Skip to content

Commit dd71b27

Browse files
SkutteOlegSkutteOleg
committed
Handle ggml aborts
1 parent c03f2b3 commit dd71b27

File tree

3 files changed

+122
-88
lines changed

3 files changed

+122
-88
lines changed

‎.gitignore‎

Lines changed: 4 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -10,4 +10,7 @@ test/
1010
*.gguf
1111
output*.png
1212
models*
13-
*.log
13+
*.log
14+
/cmake-build-debug
15+
/.idea
16+
/cmake-build-release

‎.gitmodules‎

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,3 +1,4 @@
11
[submodule "ggml"]
22
path = ggml
3-
url = https://github.com/ggerganov/ggml.git
3+
url = https://github.com/SkutteOleg/ggml.git
4+
branch = dreamio

‎stable-diffusion.cpp‎

Lines changed: 116 additions & 86 deletions
Original file line numberDiff line numberDiff line change
@@ -15,6 +15,7 @@
1515
#include "pmid.hpp"
1616
#include "tae.hpp"
1717
#include "vae.hpp"
18+
#include <setjmp.h>
1819

1920
#define STB_IMAGE_IMPLEMENTATION
2021
#define STB_IMAGE_STATIC
@@ -1061,7 +1062,7 @@ SDError new_sd_ctx(sd_ctx_t** sd_ctx,
10611062
const char* model_path_c_str,
10621063
const char* clip_l_path_c_str,
10631064
const char* clip_g_path_c_str,
1064-
const char* t5xxl_path_c_str,
1065+
const char* t5xxl_path_c_str,
10651066
const char* diffusion_model_path_c_str,
10661067
const char* vae_path_c_str,
10671068
const char* taesd_path_c_str,
@@ -1079,33 +1080,37 @@ SDError new_sd_ctx(sd_ctx_t** sd_ctx,
10791080
bool keep_clip_on_cpu,
10801081
bool keep_control_net_cpu,
10811082
bool keep_vae_on_cpu) {
1082-
try {
1083-
*sd_ctx = (sd_ctx_t*)malloc(sizeof(sd_ctx_t));
1084-
if (*sd_ctx == NULL) {
1085-
return SD_ERROR_MEMORY_ALLOCATION;
1086-
}
1083+
ggml_error_jmp_set = 1;
1084+
if (setjmp(ggml_error_jmp_buf) == 0) {
1085+
try {
1086+
*sd_ctx = (sd_ctx_t*)malloc(sizeof(sd_ctx_t));
1087+
if (*sd_ctx == NULL) {
1088+
ggml_error_jmp_set = 0;
1089+
return SD_ERROR_MEMORY_ALLOCATION;
1090+
}
10871091

1088-
std::string model_path(model_path_c_str);
1089-
std::string clip_l_path(clip_l_path_c_str);
1090-
std::string clip_g_path(clip_g_path_c_str);
1091-
std::string t5xxl_path(t5xxl_path_c_str);
1092-
std::string diffusion_model_path(diffusion_model_path_c_str);
1093-
std::string vae_path(vae_path_c_str);
1094-
std::string taesd_path(taesd_path_c_str);
1095-
std::string control_net_path(control_net_path_c_str);
1096-
std::string embd_path(embed_dir_c_str);
1097-
std::string id_embd_path(id_embed_dir_c_str);
1098-
std::string lora_model_dir(lora_model_dir_c_str);
1099-
1100-
(*sd_ctx)->sd = new StableDiffusionGGML(n_threads,
1101-
vae_decode_only,
1102-
free_params_immediately,
1103-
lora_model_dir,
1104-
rng_type);
1105-
if ((*sd_ctx)->sd == NULL) {
1106-
free(*sd_ctx);
1107-
return SD_ERROR_MEMORY_ALLOCATION;
1108-
}
1092+
std::string model_path(model_path_c_str);
1093+
std::string clip_l_path(clip_l_path_c_str);
1094+
std::string clip_g_path(clip_g_path_c_str);
1095+
std::string t5xxl_path(t5xxl_path_c_str);
1096+
std::string diffusion_model_path(diffusion_model_path_c_str);
1097+
std::string vae_path(vae_path_c_str);
1098+
std::string taesd_path(taesd_path_c_str);
1099+
std::string control_net_path(control_net_path_c_str);
1100+
std::string embd_path(embed_dir_c_str);
1101+
std::string id_embd_path(id_embed_dir_c_str);
1102+
std::string lora_model_dir(lora_model_dir_c_str);
1103+
1104+
(*sd_ctx)->sd = new StableDiffusionGGML(n_threads,
1105+
vae_decode_only,
1106+
free_params_immediately,
1107+
lora_model_dir,
1108+
rng_type);
1109+
if ((*sd_ctx)->sd == NULL) {
1110+
free(*sd_ctx);
1111+
ggml_error_jmp_set = 0;
1112+
return SD_ERROR_MEMORY_ALLOCATION;
1113+
}
11091114

