alphard/stable-diffusion.cpp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

style: format code

Browse Source
This commit is contained in:
leejet 2024-08-25 00:19:37 +08:00
parent d08d7fa632
commit c837c5d9cc
10 changed files with 937 additions and 947 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
conditioner.hpp
View File

@ -10,8 +10,8 @@ struct SDCondition {
struct ggml_tensor* c_concat = NULL; struct ggml_tensor* c_concat = NULL;
SDCondition() = default; SDCondition() = default;
SDCondition(struct ggml_tensor* c_crossattn, struct ggml_tensor* c_vector, struct ggml_tensor* c_concat) : SDCondition(struct ggml_tensor* c_crossattn, struct ggml_tensor* c_vector, struct ggml_tensor* c_concat)
c_crossattn(c_crossattn), c_vector(c_vector), c_concat(c_concat) {} : c_crossattn(c_crossattn), c_vector(c_vector), c_concat(c_concat) {}
}; };
struct Conditioner { struct Conditioner {
@ -978,7 +978,6 @@ struct SD3CLIPEmbedder : public Conditioner {
} }
}; };
struct FluxCLIPEmbedder : public Conditioner { struct FluxCLIPEmbedder : public Conditioner {
ggml_type wtype; ggml_type wtype;
CLIPTokenizer clip_l_tokenizer; CLIPTokenizer clip_l_tokenizer;
@ -987,8 +986,8 @@ struct FluxCLIPEmbedder : public Conditioner {
std::shared_ptr<T5Runner> t5; std::shared_ptr<T5Runner> t5;
FluxCLIPEmbedder(ggml_backend_t backend, FluxCLIPEmbedder(ggml_backend_t backend,
ggml_type wtype, ggml_type wtype,
int clip_skip = -1) int clip_skip = -1)
: wtype(wtype) { : wtype(wtype) {
if (clip_skip <= 0) { if (clip_skip <= 0) {
clip_skip = 2; clip_skip = 2;
@ -1085,10 +1084,10 @@ struct FluxCLIPEmbedder : public Conditioner {
auto& t5_tokens = token_and_weights[1].first; auto& t5_tokens = token_and_weights[1].first;
auto& t5_weights = token_and_weights[1].second; auto& t5_weights = token_and_weights[1].second;
int64_t t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
struct ggml_tensor* hidden_states = NULL; // [N, n_token, 4096] struct ggml_tensor* hidden_states = NULL; // [N, n_token, 4096]
struct ggml_tensor* chunk_hidden_states = NULL; // [n_token, 4096] struct ggml_tensor* chunk_hidden_states = NULL; // [n_token, 4096]
struct ggml_tensor* pooled = NULL; // [768,] struct ggml_tensor* pooled = NULL; // [768,]
std::vector<float> hidden_states_vec; std::vector<float> hidden_states_vec;
size_t chunk_len = 256; size_t chunk_len = 256;

3
denoiser.hpp
View File

@ -351,7 +351,6 @@ 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 std::exp(mu) / (std::exp(mu) + std::pow((1.0 / t - 1.0), sigma));
} }
@ -369,7 +368,7 @@ struct FluxFlowDenoiser : public Denoiser {
void set_parameters(float shift = 1.15f) { void set_parameters(float shift = 1.15f) {
this->shift = shift; this->shift = shift;
for (int i = 1; i < TIMESTEPS + 1; i++) { for (int i = 1; i < TIMESTEPS + 1; i++) {
sigmas[i - 1] = t_to_sigma(i/TIMESTEPS * TIMESTEPS); sigmas[i - 1] = t_to_sigma(i / TIMESTEPS * TIMESTEPS);
} }
} }

