_t d = 0; d < unpatched_dim; d++) {
float value = *(float*)((char*)latents->data + latent_id + (d * patch_size * patch_size + channel_offset) * latents->nb[ggml_n_dims(latents) - 1]); float value = *(float*)((char*)latents->data + latent_id + (d * patch_size * patch_size + channel_offset) * latents->nb[ggml_n_dims(latents) - 1]);
r += value * latent_rgb_proj[d][0]; r += value * latent_rgb_proj[d][0];
g += value * latent_rgb_proj[d][1]; g += value * latent_rgb_proj[d][1];

110
llm.hpp
View File

@ -195,14 +195,14 @@ namespace LLM {
tokens.insert(tokens.begin(), BOS_TOKEN_ID); tokens.insert(tokens.begin(), BOS_TOKEN_ID);
} }
if (max_length > 0 && padding) { if (max_length > 0 && padding) {
size_t n = std::ceil(tokens.size() * 1.0 / max_length); size_t n = static_cast<size_t>(std::ceil(tokens.size() * 1.f / max_length));
if (n == 0) { if (n == 0) {
n = 1; n = 1;
} }
size_t length = max_length * n; size_t length = max_length * n;
LOG_DEBUG("token length: %llu", length); LOG_DEBUG("token length: %llu", length);
tokens.insert(tokens.end(), length - tokens.size(), PAD_TOKEN_ID); tokens.insert(tokens.end(), length - tokens.size(), PAD_TOKEN_ID);
weights.insert(weights.end(), length - weights.size(), 1.0); weights.insert(weights.end(), length - weights.size(), 1.f);
} }
} }
@ -377,7 +377,7 @@ namespace LLM {
try { try {
vocab = nlohmann::json::parse(vocab_utf8_str); vocab = nlohmann::json::parse(vocab_utf8_str);
} catch (const nlohmann::json::parse_error& e) { } catch (const nlohmann::json::parse_error&) {
GGML_ABORT("invalid vocab json str"); GGML_ABORT("invalid vocab json str");
} }
for (const auto& [key, value] : vocab.items()) { for (const auto& [key, value] : vocab.items()) {
@ -386,7 +386,7 @@ namespace LLM {
encoder[token] = i; encoder[token] = i;
decoder[i] = token; decoder[i] = token;
} }
encoder_len = vocab.size(); encoder_len = static_cast<int>(vocab.size());
LOG_DEBUG("vocab size: %d", encoder_len); LOG_DEBUG("vocab size: %d", encoder_len);
auto byte_unicode_pairs = bytes_to_unicode(); auto byte_unicode_pairs = bytes_to_unicode();
@ -485,16 +485,16 @@ namespace LLM {
}; };
struct LLMVisionParams { struct LLMVisionParams {
int64_t num_layers = 32; int num_layers = 32;
int64_t hidden_size = 1280; int64_t hidden_size = 1280;
int64_t intermediate_size = 3420; int64_t intermediate_size = 3420;
int64_t num_heads = 16; int num_heads = 16;
int64_t in_channels = 3; int64_t in_channels = 3;
int64_t out_hidden_size = 3584; int64_t out_hidden_size = 3584;
int64_t temporal_patch_size = 2; int temporal_patch_size = 2;
int64_t patch_size = 14; int patch_size = 14;
int64_t spatial_merge_size = 2; int spatial_merge_size = 2;
int64_t window_size = 112; int window_size = 112;
std::set<int> fullatt_block_indexes = {7, 15, 23, 31}; std::set<int> fullatt_block_indexes = {7, 15, 23, 31};
}; };
@ -503,9 +503,9 @@ namespace LLM {
int64_t num_layers = 28; int64_t num_layers = 28;
int64_t hidden_size = 3584; int64_t hidden_size = 3584;
int64_t intermediate_size = 18944; int64_t intermediate_size = 18944;
int64_t num_heads = 28; int num_heads = 28;
int64_t num_kv_heads = 4; int num_kv_heads = 4;
int64_t head_dim = 128; int head_dim = 128;
bool qkv_bias = true; bool qkv_bias = true;
bool qk_norm = false; bool qk_norm = false;
int64_t vocab_size = 152064; int64_t vocab_size = 152064;
@ -647,15 +647,15 @@ namespace LLM {
struct VisionAttention : public GGMLBlock { struct VisionAttention : public GGMLBlock {
protected: protected:
bool llama_cpp_style; bool llama_cpp_style;
int64_t head_dim; int head_dim;
int64_t num_heads; int num_heads;
public: public:
VisionAttention(bool llama_cpp_style, VisionAttention(bool llama_cpp_style,
int64_t hidden_size, int64_t hidden_size,
int64_t num_heads) int num_heads)
: llama_cpp_style(llama_cpp_style), num_heads(num_heads) { : llama_cpp_style(llama_cpp_style), num_heads(num_heads) {
head_dim = hidden_size / num_heads; head_dim = static_cast<int>(hidden_size / num_heads);
GGML_ASSERT(num_heads * head_dim == hidden_size); GGML_ASSERT(num_heads * head_dim == hidden_size);
if (llama_cpp_style) { if (llama_cpp_style) {
blocks["q_proj"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, hidden_size)); blocks["q_proj"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, hidden_size));
@ -709,7 +709,7 @@ namespace LLM {
VisionBlock(bool llama_cpp_style, VisionBlock(bool llama_cpp_style,
int64_t hidden_size, int64_t hidden_size,
int64_t intermediate_size, int64_t intermediate_size,
int64_t num_heads, int num_heads,
float eps = 1e-6f) { float eps = 1e-6f) {
blocks["attn"] = std::shared_ptr<GGMLBlock>(new VisionAttention(llama_cpp_style, hidden_size, num_heads)); blocks["attn"] = std::shared_ptr<GGMLBlock>(new VisionAttention(llama_cpp_style, hidden_size, num_heads));
blocks["mlp"] = std::shared_ptr<GGMLBlock>(new MLP(hidden_size, intermediate_size, true)); blocks["mlp"] = std::shared_ptr<GGMLBlock>(new MLP(hidden_size, intermediate_size, true));
@ -743,22 +743,22 @@ namespace LLM {
struct VisionModel : public GGMLBlock { struct VisionModel : public GGMLBlock {
protected: protected:
int64_t num_layers; int num_layers;
int64_t spatial_merge_size; int spatial_merge_size;
std::set<int> fullatt_block_indexes; std::set<int> fullatt_block_indexes;
public: public:
VisionModel(bool llama_cpp_style, VisionModel(bool llama_cpp_style,
int64_t num_layers, int num_layers,
int64_t in_channels, int64_t in_channels,
int64_t hidden_size, int64_t hidden_size,
int64_t out_hidden_size, int64_t out_hidden_size,
int64_t intermediate_size, int64_t intermediate_size,
int64_t num_heads, int num_heads,
int64_t spatial_merge_size, int spatial_merge_size,
int64_t patch_size, int patch_size,
int64_t temporal_patch_size, int temporal_patch_size,
int64_t window_size, int window_size,
std::set<int> fullatt_block_indexes = {7, 15, 23, 31}, std::set<int> fullatt_block_indexes = {7, 15, 23, 31},
float eps = 1e-6f) float eps = 1e-6f)
: num_layers(num_layers), fullatt_block_indexes(std::move(fullatt_block_indexes)), spatial_merge_size(spatial_merge_size) { : num_layers(num_layers), fullatt_block_indexes(std::move(fullatt_block_indexes)), spatial_merge_size(spatial_merge_size) {
@ -817,7 +817,7 @@ namespace LLM {
struct Attention : public GGMLBlock { struct Attention : public GGMLBlock {
protected: protected:
LLMArch arch; LLMArch arch;
int64_t head_dim; int head_dim;
int64_t num_heads; int64_t num_heads;
int64_t num_kv_heads; int64_t num_kv_heads;
bool qk_norm; bool qk_norm;
@ -1227,11 +1227,11 @@ namespace LLM {
} }
int64_t get_num_image_tokens(int64_t t, int64_t h, int64_t w) { int64_t get_num_image_tokens(int64_t t, int64_t h, int64_t w) {
int grid_t = 1; int64_t grid_t = 1;
int grid_h = h / params.vision.patch_size; int64_t grid_h = h / params.vision.patch_size;
int grid_w = w / params.vision.patch_size; int64_t grid_w = w / params.vision.patch_size;
int llm_grid_h = grid_h / params.vision.spatial_merge_size; int64_t llm_grid_h = grid_h / params.vision.spatial_merge_size;
int llm_grid_w = grid_w / params.vision.spatial_merge_size; int64_t llm_grid_w = grid_w / params.vision.spatial_merge_size;
return grid_t * grid_h * grid_w; return grid_t * grid_h * grid_w;
} }
@ -1269,8 +1269,8 @@ namespace LLM {
GGML_ASSERT(image->ne[0] % (params.vision.patch_size * params.vision.spatial_merge_size) == 0); GGML_ASSERT(image->ne[0] % (params.vision.patch_size * params.vision.spatial_merge_size) == 0);
int grid_t = 1; int grid_t = 1;
int grid_h = image->ne[1] / params.vision.patch_size; int grid_h = static_cast<int>(image->ne[1]) / params.vision.patch_size;
int grid_w = image->ne[0] / params.vision.patch_size; int grid_w = static_cast<int>(image->ne[0]) / params.vision.patch_size;
int llm_grid_h = grid_h / params.vision.spatial_merge_size; int llm_grid_h = grid_h / params.vision.spatial_merge_size;
int llm_grid_w = grid_w / params.vision.spatial_merge_size; int llm_grid_w = grid_w / params.vision.spatial_merge_size;
int vit_merger_window_size = params.vision.window_size / params.vision.patch_size / params.vision.spatial_merge_size; int vit_merger_window_size = params.vision.window_size / params.vision.patch_size / params.vision.spatial_merge_size;
@ -1358,14 +1358,14 @@ namespace LLM {
set_backend_tensor_data(window_mask, window_mask_vec.data()); set_backend_tensor_data(window_mask, window_mask_vec.data());
// pe // pe
int head_dim = params.vision.hidden_size / params.vision.num_heads; int head_dim = static_cast<int>(params.vision.hidden_size / params.vision.num_heads);
pe_vec = Rope::gen_qwen2vl_pe(grid_h, pe_vec = Rope::gen_qwen2vl_pe(grid_h,
grid_w, grid_w,
params.vision.spatial_merge_size, params.vision.spatial_merge_size,
window_inverse_index_vec, window_inverse_index_vec,
10000.f, 10000,
{head_dim / 2, head_dim / 2}); {head_dim / 2, head_dim / 2});
int pos_len = pe_vec.size() / head_dim / 2; int pos_len = static_cast<int>(pe_vec.size() / head_dim / 2);
// LOG_DEBUG("pos_len %d", pos_len); // LOG_DEBUG("pos_len %d", pos_len);
auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, head_dim / 2, pos_len); auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, head_dim / 2, pos_len);
// pe->data = pe_vec.data(); // pe->data = pe_vec.data();
@ -1485,13 +1485,13 @@ namespace LLM {
print_ggml_tensor(image, false, "image"); print_ggml_tensor(image, false, "image");
struct ggml_tensor* out = nullptr; struct ggml_tensor* out = nullptr;
int t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
model.encode_image(8, image, &out, work_ctx); model.encode_image(8, image, &out, work_ctx);
int t1 = ggml_time_ms(); int64_t t1 = ggml_time_ms();
print_ggml_tensor(out, false, "image_embed"); print_ggml_tensor(out, false, "image_embed");
image_embed = out; image_embed = out;
LOG_DEBUG("llm encode_image test done in %dms", t1 - t0); LOG_DEBUG("llm encode_image test done in %lldms", t1 - t0);
} }
std::string placeholder = "<|image_pad|>"; std::string placeholder = "<|image_pad|>";
@ -1524,12 +1524,12 @@ namespace LLM {
auto input_ids = vector_to_ggml_tensor_i32(work_ctx, tokens); auto input_ids = vector_to_ggml_tensor_i32(work_ctx, tokens);
struct ggml_tensor* out = nullptr; struct ggml_tensor* out = nullptr;
int t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
model.compute(8, input_ids, image_embeds, {}, &out, work_ctx); model.compute(8, input_ids, image_embeds, {}, &out, work_ctx);
int t1 = ggml_time_ms(); int64_t t1 = ggml_time_ms();
print_ggml_tensor(out); print_ggml_tensor(out);
LOG_DEBUG("llm test done in %dms", t1 - t0); LOG_DEBUG("llm test done in %lldms", t1 - t0);
} else if (test_vit) { } else if (test_vit) {
// auto image = ggml_new_tensor_3d(work_ctx, GGML_TYPE_F32, 280, 280, 3); // auto image = ggml_new_tensor_3d(work_ctx, GGML_TYPE_F32, 280, 280, 3);
// ggml_set_f32(image, 0.f); // ggml_set_f32(image, 0.f);
@ -1537,16 +1537,16 @@ namespace LLM {
print_ggml_tensor(image, false, "image"); print_ggml_tensor(image, false, "image");
struct ggml_tensor* out = nullptr; struct ggml_tensor* out = nullptr;
int t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
model.encode_image(8, image, &out, work_ctx); model.encode_image(8, image, &out, work_ctx);
int t1 = ggml_time_ms(); int64_t t1 = ggml_time_ms();
print_ggml_tensor(out, false, "out"); print_ggml_tensor(out, false, "out");
// auto ref_out = load_tensor_from_file(work_ctx, "qwen2vl.bin"); // auto ref_out = load_tensor_from_file(work_ctx, "qwen2vl.bin");
// ggml_ext_tensor_diff(ref_out, out, 0.01f); // ggml_ext_tensor_diff(ref_out, out, 0.01f);
LOG_DEBUG("llm test done in %dms", t1 - t0); LOG_DEBUG("llm test done in %lldms", t1 - t0);
} else if (test_mistral) { } else if (test_mistral) {
std::pair<int, int> prompt_attn_range; std::pair<int, int> prompt_attn_range;
std::string text = "[SYSTEM_PROMPT]You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object\nattribution and actions without speculation.[/SYSTEM_PROMPT][INST]"; std::string text = "[SYSTEM_PROMPT]You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object\nattribution and actions without speculation.[/SYSTEM_PROMPT][INST]";
@ -1564,12 +1564,12 @@ namespace LLM {
auto input_ids = vector_to_ggml_tensor_i32(work_ctx, tokens); auto input_ids = vector_to_ggml_tensor_i32(work_ctx, tokens);
struct ggml_tensor* out = nullptr; struct ggml_tensor* out = nullptr;
int t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
model.compute(8, input_ids, {}, {10, 20, 30}, &out, work_ctx); model.compute(8, input_ids, {}, {10, 20, 30}, &out, work_ctx);
int t1 = ggml_time_ms(); int64_t t1 = ggml_time_ms();
print_ggml_tensor(out); print_ggml_tensor(out);
LOG_DEBUG("llm test done in %dms", t1 - t0); LOG_DEBUG("llm test done in %lldms", t1 - t0);
} else if (test_qwen3) { } else if (test_qwen3) {
std::pair<int, int> prompt_attn_range; std::pair<int, int> prompt_attn_range;
std::string text = "<|im_start|>user\n"; std::string text = "<|im_start|>user\n";
@ -1587,12 +1587,12 @@ namespace LLM {
auto input_ids = vector_to_ggml_tensor_i32(work_ctx, tokens); auto input_ids = vector_to_ggml_tensor_i32(work_ctx, tokens);
struct ggml_tensor* out = nullptr; struct ggml_tensor* out = nullptr;
int t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
model.compute(8, input_ids, {}, {35}, &out, work_ctx); model.compute(8, input_ids, {}, {35}, &out, work_ctx);
int t1 = ggml_time_ms(); int64_t t1 = ggml_time_ms();
print_ggml_tensor(out); print_ggml_tensor(out);
LOG_DEBUG("llm test done in %dms", t1 - t0); LOG_DEBUG("llm test done in %lldms", t1 - t0);
} else { } else {
std::pair<int, int> prompt_attn_range; std::pair<int, int> prompt_attn_range;
std::string text = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n"; std::string text = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n";
@ -1610,12 +1610,12 @@ namespace LLM {
auto input_ids = vector_to_ggml_tensor_i32(work_ctx, tokens); auto input_ids = vector_to_ggml_tensor_i32(work_ctx, tokens);
struct ggml_tensor* out = nullptr; struct ggml_tensor* out = nullptr;
int t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
model.compute(8, input_ids, {}, {}, &out, work_ctx); model.compute(8, input_ids, {}, {}, &out, work_ctx);
int t1 = ggml_time_ms(); int64_t t1 = ggml_time_ms();
print_ggml_tensor(out); print_ggml_tensor(out);
LOG_DEBUG("llm test done in %dms", t1 - t0); LOG_DEBUG("llm test done in %lldms", t1 - t0);
} }
} }