11101115
if (!(*sd_ctx)->sd->load_from_file(model_path,
11111116
clip_l_path,
@@ -1124,15 +1129,24 @@ SDError new_sd_ctx(sd_ctx_t** sd_ctx,
11241129
keep_vae_on_cpu)) {
11251130
delete (*sd_ctx)->sd;
11261131
free(*sd_ctx);
1132+
ggml_error_jmp_set = 0;
11271133
return SD_ERROR_PROCESSING;
11281134
}
11291135

1130-
return SD_SUCCESS;
1131-
} catch (const std::exception& e) {
1132-
LOG_ERROR("Exception in new_sd_ctx: %s", e.what());
1133-
return SD_ERROR_PROCESSING;
1134-
} catch (...) {
1135-
LOG_ERROR("Unknown exception in new_sd_ctx");
1136+
ggml_error_jmp_set = 0;
1137+
return SD_SUCCESS;
1138+
} catch (const std::exception& e) {
1139+
LOG_ERROR("Exception in new_sd_ctx: %s", e.what());
1140+
ggml_error_jmp_set = 0;
1141+
return SD_ERROR_PROCESSING;
1142+
} catch (...) {
1143+
LOG_ERROR("Unknown exception in new_sd_ctx");
1144+
ggml_error_jmp_set = 0;
1145+
return SD_ERROR_PROCESSING;
1146+
}
1147+
} else {
1148+
LOG_ERROR("GGML error in new_sd_ctx: %s", ggml_get_error_message());
1149+
ggml_error_jmp_set = 0;
11361150
return SD_ERROR_PROCESSING;
11371151
}
11381152
}
@@ -1469,55 +1483,59 @@ SDError txt2img(sd_ctx_t* sd_ctx,
14691483
float skip_layer_start,
14701484
float skip_layer_end) {
14711485
struct ggml_context* work_ctx = NULL;
1472-
try {
1473-
LOG_DEBUG("txt2img %dx%d", width, height);
1474-
if (sd_ctx == NULL) {
1475-
return SD_ERROR_INVALID_CONTEXT;
1476-
}
1486+
ggml_error_jmp_set = 1;
1487+
if (setjmp(ggml_error_jmp_buf) == 0) {
1488+
try {
1489+
LOG_DEBUG("txt2img %dx%d", width, height);
1490+
if (sd_ctx == NULL) {
1491+
ggml_error_jmp_set = 0;
1492+
return SD_ERROR_INVALID_CONTEXT;
1493+
}
14771494