7
diffusion_model.hpp
View File

@ -1,9 +1,9 @@
#ifndef __DIFFUSION_MODEL_H__ #ifndef __DIFFUSION_MODEL_H__
#define __DIFFUSION_MODEL_H__ #define __DIFFUSION_MODEL_H__
#include "flux.hpp"
#include "mmdit.hpp" #include "mmdit.hpp"
#include "unet.hpp" #include "unet.hpp"
#include "flux.hpp"
struct DiffusionModel { struct DiffusionModel {
virtual void compute(int n_threads, virtual void compute(int n_threads,
@ -124,13 +124,12 @@ struct MMDiTModel : public DiffusionModel {
} }
}; };
struct FluxModel : public DiffusionModel { struct FluxModel : public DiffusionModel {
Flux::FluxRunner flux; Flux::FluxRunner flux;
FluxModel(ggml_backend_t backend, FluxModel(ggml_backend_t backend,
ggml_type wtype, ggml_type wtype,
SDVersion version = VERSION_FLUX_DEV) SDVersion version = VERSION_FLUX_DEV)
: flux(backend, wtype, version) { : flux(backend, wtype, version) {
} }

1712
flux.hpp
View File

File diff suppressed because it is too large Load Diff

36
ggml_extend.hpp
View File

@ -541,7 +541,7 @@ __STATIC_INLINE__ void sd_tiling(ggml_tensor* input, ggml_tensor* output, const
__STATIC_INLINE__ struct ggml_tensor* ggml_group_norm_32(struct ggml_context* ctx, __STATIC_INLINE__ struct ggml_tensor* ggml_group_norm_32(struct ggml_context* ctx,
struct ggml_tensor* a) { struct ggml_tensor* a) {
const float eps = 1e-6f; // default eps parameter const float eps = 1e-6f; // default eps parameter
return ggml_group_norm(ctx, a, 32, eps); return ggml_group_norm(ctx, a, 32, eps);
} }
@ -683,27 +683,27 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
int64_t n_head, int64_t n_head,
struct ggml_tensor* mask = NULL, struct ggml_tensor* mask = NULL,
bool diag_mask_inf = false, bool diag_mask_inf = false,
bool skip_reshape = false) { bool skip_reshape = false) {
int64_t L_q; int64_t L_q;
int64_t L_k; int64_t L_k;
int64_t C ; int64_t C;
int64_t N ; int64_t N;
int64_t d_head; int64_t d_head;
if (!skip_reshape) { if (!skip_reshape) {
L_q = q->ne[1]; L_q = q->ne[1];
L_k = k->ne[1]; L_k = k->ne[1];
C = q->ne[0]; C = q->ne[0];
N = q->ne[2]; N = q->ne[2];
d_head = C / n_head; d_head = C / n_head;
q = ggml_reshape_4d(ctx, q, d_head, n_head, L_q, N); // [N, L_q, n_head, d_head] q = ggml_reshape_4d(ctx, q, d_head, n_head, L_q, N); // [N, L_q, n_head, d_head]
q = ggml_cont(ctx, ggml_permute(ctx, q, 0, 2, 1, 3)); // [N, n_head, L_q, d_head] q = ggml_cont(ctx, ggml_permute(ctx, q, 0, 2, 1, 3)); // [N, n_head, L_q, d_head]
q = ggml_reshape_3d(ctx, q, d_head, L_q, n_head * N); // [N * n_head, L_q, d_head] q = ggml_reshape_3d(ctx, q, d_head, L_q, n_head * N); // [N * n_head, L_q, d_head]
k = ggml_reshape_4d(ctx, k, d_head, n_head, L_k, N); // [N, L_k, n_head, d_head] k = ggml_reshape_4d(ctx, k, d_head, n_head, L_k, N); // [N, L_k, n_head, d_head]
k = ggml_cont(ctx, ggml_permute(ctx, k, 0, 2, 1, 3)); // [N, n_head, L_k, d_head] k = ggml_cont(ctx, ggml_permute(ctx, k, 0, 2, 1, 3)); // [N, n_head, L_k, d_head]