20
mmdit.hpp
View File

@ -97,11 +97,11 @@ public:
struct TimestepEmbedder : public GGMLBlock { struct TimestepEmbedder : public GGMLBlock {
// Embeds scalar timesteps into vector representations. // Embeds scalar timesteps into vector representations.
protected: protected:
int64_t frequency_embedding_size; int frequency_embedding_size;
public: public:
TimestepEmbedder(int64_t hidden_size, TimestepEmbedder(int64_t hidden_size,
int64_t frequency_embedding_size = 256, int frequency_embedding_size = 256,
int64_t out_channels = 0) int64_t out_channels = 0)
: frequency_embedding_size(frequency_embedding_size) { : frequency_embedding_size(frequency_embedding_size) {
if (out_channels <= 0) { if (out_channels <= 0) {
@ -167,11 +167,11 @@ public:
blocks["proj"] = std::shared_ptr<GGMLBlock>(new Linear(dim, dim)); blocks["proj"] = std::shared_ptr<GGMLBlock>(new Linear(dim, dim));
} }
if (qk_norm == "rms") { if (qk_norm == "rms") {
blocks["ln_q"] = std::shared_ptr<GGMLBlock>(new RMSNorm(d_head, 1.0e-6)); blocks["ln_q"] = std::shared_ptr<GGMLBlock>(new RMSNorm(d_head, 1.0e-6f));
blocks["ln_k"] = std::shared_ptr<GGMLBlock>(new RMSNorm(d_head, 1.0e-6)); blocks["ln_k"] = std::shared_ptr<GGMLBlock>(new RMSNorm(d_head, 1.0e-6f));
} else if (qk_norm == "ln") { } else if (qk_norm == "ln") {
blocks["ln_q"] = std::shared_ptr<GGMLBlock>(new LayerNorm(d_head, 1.0e-6)); blocks["ln_q"] = std::shared_ptr<GGMLBlock>(new LayerNorm(d_head, 1.0e-6f));
blocks["ln_k"] = std::shared_ptr<GGMLBlock>(new LayerNorm(d_head, 1.0e-6)); blocks["ln_k"] = std::shared_ptr<GGMLBlock>(new LayerNorm(d_head, 1.0e-6f));
} }
} }
@ -623,7 +623,7 @@ struct MMDiT : public GGMLBlock {
// Diffusion model with a Transformer backbone. // Diffusion model with a Transformer backbone.
protected: protected:
int64_t input_size = -1; int64_t input_size = -1;
int64_t patch_size = 2; int patch_size = 2;
int64_t in_channels = 16; int64_t in_channels = 16;
int64_t d_self = -1; // >=0 for MMdiT-X int64_t d_self = -1; // >=0 for MMdiT-X
int64_t depth = 24; int64_t depth = 24;
@ -943,12 +943,12 @@ struct MMDiTRunner : public GGMLRunner {
struct ggml_tensor* out = nullptr; struct ggml_tensor* out = nullptr;
int t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
compute(8, x, timesteps, context, y, &out, work_ctx); compute(8, x, timesteps, context, y, &out, work_ctx);
int t1 = ggml_time_ms(); int64_t t1 = ggml_time_ms();
print_ggml_tensor(out); print_ggml_tensor(out);
LOG_DEBUG("mmdit test done in %dms", t1 - t0); LOG_DEBUG("mmdit test done in %lldms", t1 - t0);
} }
} }

8
model.cpp
View File

@ -436,7 +436,7 @@ bool ModelLoader::init_from_gguf_file(const std::string& file_path, const std::s
name, name,
gguf_tensor_info.type, gguf_tensor_info.type,
gguf_tensor_info.shape.data(), gguf_tensor_info.shape.data(),
gguf_tensor_info.shape.size(), static_cast<int>(gguf_tensor_info.shape.size()),
file_index, file_index,
data_offset + gguf_tensor_info.offset); data_offset + gguf_tensor_info.offset);
@ -448,7 +448,7 @@ bool ModelLoader::init_from_gguf_file(const std::string& file_path, const std::s
return true; return true;
} }
int n_tensors = gguf_get_n_tensors(ctx_gguf_); int n_tensors = static_cast<int>(gguf_get_n_tensors(ctx_gguf_));
size_t total_size = 0; size_t total_size = 0;
size_t data_offset = gguf_get_data_offset(ctx_gguf_); size_t data_offset = gguf_get_data_offset(ctx_gguf_);
@ -1570,7 +1570,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
break; break;
} }
size_t curr_num = total_tensors_processed + current_idx; size_t curr_num = total_tensors_processed + current_idx;
pretty_progress(curr_num, total_tensors_to_process, (ggml_time_ms() - t_start) / 1000.0f / (curr_num + 1e-6f)); pretty_progress(static_cast<int>(curr_num), static_cast<int>(total_tensors_to_process), (ggml_time_ms() - t_start) / 1000.0f / (curr_num + 1e-6f));
std::this_thread::sleep_for(std::chrono::milliseconds(200)); std::this_thread::sleep_for(std::chrono::milliseconds(200));
} }
@ -1583,7 +1583,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
break; break;
} }
total_tensors_processed += file_tensors.size(); total_tensors_processed += file_tensors.size();
pretty_progress(total_tensors_processed, total_tensors_to_process, (ggml_time_ms() - t_start) / 1000.0f / (total_tensors_processed + 1e-6f)); pretty_progress(static_cast<int>(total_tensors_processed), static_cast<int>(total_tensors_to_process), (ggml_time_ms() - t_start) / 1000.0f / (total_tensors_processed + 1e-6f));
if (total_tensors_processed < total_tensors_to_process) { if (total_tensors_processed < total_tensors_to_process) {
printf("\n"); printf("\n");
} }

2
pmid.hpp
View File

@ -72,7 +72,7 @@ struct PerceiverAttention : public GGMLBlock {
int heads; // = heads int heads; // = heads
public: public:
PerceiverAttention(int dim, int dim_h = 64, int h = 8) PerceiverAttention(int dim, int dim_h = 64, int h = 8)
: scale(powf(dim_h, -0.5)), dim_head(dim_h), heads(h) { : scale(powf(static_cast<float>(dim_h), -0.5f)), dim_head(dim_h), heads(h) {
int inner_dim = dim_head * heads; int inner_dim = dim_head * heads;
blocks["norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim)); blocks["norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
blocks["norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim)); blocks["norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));

20
preprocessing.hpp
View File

@ -2,7 +2,7 @@
#define __PREPROCESSING_HPP__ #define __PREPROCESSING_HPP__
#include "ggml_extend.hpp" #include "ggml_extend.hpp"
#define M_PI_ 3.14159265358979323846 #define M_PI_ 3.14159265358979323846f
void convolve(struct ggml_tensor* input, struct ggml_tensor* output, struct ggml_tensor* kernel, int padding) { void convolve(struct ggml_tensor* input, struct ggml_tensor* output, struct ggml_tensor* kernel, int padding) {
struct ggml_init_params params; struct ggml_init_params params;
@ -20,13 +20,13 @@ void convolve(struct ggml_tensor* input, struct ggml_tensor* output, struct ggml
} }
void gaussian_kernel(struct ggml_tensor* kernel) { void gaussian_kernel(struct ggml_tensor* kernel) {
int ks_mid = kernel->ne[0] / 2; int ks_mid = static_cast<int>(kernel->ne[0] / 2);
float sigma = 1.4f; float sigma = 1.4f;
float normal = 1.f / (2.0f * M_PI_ * powf(sigma, 2.0f)); float normal = 1.f / (2.0f * M_PI_ * powf(sigma, 2.0f));
for (int y = 0; y < kernel->ne[0]; y++) { for (int y = 0; y < kernel->ne[0]; y++) {
float gx = -ks_mid + y; float gx = static_cast<float>(-ks_mid + y);
for (int x = 0; x < kernel->ne[1]; x++) { for (int x = 0; x < kernel->ne[1]; x++) {
float gy = -ks_mid + x; float gy = static_cast<float>(-ks_mid + x);
float k_ = expf(-((gx * gx + gy * gy) / (2.0f * powf(sigma, 2.0f)))) * normal; float k_ = expf(-((gx * gx + gy * gy) / (2.0f * powf(sigma, 2.0f)))) * normal;
ggml_ext_tensor_set_f32(kernel, k_, x, y); ggml_ext_tensor_set_f32(kernel, k_, x, y);
} }
@ -46,7 +46,7 @@ void grayscale(struct ggml_tensor* rgb_img, struct ggml_tensor* grayscale) {
} }
void prop_hypot(struct ggml_tensor* x, struct ggml_tensor* y, struct ggml_tensor* h) { void prop_hypot(struct ggml_tensor* x, struct ggml_tensor* y, struct ggml_tensor* h) {
int n_elements = ggml_nelements(h); int n_elements = static_cast<int>(ggml_nelements(h));
float* dx = (float*)x->data; float* dx = (float*)x->data;
float* dy = (float*)y->data; float* dy = (float*)y->data;
float* dh = (float*)h->data; float* dh = (float*)h->data;
@ -56,7 +56,7 @@ void prop_hypot(struct ggml_tensor* x, struct ggml_tensor* y, struct ggml_tensor
} }
void prop_arctan2(struct ggml_tensor* x, struct ggml_tensor* y, struct ggml_tensor* h) { void prop_arctan2(struct ggml_tensor* x, struct ggml_tensor* y, struct ggml_tensor* h) {
int n_elements = ggml_nelements(h); int n_elements = static_cast<int>(ggml_nelements(h));
float* dx = (float*)x->data; float* dx = (float*)x->data;
float* dy = (float*)y->data; float* dy = (float*)y->data;
float* dh = (float*)h->data; float* dh = (float*)h->data;
@ -66,7 +66,7 @@ void prop_arctan2(struct ggml_tensor* x, struct ggml_tensor* y, struct ggml_tens
} }
void normalize_tensor(struct ggml_tensor* g) { void normalize_tensor(struct ggml_tensor* g) {
int n_elements = ggml_nelements(g); int n_elements = static_cast<int>(ggml_nelements(g));
float* dg = (float*)g->data; float* dg = (float*)g->data;
float max = -INFINITY; float max = -INFINITY;
for (int i = 0; i < n_elements; i++) { for (int i = 0; i < n_elements; i++) {
@ -118,7 +118,7 @@ void non_max_supression(struct ggml_tensor* result, struct ggml_tensor* G, struc
} }
void threshold_hystersis(struct ggml_tensor* img, float high_threshold, float low_threshold, float weak, float strong) { void threshold_hystersis(struct ggml_tensor* img, float high_threshold, float low_threshold, float weak, float strong) {
int n_elements = ggml_nelements(img); int n_elements = static_cast<int>(ggml_nelements(img));
float* imd = (float*)img->data; float* imd = (float*)img->data;
float max = -INFINITY; float max = -INFINITY;
for (int i = 0; i < n_elements; i++) { for (int i = 0; i < n_elements; i++) {
@ -209,8 +209,8 @@ bool preprocess_canny(sd_image_t img, float high_threshold, float low_threshold,
non_max_supression(image_gray, G, tetha); non_max_supression(image_gray, G, tetha);
threshold_hystersis(image_gray, high_threshold, low_threshold, weak, strong); threshold_hystersis(image_gray, high_threshold, low_threshold, weak, strong);
// to RGB channels // to RGB channels
for (int iy = 0; iy < img.height; iy++) { for (uint32_t iy = 0; iy < img.height; iy++) {
for (int ix = 0; ix < img.width; ix++) { for (uint32_t ix = 0; ix < img.width; ix++) {
float gray = ggml_ext_tensor_get_f32(image_gray, ix, iy); float gray = ggml_ext_tensor_get_f32(image_gray, ix, iy);
gray = inverse ? 1.0f - gray : gray; gray = inverse ? 1.0f - gray : gray;
ggml_ext_tensor_set_f32(image, gray, ix, iy); ggml_ext_tensor_set_f32(image, gray, ix, iy);

26
qwen_image.hpp
View File

@ -350,16 +350,16 @@ namespace Qwen {
}; };
struct QwenImageParams { struct QwenImageParams {
int64_t patch_size = 2; int patch_size = 2;
int64_t in_channels = 64; int64_t in_channels = 64;
int64_t out_channels = 16; int64_t out_channels = 16;
int64_t num_layers = 60; int num_layers = 60;
int64_t attention_head_dim = 128; int64_t attention_head_dim = 128;
int64_t num_attention_heads = 24; int64_t num_attention_heads = 24;
int64_t joint_attention_dim = 3584; int64_t joint_attention_dim = 3584;
float theta = 10000; int theta = 10000;
std::vector<int> axes_dim = {16, 56, 56}; std::vector<int> axes_dim = {16, 56, 56};
int64_t axes_dim_sum = 128; int axes_dim_sum = 128;
bool zero_cond_t = false; bool zero_cond_t = false;
}; };
@ -514,7 +514,7 @@ namespace Qwen {
int64_t N = x->ne[3]; int64_t N = x->ne[3];
auto img = process_img(ctx, x); auto img = process_img(ctx, x);
uint64_t img_tokens = img->ne[1]; int64_t img_tokens = img->ne[1];
if (ref_latents.size() > 0) { if (ref_latents.size() > 0) {
for (ggml_tensor* ref : ref_latents) { for (ggml_tensor* ref : ref_latents) {
@ -613,18 +613,18 @@ namespace Qwen {
ref_latents[i] = to_backend(ref_latents[i]); ref_latents[i] = to_backend(ref_latents[i]);
} }
pe_vec = Rope::gen_qwen_image_pe(x->ne[1], pe_vec = Rope::gen_qwen_image_pe(static_cast<int>(x->ne[1]),
x->ne[0], static_cast<int>(x->ne[0]),
qwen_image_params.patch_size, qwen_image_params.patch_size,
x->ne[3], static_cast<int>(x->ne[3]),
context->ne[1], static_cast<int>(context->ne[1]),
ref_latents, ref_latents,
increase_ref_index, increase_ref_index,
qwen_image_params.theta, qwen_image_params.theta,
circular_y_enabled, circular_y_enabled,
circular_x_enabled, circular_x_enabled,
qwen_image_params.axes_dim); qwen_image_params.axes_dim);
int pos_len = pe_vec.size() / qwen_image_params.axes_dim_sum / 2; int pos_len = static_cast<int>(pe_vec.size() / qwen_image_params.axes_dim_sum / 2);
// LOG_DEBUG("pos_len %d", pos_len); // LOG_DEBUG("pos_len %d", pos_len);
auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, qwen_image_params.axes_dim_sum / 2, pos_len); auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, qwen_image_params.axes_dim_sum / 2, pos_len);
// pe->data = pe_vec.data(); // pe->data = pe_vec.data();
@ -715,12 +715,12 @@ namespace Qwen {
struct ggml_tensor* out = nullptr; struct ggml_tensor* out = nullptr;
int t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
compute(8, x, timesteps, context, {}, false, &out, work_ctx); compute(8, x, timesteps, context, {}, false, &out, work_ctx);
int t1 = ggml_time_ms(); int64_t t1 = ggml_time_ms();
print_ggml_tensor(out); print_ggml_tensor(out);
LOG_DEBUG("qwen_image test done in %dms", t1 - t0); LOG_DEBUG("qwen_image test done in %lldms", t1 - t0);
} }
} }

2
rng_mt19937.hpp
View File

@ -90,7 +90,7 @@ class MT19937RNG : public RNG {
float u1 = 1.0f - data[j]; float u1 = 1.0f - data[j];
float u2 = data[j + 8]; float u2 = data[j + 8];
float r = std::sqrt(-2.0f * std::log(u1)); float r = std::sqrt(-2.0f * std::log(u1));
float theta = 2.0f * 3.14159265358979323846 * u2; float theta = 2.0f * 3.14159265358979323846f * u2;
data[j] = r * std::cos(theta) * std + mean; data[j] = r * std::cos(theta) * std + mean;
data[j + 8] = r * std::sin(theta) * std + mean; data[j + 8] = r * std::sin(theta) * std + mean;
} }

58
rope.hpp
View File

@ -22,11 +22,11 @@ namespace Rope {
} }
__STATIC_INLINE__ std::vector<std::vector<float>> transpose(const std::vector<std::vector<float>>& mat) { __STATIC_INLINE__ std::vector<std::vector<float>> transpose(const std::vector<std::vector<float>>& mat) {
int rows = mat.size(); size_t rows = mat.size();
int cols = mat[0].size(); size_t cols = mat[0].size();
std::vector<std::vector<float>> transposed(cols, std::vector<float>(rows)); std::vector<std::vector<float>> transposed(cols, std::vector<float>(rows));
for (int i = 0; i < rows; ++i) { for (size_t i = 0; i < rows; ++i) {
for (int j = 0; j < cols; ++j) { for (size_t j = 0; j < cols; ++j) {
transposed[j][i] = mat[i][j]; transposed[j][i] = mat[i][j];
} }
} }
@ -52,13 +52,13 @@ namespace Rope {
std::vector<float> omega(half_dim); std::vector<float> omega(half_dim);
for (int i = 0; i < half_dim; ++i) { for (int i = 0; i < half_dim; ++i) {
omega[i] = 1.0f / std::pow(theta, scale[i]); omega[i] = 1.0f / ::powf(1.f * theta, scale[i]);
} }
int pos_size = pos.size(); size_t pos_size = pos.size();
std::vector<std::vector<float>> out(pos_size, std::vector<float>(half_dim)); std::vector<std::vector<float>> out(pos_size, std::vector<float>(half_dim));
for (int i = 0; i < pos_size; ++i) { for (size_t i = 0; i < pos_size; ++i) {
for (int j = 0; j < half_dim; ++j) { for (size_t j = 0; j < half_dim; ++j) {
float angle = pos[i] * omega[j]; float angle = pos[i] * omega[j];
if (!axis_wrap_dims.empty()) { if (!axis_wrap_dims.empty()) {
size_t wrap_size = axis_wrap_dims.size(); size_t wrap_size = axis_wrap_dims.size();
@ -99,7 +99,7 @@ namespace Rope {
for (int dim = 0; dim < axes_dim_num; dim++) { for (int dim = 0; dim < axes_dim_num; dim++) {
if (arange_dims.find(dim) != arange_dims.end()) { if (arange_dims.find(dim) != arange_dims.end()) {
for (int i = 0; i < bs * context_len; i++) { for (int i = 0; i < bs * context_len; i++) {
txt_ids[i][dim] = (i % context_len); txt_ids[i][dim] = 1.f * (i % context_len);
} }
} }
} }
@ -128,12 +128,12 @@ namespace Rope {
w_start -= w_len / 2; w_start -= w_len / 2;
} }
std::vector<float> row_ids = linspace<float>(h_start, h_start + h_len - 1, h_len); std::vector<float> row_ids = linspace<float>(1.f * h_start, 1.f * h_start + h_len - 1, h_len);
std::vector<float> col_ids = linspace<float>(w_start, w_start + w_len - 1, w_len); std::vector<float> col_ids = linspace<float>(1.f * w_start, 1.f * w_start + w_len - 1, w_len);
for (int i = 0; i < h_len; ++i) { for (int i = 0; i < h_len; ++i) {
for (int j = 0; j < w_len; ++j) { for (int j = 0; j < w_len; ++j) {
img_ids[i * w_len + j][0] = index; img_ids[i * w_len + j][0] = 1.f * index;
img_ids[i * w_len + j][1] = row_ids[i]; img_ids[i * w_len + j][1] = row_ids[i];
img_ids[i * w_len + j][2] = col_ids[j]; img_ids[i * w_len + j][2] = col_ids[j];
} }
@ -172,7 +172,7 @@ namespace Rope {
const std::vector<std::vector<int>>& wrap_dims = {}) { const std::vector<std::vector<int>>& wrap_dims = {}) {
std::vector<std::vector<float>> trans_ids = transpose(ids); std::vector<std::vector<float>> trans_ids = transpose(ids);
size_t pos_len = ids.size() / bs; size_t pos_len = ids.size() / bs;
int num_axes = axes_dim.size(); size_t num_axes = axes_dim.size();
// for (int i = 0; i < pos_len; i++) { // for (int i = 0; i < pos_len; i++) {
// std::cout << trans_ids[0][i] << " " << trans_ids[1][i] << " " << trans_ids[2][i] << std::endl; // std::cout << trans_ids[0][i] << " " << trans_ids[1][i] << " " << trans_ids[2][i] << std::endl;
// } // }
@ -182,8 +182,8 @@ namespace Rope {
emb_dim += d / 2; emb_dim += d / 2;
std::vector<std::vector<float>> emb(bs * pos_len, std::vector<float>(emb_dim * 2 * 2, 0.0)); std::vector<std::vector<float>> emb(bs * pos_len, std::vector<float>(emb_dim * 2 * 2, 0.0));
int offset = 0; size_t offset = 0;
for (int i = 0; i < num_axes; ++i) { for (size_t i = 0; i < num_axes; ++i) {
std::vector<int> axis_wrap_dims; std::vector<int> axis_wrap_dims;
if (!wrap_dims.empty() && i < (int)wrap_dims.size()) { if (!wrap_dims.empty() && i < (int)wrap_dims.size()) {
axis_wrap_dims = wrap_dims[i]; axis_wrap_dims = wrap_dims[i];
@ -211,12 +211,12 @@ namespace Rope {
float ref_index_scale, float ref_index_scale,
bool scale_rope) { bool scale_rope) {
std::vector<std::vector<float>> ids; std::vector<std::vector<float>> ids;
uint64_t curr_h_offset = 0; int curr_h_offset = 0;
uint64_t curr_w_offset = 0; int curr_w_offset = 0;
int index = 1; int index = 1;
for (ggml_tensor* ref : ref_latents) { for (ggml_tensor* ref : ref_latents) {
uint64_t h_offset = 0; int h_offset = 0;
uint64_t w_offset = 0; int w_offset = 0;
if (!increase_ref_index) { if (!increase_ref_index) {
if (ref->ne[1] + curr_h_offset > ref->ne[0] + curr_w_offset) { if (ref->ne[1] + curr_h_offset > ref->ne[0] + curr_w_offset) {
w_offset = curr_w_offset; w_offset = curr_w_offset;
@ -226,8 +226,8 @@ namespace Rope {
scale_rope = false; scale_rope = false;
} }
auto ref_ids = gen_flux_img_ids(ref->ne[1], auto ref_ids = gen_flux_img_ids(static_cast<int>(ref->ne[1]),
ref->ne[0], static_cast<int>(ref->ne[0]),
patch_size, patch_size,
bs, bs,
axes_dim_num, axes_dim_num,
@ -241,8 +241,8 @@ namespace Rope {
index++; index++;
} }
curr_h_offset = std::max(curr_h_offset, ref->ne[1] + h_offset); curr_h_offset = std::max(curr_h_offset, static_cast<int>(ref->ne[1]) + h_offset);
curr_w_offset = std::max(curr_w_offset, ref->ne[0] + w_offset); curr_w_offset = std::max(curr_w_offset, static_cast<int>(ref->ne[0]) + w_offset);
} }
return ids; return ids;
} }
@ -345,7 +345,7 @@ namespace Rope {
int h_len = (h + (patch_size / 2)) / patch_size; int h_len = (h + (patch_size / 2)) / patch_size;
int w_len = (w + (patch_size / 2)) / patch_size; int w_len = (w + (patch_size / 2)) / patch_size;
int txt_id_start = std::max(h_len, w_len); int txt_id_start = std::max(h_len, w_len);
auto txt_ids = linspace<float>(txt_id_start, context_len + txt_id_start, context_len); auto txt_ids = linspace<float>(1.f * txt_id_start, 1.f * context_len + txt_id_start, context_len);
std::vector<std::vector<float>> txt_ids_repeated(bs * context_len, std::vector<float>(3)); std::vector<std::vector<float>> txt_ids_repeated(bs * context_len, std::vector<float>(3));
for (int i = 0; i < bs; ++i) { for (int i = 0; i < bs; ++i) {
for (int j = 0; j < txt_ids.size(); ++j) { for (int j = 0; j < txt_ids.size(); ++j) {
@ -440,9 +440,9 @@ namespace Rope {
std::vector<std::vector<float>> vid_ids(t_len * h_len * w_len, std::vector<float>(3, 0.0)); std::vector<std::vector<float>> vid_ids(t_len * h_len * w_len, std::vector<float>(3, 0.0));
std::vector<float> t_ids = linspace<float>(t_offset, t_len - 1 + t_offset, t_len); std::vector<float> t_ids = linspace<float>(1.f * t_offset, 1.f * t_len - 1 + t_offset, t_len);
std::vector<float> h_ids = linspace<float>(h_offset, h_len - 1 + h_offset, h_len); std::vector<float> h_ids = linspace<float>(1.f * h_offset, 1.f * h_len - 1 + h_offset, h_len);
std::vector<float> w_ids = linspace<float>(w_offset, w_len - 1 + w_offset, w_len); std::vector<float> w_ids = linspace<float>(1.f * w_offset, 1.f * w_len - 1 + w_offset, w_len);
for (int i = 0; i < t_len; ++i) { for (int i = 0; i < t_len; ++i) {
for (int j = 0; j < h_len; ++j) { for (int j = 0; j < h_len; ++j) {
@ -493,8 +493,8 @@ namespace Rope {
GGML_ASSERT(i < grid_h * grid_w); GGML_ASSERT(i < grid_h * grid_w);
ids[i][0] = ih + iy; ids[i][0] = static_cast<float>(ih + iy);
ids[i][1] = iw + ix; ids[i][1] = static_cast<float>(iw + ix);
index++; index++;
} }
} }