k = ggml_reshape_3d(ctx, k, d_head, L_k, n_head * N); // [N * n_head, L_k, d_head] k = ggml_reshape_3d(ctx, k, d_head, L_k, n_head * N); // [N * n_head, L_k, d_head]
v = ggml_reshape_4d(ctx, v, d_head, n_head, L_k, N); // [N, L_k, n_head, d_head] v = ggml_reshape_4d(ctx, v, d_head, n_head, L_k, N); // [N, L_k, n_head, d_head]
} else { } else {
L_q = q->ne[1]; L_q = q->ne[1];
L_k = k->ne[1]; L_k = k->ne[1];
@ -712,10 +712,10 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
C = d_head * n_head; C = d_head * n_head;
} }
float scale = (1.0f / sqrt((float)d_head)); float scale = (1.0f / sqrt((float)d_head));
bool use_flash_attn = false; bool use_flash_attn = false;
ggml_tensor* kqv = NULL; ggml_tensor* kqv = NULL;
if (use_flash_attn) { if (use_flash_attn) {
v = ggml_cont(ctx, ggml_permute(ctx, v, 0, 2, 1, 3)); // [N, n_head, L_k, d_head] v = ggml_cont(ctx, ggml_permute(ctx, v, 0, 2, 1, 3)); // [N, n_head, L_k, d_head]
v = ggml_reshape_3d(ctx, v, d_head, L_k, n_head * N); // [N * n_head, L_k, d_head] v = ggml_reshape_3d(ctx, v, d_head, L_k, n_head * N); // [N * n_head, L_k, d_head]
@ -770,8 +770,8 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_group_norm(struct ggml_context* ct
b = ggml_reshape_4d(ctx, b, 1, 1, b->ne[0], 1); b = ggml_reshape_4d(ctx, b, 1, 1, b->ne[0], 1);
} }
const float eps = 1e-6f; // default eps parameter const float eps = 1e-6f; // default eps parameter
x = ggml_group_norm(ctx, x, num_groups, eps); x = ggml_group_norm(ctx, x, num_groups, eps);
if (w != NULL && b != NULL) { if (w != NULL && b != NULL) {
x = ggml_mul(ctx, x, w); x = ggml_mul(ctx, x, w);
// b = ggml_repeat(ctx, b, x); // b = ggml_repeat(ctx, b, x);
@ -781,7 +781,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_group_norm(struct ggml_context* ct
} }
__STATIC_INLINE__ void ggml_backend_tensor_get_and_sync(ggml_backend_t backend, const struct ggml_tensor* tensor, void* data, size_t offset, size_t size) { __STATIC_INLINE__ void ggml_backend_tensor_get_and_sync(ggml_backend_t backend, const struct ggml_tensor* tensor, void* data, size_t offset, size_t size) {
#if defined (SD_USE_CUBLAS) || defined (SD_USE_SYCL) #if defined(SD_USE_CUBLAS) || defined(SD_USE_SYCL)
if (!ggml_backend_is_cpu(backend)) { if (!ggml_backend_is_cpu(backend)) {
ggml_backend_tensor_get_async(backend, tensor, data, offset, size); ggml_backend_tensor_get_async(backend, tensor, data, offset, size);
ggml_backend_synchronize(backend); ggml_backend_synchronize(backend);
@ -889,7 +889,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_timestep_embedding(
struct ggml_context* ctx, struct ggml_context* ctx,
struct ggml_tensor* timesteps, struct ggml_tensor* timesteps,
int dim, int dim,
int max_period = 10000, int max_period = 10000,
float time_factor = 1.0f) { float time_factor = 1.0f) {
timesteps = ggml_scale(ctx, timesteps, time_factor); timesteps = ggml_scale(ctx, timesteps, time_factor);
return ggml_timestep_embedding(ctx, timesteps, dim, max_period); return ggml_timestep_embedding(ctx, timesteps, dim, max_period);