74
stable-diffusion.cpp
View File

@ -534,7 +534,7 @@ public:
version); version);
} else { // SD1.x SD2.x SDXL } else { // SD1.x SD2.x SDXL
std::map<std::string, std::string> embbeding_map; std::map<std::string, std::string> embbeding_map;
for (int i = 0; i < sd_ctx_params->embedding_count; i++) { for (uint32_t i = 0; i < sd_ctx_params->embedding_count; i++) {
embbeding_map.emplace(SAFE_STR(sd_ctx_params->embeddings[i].name), SAFE_STR(sd_ctx_params->embeddings[i].path)); embbeding_map.emplace(SAFE_STR(sd_ctx_params->embeddings[i].name), SAFE_STR(sd_ctx_params->embeddings[i].path));
} }
if (strstr(SAFE_STR(sd_ctx_params->photo_maker_path), "v2")) { if (strstr(SAFE_STR(sd_ctx_params->photo_maker_path), "v2")) {
@ -1191,7 +1191,7 @@ public:
void apply_loras(const sd_lora_t* loras, uint32_t lora_count) { void apply_loras(const sd_lora_t* loras, uint32_t lora_count) {
std::unordered_map<std::string, float> lora_f2m; std::unordered_map<std::string, float> lora_f2m;
for (int i = 0; i < lora_count; i++) { for (uint32_t i = 0; i < lora_count; i++) {
std::string lora_id = SAFE_STR(loras[i].path); std::string lora_id = SAFE_STR(loras[i].path);
if (loras[i].is_high_noise) { if (loras[i].is_high_noise) {
lora_id = "|high_noise|" + lora_id; lora_id = "|high_noise|" + lora_id;
@ -1443,12 +1443,12 @@ public:
void* step_callback_data, void* step_callback_data,
bool is_noisy) { bool is_noisy) {
const uint32_t channel = 3; const uint32_t channel = 3;
uint32_t width = latents->ne[0]; uint32_t width = static_cast<uint32_t>(latents->ne[0]);
uint32_t height = latents->ne[1]; uint32_t height = static_cast<uint32_t>(latents->ne[1]);
uint32_t dim = latents->ne[ggml_n_dims(latents) - 1]; uint32_t dim = static_cast<uint32_t>(latents->ne[ggml_n_dims(latents) - 1]);
if (preview_mode == PREVIEW_PROJ) { if (preview_mode == PREVIEW_PROJ) {
int64_t patch_sz = 1; int patch_sz = 1;
const float(*latent_rgb_proj)[channel] = nullptr; const float(*latent_rgb_proj)[channel] = nullptr;
float* latent_rgb_bias = nullptr; float* latent_rgb_bias = nullptr;
@ -1508,7 +1508,7 @@ public:
uint32_t frames = 1; uint32_t frames = 1;
if (ggml_n_dims(latents) == 4) { if (ggml_n_dims(latents) == 4) {
frames = latents->ne[2]; frames = static_cast<uint32_t>(latents->ne[2]);
} }
uint32_t img_width = width * patch_sz; uint32_t img_width = width * patch_sz;
@ -1518,7 +1518,7 @@ public:
preview_latent_video(data, latents, latent_rgb_proj, latent_rgb_bias, patch_sz); preview_latent_video(data, latents, latent_rgb_proj, latent_rgb_bias, patch_sz);
sd_image_t* images = (sd_image_t*)malloc(frames * sizeof(sd_image_t)); sd_image_t* images = (sd_image_t*)malloc(frames * sizeof(sd_image_t));
for (int i = 0; i < frames; i++) { for (uint32_t i = 0; i < frames; i++) {
images[i] = {img_width, img_height, channel, data + i * img_width * img_height * channel}; images[i] = {img_width, img_height, channel, data + i * img_width * img_height * channel};
} }
step_callback(step, frames, images, is_noisy, step_callback_data); step_callback(step, frames, images, is_noisy, step_callback_data);
@ -1563,22 +1563,22 @@ public:
ggml_ext_tensor_clamp_inplace(result, 0.0f, 1.0f); ggml_ext_tensor_clamp_inplace(result, 0.0f, 1.0f);
uint32_t frames = 1; uint32_t frames = 1;
if (ggml_n_dims(latents) == 4) { if (ggml_n_dims(latents) == 4) {
frames = result->ne[2]; frames = static_cast<uint32_t>(result->ne[2]);
} }
sd_image_t* images = (sd_image_t*)malloc(frames * sizeof(sd_image_t)); sd_image_t* images = (sd_image_t*)malloc(frames * sizeof(sd_image_t));
// print_ggml_tensor(result,true); // print_ggml_tensor(result,true);
for (size_t i = 0; i < frames; i++) { for (size_t i = 0; i < frames; i++) {
images[i].width = result->ne[0]; images[i].width = static_cast<uint32_t>(result->ne[0]);
images[i].height = result->ne[1]; images[i].height = static_cast<uint32_t>(result->ne[1]);
images[i].channel = 3; images[i].channel = 3;
images[i].data = ggml_tensor_to_sd_image(result, i, ggml_n_dims(latents) == 4); images[i].data = ggml_tensor_to_sd_image(result, static_cast<int>(i), ggml_n_dims(latents) == 4);
} }
step_callback(step, frames, images, is_noisy, step_callback_data); step_callback(step, frames, images, is_noisy, step_callback_data);
ggml_ext_tensor_scale_inplace(result, 0); ggml_ext_tensor_scale_inplace(result, 0);
for (int i = 0; i < frames; i++) { for (uint32_t i = 0; i < frames; i++) {
free(images[i].data); free(images[i].data);
} }
@ -1800,7 +1800,7 @@ public:
int64_t H = x->ne[1] * get_vae_scale_factor(); int64_t H = x->ne[1] * get_vae_scale_factor();
if (ggml_n_dims(x) == 4) { if (ggml_n_dims(x) == 4) {
// assuming video mode (if batch processing gets implemented this will break) // assuming video mode (if batch processing gets implemented this will break)
int T = x->ne[2]; int64_t T = x->ne[2];
if (sd_version_is_wan(version)) { if (sd_version_is_wan(version)) {
T = ((T - 1) * 4) + 1; T = ((T - 1) * 4) + 1;
} }
@ -2077,7 +2077,7 @@ public:
img_cond_data = (float*)out_img_cond->data; img_cond_data = (float*)out_img_cond->data;
} }
int step_count = sigmas.size(); int step_count = static_cast<int>(sigmas.size());
bool is_skiplayer_step = has_skiplayer && step > (int)(guidance.slg.layer_start * step_count) && step < (int)(guidance.slg.layer_end * step_count); bool is_skiplayer_step = has_skiplayer && step > (int)(guidance.slg.layer_start * step_count) && step < (int)(guidance.slg.layer_end * step_count);
float* skip_layer_data = has_skiplayer ? (float*)out_skip->data : nullptr; float* skip_layer_data = has_skiplayer ? (float*)out_skip->data : nullptr;
if (is_skiplayer_step) { if (is_skiplayer_step) {
@ -2449,11 +2449,11 @@ public:
int& tile_size_y, int& tile_size_y,
float& tile_overlap, float& tile_overlap,
const sd_tiling_params_t& params, const sd_tiling_params_t& params,
int latent_x, int64_t latent_x,
int latent_y, int64_t latent_y,
float encoding_factor = 1.0f) { float encoding_factor = 1.0f) {
tile_overlap = std::max(std::min(params.target_overlap, 0.5f), 0.0f); tile_overlap = std::max(std::min(params.target_overlap, 0.5f), 0.0f);
auto get_tile_size = [&](int requested_size, float factor, int latent_size) { auto get_tile_size = [&](int requested_size, float factor, int64_t latent_size) {
const int default_tile_size = 32; const int default_tile_size = 32;
const int min_tile_dimension = 4; const int min_tile_dimension = 4;
int tile_size = default_tile_size; int tile_size = default_tile_size;
@ -2462,12 +2462,12 @@ public:
if (factor > 0.f) { if (factor > 0.f) {
if (factor > 1.0) if (factor > 1.0)
factor = 1 / (factor - factor * tile_overlap + tile_overlap); factor = 1 / (factor - factor * tile_overlap + tile_overlap);
tile_size = std::round(latent_size * factor); tile_size = static_cast<int>(std::round(latent_size * factor));
} else if (requested_size >= min_tile_dimension) { } else if (requested_size >= min_tile_dimension) {
tile_size = requested_size; tile_size = requested_size;
} }
tile_size *= encoding_factor; tile_size = static_cast<int>(tile_size * encoding_factor);
return std::max(std::min(tile_size, latent_size), min_tile_dimension); return std::max(std::min(tile_size, static_cast<int>(latent_size)), min_tile_dimension);
}; };
tile_size_x = get_tile_size(params.tile_size_x, params.rel_size_x, latent_x); tile_size_x = get_tile_size(params.tile_size_x, params.rel_size_x, latent_x);
@ -2478,13 +2478,13 @@ public:
int64_t t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
ggml_tensor* result = nullptr; ggml_tensor* result = nullptr;
const int vae_scale_factor = get_vae_scale_factor(); const int vae_scale_factor = get_vae_scale_factor();
int W = x->ne[0] / vae_scale_factor; int64_t W = x->ne[0] / vae_scale_factor;
int H = x->ne[1] / vae_scale_factor; int64_t H = x->ne[1] / vae_scale_factor;
int C = get_latent_channel(); int64_t C = get_latent_channel();
if (vae_tiling_params.enabled && !encode_video) { if (vae_tiling_params.enabled && !encode_video) {
// TODO wan2.2 vae support? // TODO wan2.2 vae support?
int ne2; int64_t ne2;
int ne3; int64_t ne3;
if (sd_version_is_qwen_image(version)) { if (sd_version_is_qwen_image(version)) {
ne2 = 1; ne2 = 1;
ne3 = C * x->ne[3]; ne3 = C * x->ne[3];
@ -2608,7 +2608,7 @@ public:
int64_t C = 3; int64_t C = 3;
ggml_tensor* result = nullptr; ggml_tensor* result = nullptr;
if (decode_video) { if (decode_video) {
int T = x->ne[2]; int64_t T = x->ne[2];
if (sd_version_is_wan(version)) { if (sd_version_is_wan(version)) {
T = ((T - 1) * 4) + 1; T = ((T - 1) * 4) + 1;
} }
@ -3193,7 +3193,7 @@ sd_image_t* generate_image_internal(sd_ctx_t* sd_ctx,
guidance.img_cfg = guidance.txt_cfg; guidance.img_cfg = guidance.txt_cfg;
} }
int sample_steps = sigmas.size() - 1; int sample_steps = static_cast<int>(sigmas.size() - 1);
int64_t t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
@ -3203,7 +3203,7 @@ sd_image_t* generate_image_internal(sd_ctx_t* sd_ctx,
condition_params.width = width; condition_params.width = width;
condition_params.height = height; condition_params.height = height;
condition_params.ref_images = ref_images; condition_params.ref_images = ref_images;
condition_params.adm_in_channels = sd_ctx->sd->diffusion_model->get_adm_in_channels(); condition_params.adm_in_channels = static_cast<int>(sd_ctx->sd->diffusion_model->get_adm_in_channels());
// Photo Maker // Photo Maker
SDCondition id_cond = sd_ctx->sd->get_pmid_conditon(work_ctx, pm_params, condition_params); SDCondition id_cond = sd_ctx->sd->get_pmid_conditon(work_ctx, pm_params, condition_params);
@ -3799,7 +3799,7 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
// timesteps ∝ sigmas for Flow models (like wan2.2 a14b) // timesteps ∝ sigmas for Flow models (like wan2.2 a14b)
for (size_t i = 0; i < sigmas.size(); ++i) { for (size_t i = 0; i < sigmas.size(); ++i) {
if (sigmas[i] < sd_vid_gen_params->moe_boundary) { if (sigmas[i] < sd_vid_gen_params->moe_boundary) {
high_noise_sample_steps = i; high_noise_sample_steps = static_cast<int>(i);
break; break;
} }
} }
@ -3977,7 +3977,7 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
int64_t length = inactive->ne[2]; int64_t length = inactive->ne[2];
if (ref_image_latent) { if (ref_image_latent) {
length += 1; length += 1;
frames = (length - 1) * 4 + 1; frames = static_cast<int>((length - 1) * 4 + 1);
ref_image_num = 1; ref_image_num = 1;
} }
vace_context = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, inactive->ne[0], inactive->ne[1], length, 96); // [b*96, t, h/vae_scale_factor, w/vae_scale_factor] vace_context = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, inactive->ne[0], inactive->ne[1], length, 96); // [b*96, t, h/vae_scale_factor, w/vae_scale_factor]
@ -4043,7 +4043,7 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
int W = width / vae_scale_factor; int W = width / vae_scale_factor;
int H = height / vae_scale_factor; int H = height / vae_scale_factor;
int T = init_latent->ne[2]; int T = static_cast<int>(init_latent->ne[2]);
int C = sd_ctx->sd->get_latent_channel(); int C = sd_ctx->sd->get_latent_channel();
struct ggml_tensor* final_latent; struct ggml_tensor* final_latent;
@ -4162,13 +4162,13 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
ggml_free(work_ctx); ggml_free(work_ctx);
return nullptr; return nullptr;
} }
*num_frames_out = vid->ne[2]; *num_frames_out = static_cast<int>(vid->ne[2]);
for (size_t i = 0; i < vid->ne[2]; i++) { for (int64_t i = 0; i < vid->ne[2]; i++) {
result_images[i].width = vid->ne[0]; result_images[i].width = static_cast<uint32_t>(vid->ne[0]);
result_images[i].height = vid->ne[1]; result_images[i].height = static_cast<uint32_t>(vid->ne[1]);
result_images[i].channel = 3; result_images[i].channel = 3;
result_images[i].data = ggml_tensor_to_sd_image(vid, i, true); result_images[i].data = ggml_tensor_to_sd_image(vid, static_cast<int>(i), true);
} }
ggml_free(work_ctx); ggml_free(work_ctx);

24
t5.hpp
View File

@ -96,7 +96,7 @@ protected:
try { try {
data = nlohmann::json::parse(json_str); data = nlohmann::json::parse(json_str);
} catch (const nlohmann::json::parse_error& e) { } catch (const nlohmann::json::parse_error&) {
status_ = INVLIAD_JSON; status_ = INVLIAD_JSON;
return; return;
} }
@ -168,9 +168,9 @@ protected:
kMaxTrieResultsSize); kMaxTrieResultsSize);
trie_results_size_ = 0; trie_results_size_ = 0;
for (const auto& p : *pieces) { for (const auto& p : *pieces) {
const int num_nodes = trie_->commonPrefixSearch( const size_t num_nodes = trie_->commonPrefixSearch(
p.first.data(), results.data(), results.size(), p.first.size()); p.first.data(), results.data(), results.size(), p.first.size());
trie_results_size_ = std::max(trie_results_size_, num_nodes); trie_results_size_ = std::max(trie_results_size_, static_cast<int>(num_nodes));
} }
if (trie_results_size_ == 0) if (trie_results_size_ == 0)
@ -268,7 +268,7 @@ protected:
-1; // The starting position (in utf-8) of this node. The entire best -1; // The starting position (in utf-8) of this node. The entire best
// path can be constructed by backtracking along this link. // path can be constructed by backtracking along this link.
}; };
const int size = normalized.size(); const int size = static_cast<int>(normalized.size());
const float unk_score = min_score() - kUnkPenalty; const float unk_score = min_score() - kUnkPenalty;
// The ends are exclusive. // The ends are exclusive.
std::vector<BestPathNode> best_path_ends_at(size + 1); std::vector<BestPathNode> best_path_ends_at(size + 1);
@ -281,7 +281,7 @@ protected:
best_path_ends_at[starts_at].best_path_score; best_path_ends_at[starts_at].best_path_score;
bool has_single_node = false; bool has_single_node = false;
const int mblen = const int mblen =
std::min<int>(OneCharLen(normalized.data() + starts_at), std::min<int>(static_cast<int>(OneCharLen(normalized.data() + starts_at)),
size - starts_at); size - starts_at);
while (key_pos < size) { while (key_pos < size) {
const int ret = const int ret =
@ -302,7 +302,7 @@ protected:
score + best_path_score_till_here; score + best_path_score_till_here;
if (target_node.starts_at == -1 || if (target_node.starts_at == -1 ||
candidate_best_path_score > target_node.best_path_score) { candidate_best_path_score > target_node.best_path_score) {
target_node.best_path_score = candidate_best_path_score; target_node.best_path_score = static_cast<float>(candidate_best_path_score);
target_node.starts_at = starts_at; target_node.starts_at = starts_at;
target_node.id = ret; target_node.id = ret;
} }
@ -394,7 +394,7 @@ public:
bool padding = false) { bool padding = false) {
if (max_length > 0 && padding) { if (max_length > 0 && padding) {
size_t orig_token_num = tokens.size() - 1; size_t orig_token_num = tokens.size() - 1;
size_t n = std::ceil(orig_token_num * 1.0 / (max_length - 1)); size_t n = static_cast<size_t>(std::ceil(orig_token_num * 1.0 / (max_length - 1)));
if (n == 0) { if (n == 0) {
n = 1; n = 1;
} }
@ -608,7 +608,7 @@ public:
} }
} }
k = ggml_scale_inplace(ctx->ggml_ctx, k, sqrt(d_head)); k = ggml_scale_inplace(ctx->ggml_ctx, k, ::sqrtf(static_cast<float>(d_head)));
x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, num_heads, mask); // [N, n_token, d_head * n_head] x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, num_heads, mask); // [N, n_token, d_head * n_head]
@ -797,7 +797,7 @@ struct T5Runner : public GGMLRunner {
input_ids = to_backend(input_ids); input_ids = to_backend(input_ids);
attention_mask = to_backend(attention_mask); attention_mask = to_backend(attention_mask);
relative_position_bucket_vec = compute_relative_position_bucket(input_ids->ne[0], input_ids->ne[0]); relative_position_bucket_vec = compute_relative_position_bucket(static_cast<int>(input_ids->ne[0]), static_cast<int>(input_ids->ne[0]));
// for (int i = 0; i < relative_position_bucket_vec.size(); i++) { // for (int i = 0; i < relative_position_bucket_vec.size(); i++) {
// if (i % 77 == 0) { // if (i % 77 == 0) {
@ -984,12 +984,12 @@ struct T5Embedder {
auto attention_mask = vector_to_ggml_tensor(work_ctx, masks); auto attention_mask = vector_to_ggml_tensor(work_ctx, masks);
struct ggml_tensor* out = nullptr; struct ggml_tensor* out = nullptr;
int t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
model.compute(8, input_ids, attention_mask, &out, work_ctx); model.compute(8, input_ids, attention_mask, &out, work_ctx);
int t1 = ggml_time_ms(); int64_t t1 = ggml_time_ms();
print_ggml_tensor(out); print_ggml_tensor(out);
LOG_DEBUG("t5 test done in %dms", t1 - t0); LOG_DEBUG("t5 test done in %lldms", t1 - t0);
} }
} }

18
thirdparty/darts.h vendored
View File

@ -845,7 +845,7 @@ inline void BitVector::build() {
num_ones_ = 0; num_ones_ = 0;
for (std::size_t i = 0; i < units_.size(); ++i) { for (std::size_t i = 0; i < units_.size(); ++i) {
ranks_[i] = num_ones_; ranks_[i] = static_cast<id_type>(num_ones_);
num_ones_ += pop_count(units_[i]); num_ones_ += pop_count(units_[i]);
} }
} }
@ -1769,7 +1769,7 @@ id_type DoubleArrayBuilder::arrange_from_keyset(const Keyset<T> &keyset,
inline id_type DoubleArrayBuilder::find_valid_offset(id_type id) const { inline id_type DoubleArrayBuilder::find_valid_offset(id_type id) const {
if (extras_head_ >= units_.size()) { if (extras_head_ >= units_.size()) {
return units_.size() | (id & LOWER_MASK); return static_cast<id_type>(units_.size()) | (id & LOWER_MASK);
} }
id_type unfixed_id = extras_head_; id_type unfixed_id = extras_head_;
@ -1781,7 +1781,7 @@ inline id_type DoubleArrayBuilder::find_valid_offset(id_type id) const {
unfixed_id = extras(unfixed_id).next(); unfixed_id = extras(unfixed_id).next();
} while (unfixed_id != extras_head_); } while (unfixed_id != extras_head_);
return units_.size() | (id & LOWER_MASK); return static_cast<id_type>(units_.size()) | (id & LOWER_MASK);
} }
inline bool DoubleArrayBuilder::is_valid_offset(id_type id, inline bool DoubleArrayBuilder::is_valid_offset(id_type id,
@ -1812,7 +1812,7 @@ inline void DoubleArrayBuilder::reserve_id(id_type id) {
if (id == extras_head_) { if (id == extras_head_) {
extras_head_ = extras(id).next(); extras_head_ = extras(id).next();
if (extras_head_ == id) { if (extras_head_ == id) {
extras_head_ = units_.size(); extras_head_ = static_cast<id_type>(units_.size());
} }
} }
extras(extras(id).prev()).set_next(extras(id).next()); extras(extras(id).prev()).set_next(extras(id).next());
@ -1821,8 +1821,8 @@ inline void DoubleArrayBuilder::reserve_id(id_type id) {
} }
inline void DoubleArrayBuilder::expand_units() { inline void DoubleArrayBuilder::expand_units() {
id_type src_num_units = units_.size(); id_type src_num_units = static_cast<id_type>(units_.size());
id_type src_num_blocks = num_blocks(); id_type src_num_blocks = static_cast<id_type>(num_blocks());
id_type dest_num_units = src_num_units + BLOCK_SIZE; id_type dest_num_units = src_num_units + BLOCK_SIZE;
id_type dest_num_blocks = src_num_blocks + 1; id_type dest_num_blocks = src_num_blocks + 1;
@ -1834,7 +1834,7 @@ inline void DoubleArrayBuilder::expand_units() {
units_.resize(dest_num_units); units_.resize(dest_num_units);
if (dest_num_blocks > NUM_EXTRA_BLOCKS) { if (dest_num_blocks > NUM_EXTRA_BLOCKS) {
for (std::size_t id = src_num_units; id < dest_num_units; ++id) { for (id_type id = src_num_units; id < dest_num_units; ++id) {
extras(id).set_is_used(false); extras(id).set_is_used(false);
extras(id).set_is_fixed(false); extras(id).set_is_fixed(false);
} }
@ -1858,9 +1858,9 @@ inline void DoubleArrayBuilder::expand_units() {
inline void DoubleArrayBuilder::fix_all_blocks() { inline void DoubleArrayBuilder::fix_all_blocks() {
id_type begin = 0; id_type begin = 0;
if (num_blocks() > NUM_EXTRA_BLOCKS) { if (num_blocks() > NUM_EXTRA_BLOCKS) {
begin = num_blocks() - NUM_EXTRA_BLOCKS; begin = static_cast<id_type>(num_blocks() - NUM_EXTRA_BLOCKS);
} }
id_type end = num_blocks(); id_type end = static_cast<id_type>(num_blocks());
for (id_type block_id = begin; block_id != end; ++block_id) { for (id_type block_id = begin; block_id != end; ++block_id) {
fix_block(block_id); fix_block(block_id);