10
lora.hpp
View File

@ -10,10 +10,10 @@ struct LoraModel : public GGMLRunner {
std::map<std::string, struct ggml_tensor*> lora_tensors; std::map<std::string, struct ggml_tensor*> lora_tensors;
std::string file_path; std::string file_path;
ModelLoader model_loader; ModelLoader model_loader;
bool load_failed = false; bool load_failed = false;
bool applied = false; bool applied = false;
std::vector<int> zero_index_vec = {0}; std::vector<int> zero_index_vec = {0};
ggml_tensor* zero_index = NULL; ggml_tensor* zero_index = NULL;
LoraModel(ggml_backend_t backend, LoraModel(ggml_backend_t backend,
ggml_type wtype, ggml_type wtype,
@ -72,8 +72,8 @@ struct LoraModel : public GGMLRunner {
ggml_tensor* to_f32(ggml_context* ctx, ggml_tensor* a) { ggml_tensor* to_f32(ggml_context* ctx, ggml_tensor* a) {
auto out = ggml_reshape_1d(ctx, a, ggml_nelements(a)); auto out = ggml_reshape_1d(ctx, a, ggml_nelements(a));
out = ggml_get_rows(ctx, out, zero_index); out = ggml_get_rows(ctx, out, zero_index);
out = ggml_reshape(ctx, out, a); out = ggml_reshape(ctx, out, a);
return out; return out;
} }

19
model.cpp
View File

@ -567,10 +567,10 @@ uint16_t f8_e4m3_to_f16(uint8_t f8) {
return ggml_fp32_to_fp16(NAN); return ggml_fp32_to_fp16(NAN);
} }
uint32_t sign = f8 & 0x80; uint32_t sign = f8 & 0x80;
uint32_t exponent = (f8 & 0x78) >> 3; uint32_t exponent = (f8 & 0x78) >> 3;
uint32_t mantissa = f8 & 0x07; uint32_t mantissa = f8 & 0x07;
uint32_t result = sign << 24; uint32_t result = sign << 24;
if (exponent == 0) { if (exponent == 0) {
if (mantissa > 0) { if (mantissa > 0) {
exponent = 0x7f - exponent_bias; exponent = 0x7f - exponent_bias;
@ -1399,8 +1399,8 @@ ggml_type ModelLoader::get_sd_wtype() {
if (tensor_storage.name.find(".weight") != std::string::npos && if (tensor_storage.name.find(".weight") != std::string::npos &&
(tensor_storage.name.find("time_embed") != std::string::npos || (tensor_storage.name.find("time_embed") != std::string::npos ||
tensor_storage.name.find("context_embedder") != std::string::npos || tensor_storage.name.find("context_embedder") != std::string::npos ||
tensor_storage.name.find("time_in") != std::string::npos)) { tensor_storage.name.find("time_in") != std::string::npos)) {
return tensor_storage.type; return tensor_storage.type;
} }
} }
@ -1414,9 +1414,9 @@ ggml_type ModelLoader::get_conditioner_wtype() {
} }
if ((tensor_storage.name.find("text_encoders") == std::string::npos && if ((tensor_storage.name.find("text_encoders") == std::string::npos &&
tensor_storage.name.find("cond_stage_model") == std::string::npos && tensor_storage.name.find("cond_stage_model") == std::string::npos &&
tensor_storage.name.find("te.text_model.") == std::string::npos && tensor_storage.name.find("te.text_model.") == std::string::npos &&
tensor_storage.name.find("conditioner") == std::string::npos)) { tensor_storage.name.find("conditioner") == std::string::npos)) {
continue; continue;
} }
@ -1427,7 +1427,6 @@ ggml_type ModelLoader::get_conditioner_wtype() {
return GGML_TYPE_COUNT; return GGML_TYPE_COUNT;
} }
ggml_type ModelLoader::get_diffusion_model_wtype() { ggml_type ModelLoader::get_diffusion_model_wtype() {
for (auto& tensor_storage : tensor_storages) { for (auto& tensor_storage : tensor_storages) {
if (is_unused_tensor(tensor_storage.name)) { if (is_unused_tensor(tensor_storage.name)) {
@ -1440,8 +1439,8 @@ ggml_type ModelLoader::get_diffusion_model_wtype() {
if (tensor_storage.name.find(".weight") != std::string::npos && if (tensor_storage.name.find(".weight") != std::string::npos &&
(tensor_storage.name.find("time_embed") != std::string::npos || (tensor_storage.name.find("time_embed") != std::string::npos ||
tensor_storage.name.find("context_embedder") != std::string::npos || tensor_storage.name.find("context_embedder") != std::string::npos ||
tensor_storage.name.find("time_in") != std::string::npos)) { tensor_storage.name.find("time_in") != std::string::npos)) {
return tensor_storage.type; return tensor_storage.type;
} }
} }

1
model.h
View File

@ -165,4 +165,3 @@ public:
}; };
#endif // __MODEL_H__ #endif // __MODEL_H__

21
stable-diffusion.cpp
View File

@ -69,11 +69,10 @@ public:
ggml_backend_t clip_backend = NULL; ggml_backend_t clip_backend = NULL;
ggml_backend_t control_net_backend = NULL; ggml_backend_t control_net_backend = NULL;
ggml_backend_t vae_backend = NULL; ggml_backend_t vae_backend = NULL;
ggml_type model_wtype = GGML_TYPE_COUNT; ggml_type model_wtype = GGML_TYPE_COUNT;
ggml_type conditioner_wtype = GGML_TYPE_COUNT; ggml_type conditioner_wtype = GGML_TYPE_COUNT;
ggml_type diffusion_model_wtype = GGML_TYPE_COUNT; ggml_type diffusion_model_wtype = GGML_TYPE_COUNT;
ggml_type vae_wtype = GGML_TYPE_COUNT; ggml_type vae_wtype = GGML_TYPE_COUNT;
SDVersion version; SDVersion version;
bool vae_decode_only = false; bool vae_decode_only = false;
@ -171,7 +170,7 @@ public:
backend = ggml_backend_cpu_init(); backend = ggml_backend_cpu_init();
} }
#ifdef SD_USE_FLASH_ATTENTION #ifdef SD_USE_FLASH_ATTENTION
#if defined(SD_USE_CUBLAS) || defined(SD_USE_METAL) || defined (SD_USE_SYCL) #if defined(SD_USE_CUBLAS) || defined(SD_USE_METAL) || defined(SD_USE_SYCL)
LOG_WARN("Flash Attention not supported with GPU Backend"); LOG_WARN("Flash Attention not supported with GPU Backend");
#else #else
LOG_INFO("Flash Attention enabled"); LOG_INFO("Flash Attention enabled");
@ -243,10 +242,10 @@ public:
vae_wtype = wtype; vae_wtype = wtype;
} }
} else { } else {
model_wtype = wtype; model_wtype = wtype;
conditioner_wtype = wtype; conditioner_wtype = wtype;
diffusion_model_wtype = wtype; diffusion_model_wtype = wtype;
vae_wtype = wtype; vae_wtype = wtype;
} }
if (version == VERSION_SDXL) { if (version == VERSION_SDXL) {
@ -290,7 +289,7 @@ public:
first_stage_model->alloc_params_buffer(); first_stage_model->alloc_params_buffer();
first_stage_model->get_param_tensors(tensors, "first_stage_model"); first_stage_model->get_param_tensors(tensors, "first_stage_model");
} else { } else {
clip_backend = backend; clip_backend = backend;
bool use_t5xxl = false; bool use_t5xxl = false;
if (version == VERSION_SD3_2B || version == VERSION_FLUX_DEV || version == VERSION_FLUX_SCHNELL) { if (version == VERSION_SD3_2B || version == VERSION_FLUX_DEV || version == VERSION_FLUX_SCHNELL) {
use_t5xxl = true; use_t5xxl = true;
@ -508,7 +507,7 @@ public:
LOG_INFO("running in Flux FLOW mode"); LOG_INFO("running in Flux FLOW mode");
float shift = 1.15f; float shift = 1.15f;
if (version == VERSION_FLUX_SCHNELL) { if (version == VERSION_FLUX_SCHNELL) {
shift = 1.0f; // TODO: validate shift = 1.0f; // TODO: validate
} }
denoiser = std::make_shared<FluxFlowDenoiser>(shift); denoiser = std::make_shared<FluxFlowDenoiser>(shift);
} else if (is_using_v_parameterization) { } else if (is_using_v_parameterization) {