14
thirdparty/stb_image_write.h vendored
View File

@ -257,6 +257,10 @@ int stbi_write_tga_with_rle = 1;
int stbi_write_force_png_filter = -1; int stbi_write_force_png_filter = -1;
#endif #endif
#ifndef STBMIN
#define STBMIN(a, b) ((a) < (b) ? (a) : (b))
#endif // STBMIN
static int stbi__flip_vertically_on_write = 0; static int stbi__flip_vertically_on_write = 0;
STBIWDEF void stbi_flip_vertically_on_write(int flag) STBIWDEF void stbi_flip_vertically_on_write(int flag)
@ -1179,8 +1183,8 @@ STBIWDEF unsigned char *stbi_write_png_to_mem(const unsigned char *pixels, int s
if (!zlib) return 0; if (!zlib) return 0;
if(parameters != NULL) { if(parameters != NULL) {
param_length = strlen(parameters); param_length = (int)strlen(parameters);
param_length += strlen("parameters") + 1; // For the name and the null-byte param_length += (int)strlen("parameters") + 1; // For the name and the null-byte
} }
// each tag requires 12 bytes of overhead // each tag requires 12 bytes of overhead
@ -1526,11 +1530,11 @@ static int stbi_write_jpg_core(stbi__write_context *s, int width, int height, in
if(parameters != NULL) { if(parameters != NULL) {
stbiw__putc(s, 0xFF /* comnent */ ); stbiw__putc(s, 0xFF /* comnent */ );
stbiw__putc(s, 0xFE /* marker */ ); stbiw__putc(s, 0xFE /* marker */ );
size_t param_length = std::min(2 + strlen("parameters") + 1 + strlen(parameters) + 1, (size_t) 0xFFFF); int param_length = STBMIN(2 + (int)strlen("parameters") + 1 + (int)strlen(parameters) + 1, 0xFFFF);
stbiw__putc(s, param_length >> 8); // no need to mask, length < 65536 stbiw__putc(s, param_length >> 8); // no need to mask, length < 65536
stbiw__putc(s, param_length & 0xFF); stbiw__putc(s, param_length & 0xFF);
s->func(s->context, (void*)"parameters", strlen("parameters") + 1); // std::string is zero-terminated s->func(s->context, (void*)"parameters", (int)strlen("parameters") + 1); // std::string is zero-terminated
s->func(s->context, (void*)parameters, std::min(param_length, (size_t) 65534) - 2 - strlen("parameters") - 1); s->func(s->context, (void*)parameters, STBMIN(param_length, 65534) - 2 - (int)strlen("parameters") - 1);
if(param_length > 65534) stbiw__putc(s, 0); // always zero-terminate for safety if(param_length > 65534) stbiw__putc(s, 0); // always zero-terminate for safety
if(param_length & 1) stbiw__putc(s, 0xFF); // pad to even length if(param_length & 1) stbiw__putc(s, 0xFF); // pad to even length
} }

18
unet.hpp
View File

@ -12,7 +12,7 @@
class SpatialVideoTransformer : public SpatialTransformer { class SpatialVideoTransformer : public SpatialTransformer {
protected: protected:
int64_t time_depth; int64_t time_depth;
int64_t max_time_embed_period; int max_time_embed_period;
public: public:
SpatialVideoTransformer(int64_t in_channels, SpatialVideoTransformer(int64_t in_channels,
@ -22,7 +22,7 @@ public:
int64_t context_dim, int64_t context_dim,
bool use_linear, bool use_linear,
int64_t time_depth = 1, int64_t time_depth = 1,
int64_t max_time_embed_period = 10000) int max_time_embed_period = 10000)
: SpatialTransformer(in_channels, n_head, d_head, depth, context_dim, use_linear), : SpatialTransformer(in_channels, n_head, d_head, depth, context_dim, use_linear),
max_time_embed_period(max_time_embed_period) { max_time_embed_period(max_time_embed_period) {
// We will convert unet transformer linear to conv2d 1x1 when loading the weights, so use_linear is always False // We will convert unet transformer linear to conv2d 1x1 when loading the weights, so use_linear is always False
@ -112,9 +112,9 @@ public:
x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 2, 0, 3)); // [N, h, w, inner_dim] x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 2, 0, 3)); // [N, h, w, inner_dim]
x = ggml_reshape_3d(ctx->ggml_ctx, x, inner_dim, w * h, n); // [N, h * w, inner_dim] x = ggml_reshape_3d(ctx->ggml_ctx, x, inner_dim, w * h, n); // [N, h * w, inner_dim]
auto num_frames = ggml_arange(ctx->ggml_ctx, 0, timesteps, 1); auto num_frames = ggml_arange(ctx->ggml_ctx, 0.f, static_cast<float>(timesteps), 1.f);
// since b is 1, no need to do repeat // since b is 1, no need to do repeat
auto t_emb = ggml_ext_timestep_embedding(ctx->ggml_ctx, num_frames, in_channels, max_time_embed_period); // [N, in_channels] auto t_emb = ggml_ext_timestep_embedding(ctx->ggml_ctx, num_frames, static_cast<int>(in_channels), max_time_embed_period); // [N, in_channels]
auto emb = time_pos_embed_0->forward(ctx, t_emb); auto emb = time_pos_embed_0->forward(ctx, t_emb);
emb = ggml_silu_inplace(ctx->ggml_ctx, emb); emb = ggml_silu_inplace(ctx->ggml_ctx, emb);
@ -526,7 +526,7 @@ public:
auto cs = ggml_scale_inplace(ctx->ggml_ctx, controls[controls.size() - 1], control_strength); auto cs = ggml_scale_inplace(ctx->ggml_ctx, controls[controls.size() - 1], control_strength);
h = ggml_add(ctx->ggml_ctx, h, cs); // middle control h = ggml_add(ctx->ggml_ctx, h, cs); // middle control
} }
int control_offset = controls.size() - 2; int control_offset = static_cast<int>(controls.size() - 2);
// output_blocks // output_blocks
int output_block_idx = 0; int output_block_idx = 0;
@ -615,7 +615,7 @@ struct UNetModelRunner : public GGMLRunner {
struct ggml_cgraph* gf = new_graph_custom(UNET_GRAPH_SIZE); struct ggml_cgraph* gf = new_graph_custom(UNET_GRAPH_SIZE);
if (num_video_frames == -1) { if (num_video_frames == -1) {
num_video_frames = x->ne[3]; num_video_frames = static_cast<int>(x->ne[3]);
} }
x = to_backend(x); x = to_backend(x);
@ -700,12 +700,12 @@ struct UNetModelRunner : public GGMLRunner {
struct ggml_tensor* out = nullptr; struct ggml_tensor* out = nullptr;
int t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
compute(8, x, timesteps, context, nullptr, y, num_video_frames, {}, 0.f, &out, work_ctx); compute(8, x, timesteps, context, nullptr, y, num_video_frames, {}, 0.f, &out, work_ctx);
int t1 = ggml_time_ms(); int64_t t1 = ggml_time_ms();
print_ggml_tensor(out); print_ggml_tensor(out);
LOG_DEBUG("unet test done in %dms", t1 - t0); LOG_DEBUG("unet test done in %lldms", t1 - t0);
} }
} }
}; };

32
util.cpp
View File

@ -488,7 +488,7 @@ sd_image_f32_t sd_image_t_to_sd_image_f32_t(sd_image_t image) {
// Allocate memory for float data // Allocate memory for float data
converted_image.data = (float*)malloc(image.width * image.height * image.channel * sizeof(float)); converted_image.data = (float*)malloc(image.width * image.height * image.channel * sizeof(float));
for (int i = 0; i < image.width * image.height * image.channel; i++) { for (uint32_t i = 0; i < image.width * image.height * image.channel; i++) {
// Convert uint8_t to float // Convert uint8_t to float
converted_image.data[i] = (float)image.data[i]; converted_image.data[i] = (float)image.data[i];
} }
@ -520,7 +520,7 @@ sd_image_f32_t resize_sd_image_f32_t(sd_image_f32_t image, int target_width, int
uint32_t x2 = std::min(x1 + 1, image.width - 1); uint32_t x2 = std::min(x1 + 1, image.width - 1);
uint32_t y2 = std::min(y1 + 1, image.height - 1); uint32_t y2 = std::min(y1 + 1, image.height - 1);
for (int k = 0; k < image.channel; k++) { for (uint32_t k = 0; k < image.channel; k++) {
float v1 = *(image.data + y1 * image.width * image.channel + x1 * image.channel + k); float v1 = *(image.data + y1 * image.width * image.channel + x1 * image.channel + k);
float v2 = *(image.data + y1 * image.width * image.channel + x2 * image.channel + k); float v2 = *(image.data + y1 * image.width * image.channel + x2 * image.channel + k);
float v3 = *(image.data + y2 * image.width * image.channel + x1 * image.channel + k); float v3 = *(image.data + y2 * image.width * image.channel + x1 * image.channel + k);
@ -540,9 +540,9 @@ sd_image_f32_t resize_sd_image_f32_t(sd_image_f32_t image, int target_width, int
} }
void normalize_sd_image_f32_t(sd_image_f32_t image, float means[3], float stds[3]) { void normalize_sd_image_f32_t(sd_image_f32_t image, float means[3], float stds[3]) {
for (int y = 0; y < image.height; y++) { for (uint32_t y = 0; y < image.height; y++) {
for (int x = 0; x < image.width; x++) { for (uint32_t x = 0; x < image.width; x++) {
for (int k = 0; k < image.channel; k++) { for (uint32_t k = 0; k < image.channel; k++) {
int index = (y * image.width + x) * image.channel + k; int index = (y * image.width + x) * image.channel + k;
image.data[index] = (image.data[index] - means[k]) / stds[k]; image.data[index] = (image.data[index] - means[k]) / stds[k];
} }
@ -551,8 +551,8 @@ void normalize_sd_image_f32_t(sd_image_f32_t image, float means[3], float stds[3
} }
// Constants for means and std // Constants for means and std
float means[3] = {0.48145466, 0.4578275, 0.40821073}; float means[3] = {0.48145466f, 0.4578275f, 0.40821073f};
float stds[3] = {0.26862954, 0.26130258, 0.27577711}; float stds[3] = {0.26862954f, 0.26130258f, 0.27577711f};
// Function to clip and preprocess sd_image_f32_t // Function to clip and preprocess sd_image_f32_t
sd_image_f32_t clip_preprocess(sd_image_f32_t image, int target_width, int target_height) { sd_image_f32_t clip_preprocess(sd_image_f32_t image, int target_width, int target_height) {
@ -576,7 +576,7 @@ sd_image_f32_t clip_preprocess(sd_image_f32_t image, int target_width, int targe
uint32_t x2 = std::min(x1 + 1, image.width - 1); uint32_t x2 = std::min(x1 + 1, image.width - 1);
uint32_t y2 = std::min(y1 + 1, image.height - 1); uint32_t y2 = std::min(y1 + 1, image.height - 1);
for (int k = 0; k < image.channel; k++) { for (uint32_t k = 0; k < image.channel; k++) {
float v1 = *(image.data + y1 * image.width * image.channel + x1 * image.channel + k); float v1 = *(image.data + y1 * image.width * image.channel + x1 * image.channel + k);
float v2 = *(image.data + y1 * image.width * image.channel + x2 * image.channel + k); float v2 = *(image.data + y1 * image.width * image.channel + x2 * image.channel + k);
float v3 = *(image.data + y2 * image.width * image.channel + x1 * image.channel + k); float v3 = *(image.data + y2 * image.width * image.channel + x1 * image.channel + k);
@ -602,11 +602,11 @@ sd_image_f32_t clip_preprocess(sd_image_f32_t image, int target_width, int targe
result.channel = image.channel; result.channel = image.channel;
result.data = (float*)malloc(target_height * target_width * image.channel * sizeof(float)); result.data = (float*)malloc(target_height * target_width * image.channel * sizeof(float));
for (int k = 0; k < image.channel; k++) { for (uint32_t k = 0; k < image.channel; k++) {
for (int i = 0; i < result.height; i++) { for (uint32_t i = 0; i < result.height; i++) {
for (int j = 0; j < result.width; j++) { for (uint32_t j = 0; j < result.width; j++) {
int src_y = std::min(i + h_offset, resized_height - 1); int src_y = std::min(static_cast<int>(i + h_offset), resized_height - 1);
int src_x = std::min(j + w_offset, resized_width - 1); int src_x = std::min(static_cast<int>(j + w_offset), resized_width - 1);
*(result.data + i * result.width * image.channel + j * image.channel + k) = *(result.data + i * result.width * image.channel + j * image.channel + k) =
fmin(fmax(*(resized_data + src_y * resized_width * image.channel + src_x * image.channel + k), 0.0f), 255.0f) / 255.0f; fmin(fmax(*(resized_data + src_y * resized_width * image.channel + src_x * image.channel + k), 0.0f), 255.0f) / 255.0f;
} }
@ -617,9 +617,9 @@ sd_image_f32_t clip_preprocess(sd_image_f32_t image, int target_width, int targe
free(resized_data); free(resized_data);
// Normalize // Normalize
for (int k = 0; k < image.channel; k++) { for (uint32_t k = 0; k < image.channel; k++) {
for (int i = 0; i < result.height; i++) { for (uint32_t i = 0; i < result.height; i++) {
for (int j = 0; j < result.width; j++) { for (uint32_t j = 0; j < result.width; j++) {
// *(result.data + i * size * image.channel + j * image.channel + k) = 0.5f; // *(result.data + i * size * image.channel + j * image.channel + k) = 0.5f;
int offset = i * result.width * image.channel + j * image.channel + k; int offset = i * result.width * image.channel + j * image.channel + k;
float value = *(result.data + offset); float value = *(result.data + offset);

30
vae.hpp
View File

@ -166,18 +166,18 @@ public:
AE3DConv(int64_t in_channels, AE3DConv(int64_t in_channels,
int64_t out_channels, int64_t out_channels,
std::pair<int, int> kernel_size, std::pair<int, int> kernel_size,
int64_t video_kernel_size = 3, int video_kernel_size = 3,
std::pair<int, int> stride = {1, 1}, std::pair<int, int> stride = {1, 1},
std::pair<int, int> padding = {0, 0}, std::pair<int, int> padding = {0, 0},
std::pair<int, int> dilation = {1, 1}, std::pair<int, int> dilation = {1, 1},
bool bias = true) bool bias = true)
: Conv2d(in_channels, out_channels, kernel_size, stride, padding, dilation, bias) { : Conv2d(in_channels, out_channels, kernel_size, stride, padding, dilation, bias) {
int64_t kernel_padding = video_kernel_size / 2; int kernel_padding = video_kernel_size / 2;
blocks["time_mix_conv"] = std::shared_ptr<GGMLBlock>(new Conv3dnx1x1(out_channels, blocks["time_mix_conv"] = std::shared_ptr<GGMLBlock>(new Conv3d(out_channels,
out_channels, out_channels,
video_kernel_size, {video_kernel_size, 1, 1},
1, {1, 1, 1},
kernel_padding)); {kernel_padding, 0, 0}));
} }
struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* forward(GGMLRunnerContext* ctx,
@ -186,7 +186,7 @@ public:
// skip_video always False // skip_video always False
// x: [N, IC, IH, IW] // x: [N, IC, IH, IW]
// result: [N, OC, OH, OW] // result: [N, OC, OH, OW]
auto time_mix_conv = std::dynamic_pointer_cast<Conv3dnx1x1>(blocks["time_mix_conv"]); auto time_mix_conv = std::dynamic_pointer_cast<Conv3d>(blocks["time_mix_conv"]);
x = Conv2d::forward(ctx, x); x = Conv2d::forward(ctx, x);
// timesteps = x.shape[0] // timesteps = x.shape[0]
@ -409,7 +409,7 @@ public:
z_channels(z_channels), z_channels(z_channels),
video_decoder(video_decoder), video_decoder(video_decoder),
video_kernel_size(video_kernel_size) { video_kernel_size(video_kernel_size) {
size_t num_resolutions = ch_mult.size(); int num_resolutions = static_cast<int>(ch_mult.size());
int block_in = ch * ch_mult[num_resolutions - 1]; int block_in = ch * ch_mult[num_resolutions - 1];
blocks["conv_in"] = std::shared_ptr<GGMLBlock>(new Conv2d(z_channels, block_in, {3, 3}, {1, 1}, {1, 1})); blocks["conv_in"] = std::shared_ptr<GGMLBlock>(new Conv2d(z_channels, block_in, {3, 3}, {1, 1}, {1, 1}));
@ -461,7 +461,7 @@ public:
h = mid_block_2->forward(ctx, h); // [N, block_in, h, w] h = mid_block_2->forward(ctx, h); // [N, block_in, h, w]
// upsampling // upsampling
size_t num_resolutions = ch_mult.size(); int num_resolutions = static_cast<int>(ch_mult.size());
for (int i = num_resolutions - 1; i >= 0; i--) { for (int i = num_resolutions - 1; i >= 0; i--) {
for (int j = 0; j < num_res_blocks + 1; j++) { for (int j = 0; j < num_res_blocks + 1; j++) {
std::string name = "up." + std::to_string(i) + ".block." + std::to_string(j); std::string name = "up." + std::to_string(i) + ".block." + std::to_string(j);
@ -745,12 +745,12 @@ struct AutoEncoderKL : public VAE {
print_ggml_tensor(x); print_ggml_tensor(x);
struct ggml_tensor* out = nullptr; struct ggml_tensor* out = nullptr;
int t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
compute(8, x, false, &out, work_ctx); compute(8, x, false, &out, work_ctx);
int t1 = ggml_time_ms(); int64_t t1 = ggml_time_ms();
print_ggml_tensor(out); print_ggml_tensor(out);
LOG_DEBUG("encode test done in %dms", t1 - t0); LOG_DEBUG("encode test done in %lldms", t1 - t0);
} }
if (false) { if (false) {
@ -763,12 +763,12 @@ struct AutoEncoderKL : public VAE {
print_ggml_tensor(z); print_ggml_tensor(z);
struct ggml_tensor* out = nullptr; struct ggml_tensor* out = nullptr;
int t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
compute(8, z, true, &out, work_ctx); compute(8, z, true, &out, work_ctx);
int t1 = ggml_time_ms(); int64_t t1 = ggml_time_ms();
print_ggml_tensor(out); print_ggml_tensor(out);
LOG_DEBUG("decode test done in %dms", t1 - t0); LOG_DEBUG("decode test done in %lldms", t1 - t0);
} }
}; };
}; };