58
stable-diffusion.h
View File

@ -55,37 +55,37 @@ enum schedule_t {
// same as enum ggml_type // same as enum ggml_type
enum sd_type_t { enum sd_type_t {
SD_TYPE_F32 = 0, SD_TYPE_F32 = 0,
SD_TYPE_F16 = 1, SD_TYPE_F16 = 1,
SD_TYPE_Q4_0 = 2, SD_TYPE_Q4_0 = 2,
SD_TYPE_Q4_1 = 3, SD_TYPE_Q4_1 = 3,
// SD_TYPE_Q4_2 = 4, support has been removed // SD_TYPE_Q4_2 = 4, support has been removed
// SD_TYPE_Q4_3 = 5, support has been removed // SD_TYPE_Q4_3 = 5, support has been removed
SD_TYPE_Q5_0 = 6, SD_TYPE_Q5_0 = 6,
SD_TYPE_Q5_1 = 7, SD_TYPE_Q5_1 = 7,
SD_TYPE_Q8_0 = 8, SD_TYPE_Q8_0 = 8,
SD_TYPE_Q8_1 = 9, SD_TYPE_Q8_1 = 9,
SD_TYPE_Q2_K = 10, SD_TYPE_Q2_K = 10,
SD_TYPE_Q3_K = 11, SD_TYPE_Q3_K = 11,
SD_TYPE_Q4_K = 12, SD_TYPE_Q4_K = 12,
SD_TYPE_Q5_K = 13, SD_TYPE_Q5_K = 13,
SD_TYPE_Q6_K = 14, SD_TYPE_Q6_K = 14,
SD_TYPE_Q8_K = 15, SD_TYPE_Q8_K = 15,
SD_TYPE_IQ2_XXS = 16, SD_TYPE_IQ2_XXS = 16,
SD_TYPE_IQ2_XS = 17, SD_TYPE_IQ2_XS = 17,
SD_TYPE_IQ3_XXS = 18, SD_TYPE_IQ3_XXS = 18,
SD_TYPE_IQ1_S = 19, SD_TYPE_IQ1_S = 19,
SD_TYPE_IQ4_NL = 20, SD_TYPE_IQ4_NL = 20,
SD_TYPE_IQ3_S = 21, SD_TYPE_IQ3_S = 21,
SD_TYPE_IQ2_S = 22, SD_TYPE_IQ2_S = 22,
SD_TYPE_IQ4_XS = 23, SD_TYPE_IQ4_XS = 23,
SD_TYPE_I8 = 24, SD_TYPE_I8 = 24,
SD_TYPE_I16 = 25, SD_TYPE_I16 = 25,
SD_TYPE_I32 = 26, SD_TYPE_I32 = 26,
SD_TYPE_I64 = 27, SD_TYPE_I64 = 27,
SD_TYPE_F64 = 28, SD_TYPE_F64 = 28,
SD_TYPE_IQ1_M = 29, SD_TYPE_IQ1_M = 29,
SD_TYPE_BF16 = 30, SD_TYPE_BF16 = 30,
SD_TYPE_Q4_0_4_4 = 31, SD_TYPE_Q4_0_4_4 = 31,
SD_TYPE_Q4_0_4_8 = 32, SD_TYPE_Q4_0_4_8 = 32,
SD_TYPE_Q4_0_8_8 = 33, SD_TYPE_Q4_0_8_8 = 33,