62
wan.hpp
View File

@ -108,7 +108,7 @@ namespace WAN {
struct ggml_tensor* w = params["gamma"]; struct ggml_tensor* w = params["gamma"];
w = ggml_reshape_1d(ctx->ggml_ctx, w, ggml_nelements(w)); w = ggml_reshape_1d(ctx->ggml_ctx, w, ggml_nelements(w));
auto h = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 3, 0, 1, 2)); // [ID, IH, IW, N*IC] auto h = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 3, 0, 1, 2)); // [ID, IH, IW, N*IC]
h = ggml_rms_norm(ctx->ggml_ctx, h, 1e-12); h = ggml_rms_norm(ctx->ggml_ctx, h, 1e-12f);
h = ggml_mul(ctx->ggml_ctx, h, w); h = ggml_mul(ctx->ggml_ctx, h, w);
h = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, h, 1, 2, 3, 0)); h = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, h, 1, 2, 3, 0));
@ -243,13 +243,13 @@ namespace WAN {
protected: protected:
int64_t in_channels; int64_t in_channels;
int64_t out_channels; int64_t out_channels;
int64_t factor_t; int factor_t;
int64_t factor_s; int factor_s;
int64_t factor; int factor;
int64_t group_size; int64_t group_size;
public: public:
AvgDown3D(int64_t in_channels, int64_t out_channels, int64_t factor_t, int64_t factor_s = 1) AvgDown3D(int64_t in_channels, int64_t out_channels, int factor_t, int factor_s = 1)
: in_channels(in_channels), out_channels(out_channels), factor_t(factor_t), factor_s(factor_s) { : in_channels(in_channels), out_channels(out_channels), factor_t(factor_t), factor_s(factor_s) {
factor = factor_t * factor_s * factor_s; factor = factor_t * factor_s * factor_s;
GGML_ASSERT(in_channels * factor % out_channels == 0); GGML_ASSERT(in_channels * factor % out_channels == 0);
@ -266,7 +266,7 @@ namespace WAN {
int64_t H = x->ne[1]; int64_t H = x->ne[1];
int64_t W = x->ne[0]; int64_t W = x->ne[0];
int64_t pad_t = (factor_t - T % factor_t) % factor_t; int pad_t = (factor_t - T % factor_t) % factor_t;
x = ggml_pad_ext(ctx->ggml_ctx, x, 0, 0, 0, 0, pad_t, 0, 0, 0); x = ggml_pad_ext(ctx->ggml_ctx, x, 0, 0, 0, 0, pad_t, 0, 0, 0);
T = x->ne[2]; T = x->ne[2];
@ -1071,7 +1071,7 @@ namespace WAN {
int64_t iter_ = z->ne[2]; int64_t iter_ = z->ne[2];
auto x = conv2->forward(ctx, z); auto x = conv2->forward(ctx, z);
struct ggml_tensor* out; struct ggml_tensor* out;
for (int64_t i = 0; i < iter_; i++) { for (int i = 0; i < iter_; i++) {
_conv_idx = 0; _conv_idx = 0;
if (i == 0) { if (i == 0) {
auto in = ggml_ext_slice(ctx->ggml_ctx, x, 2, i, i + 1); // [b*c, 1, h, w] auto in = ggml_ext_slice(ctx->ggml_ctx, x, 2, i, i + 1); // [b*c, 1, h, w]
@ -1091,7 +1091,7 @@ namespace WAN {
struct ggml_tensor* decode_partial(GGMLRunnerContext* ctx, struct ggml_tensor* decode_partial(GGMLRunnerContext* ctx,
struct ggml_tensor* z, struct ggml_tensor* z,
int64_t i, int i,
int64_t b = 1) { int64_t b = 1) {
// z: [b*c, t, h, w] // z: [b*c, t, h, w]
GGML_ASSERT(b == 1); GGML_ASSERT(b == 1);
@ -1146,12 +1146,12 @@ namespace WAN {
return gf; return gf;
} }
struct ggml_cgraph* build_graph_partial(struct ggml_tensor* z, bool decode_graph, int64_t i) { struct ggml_cgraph* build_graph_partial(struct ggml_tensor* z, bool decode_graph, int i) {
struct ggml_cgraph* gf = new_graph_custom(20480); struct ggml_cgraph* gf = new_graph_custom(20480);
ae.clear_cache(); ae.clear_cache();
for (int64_t feat_idx = 0; feat_idx < ae._feat_map.size(); feat_idx++) { for (size_t feat_idx = 0; feat_idx < ae._feat_map.size(); feat_idx++) {
auto feat_cache = get_cache_tensor_by_name("feat_idx:" + std::to_string(feat_idx)); auto feat_cache = get_cache_tensor_by_name("feat_idx:" + std::to_string(feat_idx));
ae._feat_map[feat_idx] = feat_cache; ae._feat_map[feat_idx] = feat_cache;
} }
@ -1162,7 +1162,7 @@ namespace WAN {
struct ggml_tensor* out = decode_graph ? ae.decode_partial(&runner_ctx, z, i) : ae.encode(&runner_ctx, z); struct ggml_tensor* out = decode_graph ? ae.decode_partial(&runner_ctx, z, i) : ae.encode(&runner_ctx, z);
for (int64_t feat_idx = 0; feat_idx < ae._feat_map.size(); feat_idx++) { for (size_t feat_idx = 0; feat_idx < ae._feat_map.size(); feat_idx++) {
ggml_tensor* feat_cache = ae._feat_map[feat_idx]; ggml_tensor* feat_cache = ae._feat_map[feat_idx];
if (feat_cache != nullptr) { if (feat_cache != nullptr) {
cache("feat_idx:" + std::to_string(feat_idx), feat_cache); cache("feat_idx:" + std::to_string(feat_idx), feat_cache);
@ -1188,7 +1188,7 @@ namespace WAN {
} else { // chunk 1 result is weird } else { // chunk 1 result is weird
ae.clear_cache(); ae.clear_cache();
int64_t t = z->ne[2]; int64_t t = z->ne[2];
int64_t i = 0; int i = 0;
auto get_graph = [&]() -> struct ggml_cgraph* { auto get_graph = [&]() -> struct ggml_cgraph* {
return build_graph_partial(z, decode_graph, i); return build_graph_partial(z, decode_graph, i);
}; };
@ -1499,7 +1499,7 @@ namespace WAN {
class WanAttentionBlock : public GGMLBlock { class WanAttentionBlock : public GGMLBlock {
protected: protected:
int dim; int64_t dim;
void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override { void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
enum ggml_type wtype = get_type(prefix + "weight", tensor_storage_map, GGML_TYPE_F32); enum ggml_type wtype = get_type(prefix + "weight", tensor_storage_map, GGML_TYPE_F32);
@ -1639,7 +1639,7 @@ namespace WAN {
class Head : public GGMLBlock { class Head : public GGMLBlock {
protected: protected:
int dim; int64_t dim;
void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override { void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
enum ggml_type wtype = get_type(prefix + "weight", tensor_storage_map, GGML_TYPE_F32); enum ggml_type wtype = get_type(prefix + "weight", tensor_storage_map, GGML_TYPE_F32);
@ -1685,8 +1685,8 @@ namespace WAN {
class MLPProj : public GGMLBlock { class MLPProj : public GGMLBlock {
protected: protected:
int in_dim; int64_t in_dim;
int flf_pos_embed_token_number; int64_t flf_pos_embed_token_number;
void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override { void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
if (flf_pos_embed_token_number > 0) { if (flf_pos_embed_token_number > 0) {
@ -1739,17 +1739,17 @@ namespace WAN {
int64_t in_dim = 16; int64_t in_dim = 16;
int64_t dim = 2048; int64_t dim = 2048;
int64_t ffn_dim = 8192; int64_t ffn_dim = 8192;
int64_t freq_dim = 256; int freq_dim = 256;
int64_t text_dim = 4096; int64_t text_dim = 4096;
int64_t out_dim = 16; int64_t out_dim = 16;
int64_t num_heads = 16; int64_t num_heads = 16;
int64_t num_layers = 32; int num_layers = 32;
int64_t vace_layers = 0; int vace_layers = 0;
int64_t vace_in_dim = 96; int64_t vace_in_dim = 96;
std::map<int, int> vace_layers_mapping = {}; std::map<int, int> vace_layers_mapping = {};
bool qk_norm = true; bool qk_norm = true;
bool cross_attn_norm = true; bool cross_attn_norm = true;
float eps = 1e-6; float eps = 1e-6f;
int64_t flf_pos_embed_token_number = 0; int64_t flf_pos_embed_token_number = 0;
int theta = 10000; int theta = 10000;
// wan2.1 1.3B: 1536/12, wan2.1/2.2 14B: 5120/40, wan2.2 5B: 3074/24 // wan2.1 1.3B: 1536/12, wan2.1/2.2 14B: 5120/40, wan2.2 5B: 3074/24
@ -2066,7 +2066,7 @@ namespace WAN {
if (version == VERSION_WAN2_2_TI2V) { if (version == VERSION_WAN2_2_TI2V) {
desc = "Wan2.2-TI2V-5B"; desc = "Wan2.2-TI2V-5B";
wan_params.dim = 3072; wan_params.dim = 3072;
wan_params.eps = 1e-06; wan_params.eps = 1e-06f;
wan_params.ffn_dim = 14336; wan_params.ffn_dim = 14336;
wan_params.freq_dim = 256; wan_params.freq_dim = 256;
wan_params.in_dim = 48; wan_params.in_dim = 48;
@ -2085,7 +2085,7 @@ namespace WAN {
wan_params.in_dim = 16; wan_params.in_dim = 16;
} }
wan_params.dim = 1536; wan_params.dim = 1536;
wan_params.eps = 1e-06; wan_params.eps = 1e-06f;
wan_params.ffn_dim = 8960; wan_params.ffn_dim = 8960;
wan_params.freq_dim = 256; wan_params.freq_dim = 256;
wan_params.num_heads = 12; wan_params.num_heads = 12;
@ -2114,14 +2114,14 @@ namespace WAN {
} }
} }
wan_params.dim = 5120; wan_params.dim = 5120;
wan_params.eps = 1e-06; wan_params.eps = 1e-06f;
wan_params.ffn_dim = 13824; wan_params.ffn_dim = 13824;
wan_params.freq_dim = 256; wan_params.freq_dim = 256;
wan_params.num_heads = 40; wan_params.num_heads = 40;
wan_params.out_dim = 16; wan_params.out_dim = 16;
wan_params.text_len = 512; wan_params.text_len = 512;
} else { } else {
GGML_ABORT("invalid num_layers(%ld) of wan", wan_params.num_layers); GGML_ABORT("invalid num_layers(%d) of wan", wan_params.num_layers);
} }
LOG_INFO("%s", desc.c_str()); LOG_INFO("%s", desc.c_str());
@ -2156,16 +2156,16 @@ namespace WAN {
time_dim_concat = to_backend(time_dim_concat); time_dim_concat = to_backend(time_dim_concat);
vace_context = to_backend(vace_context); vace_context = to_backend(vace_context);
pe_vec = Rope::gen_wan_pe(x->ne[2], pe_vec = Rope::gen_wan_pe(static_cast<int>(x->ne[2]),
x->ne[1], static_cast<int>(x->ne[1]),
x->ne[0], static_cast<int>(x->ne[0]),
std::get<0>(wan_params.patch_size), std::get<0>(wan_params.patch_size),
std::get<1>(wan_params.patch_size), std::get<1>(wan_params.patch_size),
std::get<2>(wan_params.patch_size), std::get<2>(wan_params.patch_size),
1, 1,
wan_params.theta, wan_params.theta,
wan_params.axes_dim); wan_params.axes_dim);
int pos_len = pe_vec.size() / wan_params.axes_dim_sum / 2; int pos_len = static_cast<int>(pe_vec.size() / wan_params.axes_dim_sum / 2);
// LOG_DEBUG("pos_len %d", pos_len); // LOG_DEBUG("pos_len %d", pos_len);
auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, wan_params.axes_dim_sum / 2, pos_len); auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, wan_params.axes_dim_sum / 2, pos_len);
// pe->data = pe_vec.data(); // pe->data = pe_vec.data();
@ -2243,12 +2243,12 @@ namespace WAN {
struct ggml_tensor* out = nullptr; struct ggml_tensor* out = nullptr;
int t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
compute(8, x, timesteps, context, nullptr, nullptr, nullptr, nullptr, 1.f, &out, work_ctx); compute(8, x, timesteps, context, nullptr, nullptr, nullptr, nullptr, 1.f, &out, work_ctx);
int t1 = ggml_time_ms(); int64_t t1 = ggml_time_ms();
print_ggml_tensor(out); print_ggml_tensor(out);
LOG_DEBUG("wan test done in %dms", t1 - t0); LOG_DEBUG("wan test done in %lldms", t1 - t0);
} }
} }

26
z_image.hpp
View File

@ -239,7 +239,7 @@ namespace ZImage {
}; };
struct ZImageParams { struct ZImageParams {
int64_t patch_size = 2; int patch_size = 2;
int64_t hidden_size = 3840; int64_t hidden_size = 3840;
int64_t in_channels = 16; int64_t in_channels = 16;
int64_t out_channels = 16; int64_t out_channels = 16;
@ -249,11 +249,11 @@ namespace ZImage {
int64_t num_heads = 30; int64_t num_heads = 30;
int64_t num_kv_heads = 30; int64_t num_kv_heads = 30;
int64_t multiple_of = 256; int64_t multiple_of = 256;
float ffn_dim_multiplier = 8.0 / 3.0f; float ffn_dim_multiplier = 8.0f / 3.0f;
float norm_eps = 1e-5f; float norm_eps = 1e-5f;
bool qk_norm = true; bool qk_norm = true;
int64_t cap_feat_dim = 2560; int64_t cap_feat_dim = 2560;
float theta = 256.f; int theta = 256;
std::vector<int> axes_dim = {32, 48, 48}; std::vector<int> axes_dim = {32, 48, 48};
int64_t axes_dim_sum = 128; int64_t axes_dim_sum = 128;
}; };
@ -411,13 +411,13 @@ namespace ZImage {
auto txt = cap_embedder_1->forward(ctx, cap_embedder_0->forward(ctx, context)); // [N, n_txt_token, hidden_size] auto txt = cap_embedder_1->forward(ctx, cap_embedder_0->forward(ctx, context)); // [N, n_txt_token, hidden_size]
auto img = x_embedder->forward(ctx, x); // [N, n_img_token, hidden_size] auto img = x_embedder->forward(ctx, x); // [N, n_img_token, hidden_size]
int64_t n_txt_pad_token = Rope::bound_mod(n_txt_token, SEQ_MULTI_OF); int64_t n_txt_pad_token = Rope::bound_mod(static_cast<int>(n_txt_token), SEQ_MULTI_OF);
if (n_txt_pad_token > 0) { if (n_txt_pad_token > 0) {
auto txt_pad_tokens = ggml_repeat_4d(ctx->ggml_ctx, txt_pad_token, txt_pad_token->ne[0], n_txt_pad_token, N, 1); auto txt_pad_tokens = ggml_repeat_4d(ctx->ggml_ctx, txt_pad_token, txt_pad_token->ne[0], n_txt_pad_token, N, 1);
txt = ggml_concat(ctx->ggml_ctx, txt, txt_pad_tokens, 1); // [N, n_txt_token + n_txt_pad_token, hidden_size] txt = ggml_concat(ctx->ggml_ctx, txt, txt_pad_tokens, 1); // [N, n_txt_token + n_txt_pad_token, hidden_size]
} }
int64_t n_img_pad_token = Rope::bound_mod(n_img_token, SEQ_MULTI_OF); int64_t n_img_pad_token = Rope::bound_mod(static_cast<int>(n_img_token), SEQ_MULTI_OF);
if (n_img_pad_token > 0) { if (n_img_pad_token > 0) {
auto img_pad_tokens = ggml_repeat_4d(ctx->ggml_ctx, img_pad_token, img_pad_token->ne[0], n_img_pad_token, N, 1); auto img_pad_tokens = ggml_repeat_4d(ctx->ggml_ctx, img_pad_token, img_pad_token->ne[0], n_img_pad_token, N, 1);
img = ggml_concat(ctx->ggml_ctx, img, img_pad_tokens, 1); // [N, n_img_token + n_img_pad_token, hidden_size] img = ggml_concat(ctx->ggml_ctx, img, img_pad_tokens, 1); // [N, n_img_token + n_img_pad_token, hidden_size]
@ -543,11 +543,11 @@ namespace ZImage {
ref_latents[i] = to_backend(ref_latents[i]); ref_latents[i] = to_backend(ref_latents[i]);
} }
pe_vec = Rope::gen_z_image_pe(x->ne[1], pe_vec = Rope::gen_z_image_pe(static_cast<int>(x->ne[1]),
x->ne[0], static_cast<int>(x->ne[0]),
z_image_params.patch_size, z_image_params.patch_size,
x->ne[3], static_cast<int>(x->ne[3]),
context->ne[1], static_cast<int>(context->ne[1]),
SEQ_MULTI_OF, SEQ_MULTI_OF,
ref_latents, ref_latents,
increase_ref_index, increase_ref_index,
@ -555,7 +555,7 @@ namespace ZImage {
circular_y_enabled, circular_y_enabled,
circular_x_enabled, circular_x_enabled,
z_image_params.axes_dim); z_image_params.axes_dim);
int pos_len = pe_vec.size() / z_image_params.axes_dim_sum / 2; int pos_len = static_cast<int>(pe_vec.size() / z_image_params.axes_dim_sum / 2);
// LOG_DEBUG("pos_len %d", pos_len); // LOG_DEBUG("pos_len %d", pos_len);
auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, z_image_params.axes_dim_sum / 2, pos_len); auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, z_image_params.axes_dim_sum / 2, pos_len);
// pe->data = pe_vec.data(); // pe->data = pe_vec.data();
@ -619,12 +619,12 @@ namespace ZImage {
struct ggml_tensor* out = nullptr; struct ggml_tensor* out = nullptr;
int t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
compute(8, x, timesteps, context, {}, false, &out, work_ctx); compute(8, x, timesteps, context, {}, false, &out, work_ctx);
int t1 = ggml_time_ms(); int64_t t1 = ggml_time_ms();
print_ggml_tensor(out); print_ggml_tensor(out);
LOG_DEBUG("z_image test done in %dms", t1 - t0); LOG_DEBUG("z_image test done in %lldms", t1 - t0);
} }
} }