1478-
struct ggml_init_params params;
1479-
params.mem_size = static_cast<size_t>(10 * 1024 * 1024); // 10 MB
1480-
if (sd_ctx->sd->version == VERSION_SD3_2B || sd_ctx->sd->version == VERSION_SD3_5_8B || sd_ctx->sd->version == VERSION_SD3_5_2B) {
1481-
params.mem_size *= 3;
1495+
struct ggml_init_params params;
1496+
params.mem_size = static_cast<size_t>(10 * 1024 * 1024); // 10 MB
1497+
if (sd_ctx->sd->version == VERSION_SD3_2B || sd_ctx->sd->version == VERSION_SD3_5_8B || sd_ctx->sd->version == VERSION_SD3_5_2B) {
1498+
params.mem_size *= 3;
14821499
}
14831500
if (sd_ctx->sd->version == VERSION_FLUX_DEV || sd_ctx->sd->version == VERSION_FLUX_SCHNELL) {
14841501
params.mem_size *= 4;
1485-
}
1486-
if (sd_ctx->sd->stacked_id) {
1487-
params.mem_size += static_cast<size_t>(10 * 1024 * 1024); // 10 MB
1488-
}
1489-
params.mem_size += width * height * 3 * sizeof(float);
1490-
params.mem_size *= batch_count;
1491-
params.mem_buffer = NULL;
1492-
params.no_alloc = false;
1493-
// LOG_DEBUG("mem_size %u ", params.mem_size);
1494-
1495-
work_ctx = ggml_init(params);
1496-
if (!work_ctx) {
1497-
LOG_ERROR("ggml_init() failed");
1498-
return SD_ERROR_MEMORY_ALLOCATION;
1499-
}
1502+
}
1503+
if (sd_ctx->sd->stacked_id) {
1504+
params.mem_size += static_cast<size_t>(10 * 1024 * 1024); // 10 MB
1505+
}
1506+
params.mem_size += width * height * 3 * sizeof(float);
1507+
params.mem_size *= batch_count;
1508+
params.mem_buffer = NULL;
1509+
params.no_alloc = false;
1510+
// LOG_DEBUG("mem_size %u ", params.mem_size);
1511+
1512+
work_ctx = ggml_init(params);
1513+
if (!work_ctx) {
1514+
LOG_ERROR("ggml_init() failed");
1515+
ggml_error_jmp_set = 0;
1516+
return SD_ERROR_MEMORY_ALLOCATION;
1517+
}
15001518

1501-
size_t t0 = ggml_time_ms();
1519+
size_t t0 = ggml_time_ms();
15021520

1503-
std::vector<float> sigmas = sd_ctx->sd->denoiser->get_sigmas(sample_steps);
1521+
std::vector<float> sigmas = sd_ctx->sd->denoiser->get_sigmas(sample_steps);
15041522

1505-
int C = 4;
1506-
if (sd_ctx->sd->version == VERSION_SD3_2B || sd_ctx->sd->version == VERSION_SD3_5_8B || sd_ctx->sd->version == VERSION_SD3_5_2B) {
1523+
int C = 4;
1524+
if (sd_ctx->sd->version == VERSION_SD3_2B || sd_ctx->sd->version == VERSION_SD3_5_8B || sd_ctx->sd->version == VERSION_SD3_5_2B) {
15071525
C = 16;
15081526
} else if (sd_ctx->sd->version == VERSION_FLUX_DEV || sd_ctx->sd->version == VERSION_FLUX_SCHNELL) {
1509-
C = 16;
1510-
}
1511-
int W = width / 8;
1512-
int H = height / 8;
1513-
ggml_tensor* init_latent = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, W, H, C, 1);
1514-
if (sd_ctx->sd->version == VERSION_SD3_2B || sd_ctx->sd->version == VERSION_SD3_5_8B || sd_ctx->sd->version == VERSION_SD3_5_2B) {
1515-
ggml_set_f32(init_latent, 0.0609f);
1527+
C = 16;
1528+
}
1529+
int W = width / 8;
1530+
int H = height / 8;
1531+
ggml_tensor* init_latent = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, W, H, C, 1);
1532+
if (sd_ctx->sd->version == VERSION_SD3_2B || sd_ctx->sd->version == VERSION_SD3_5_8B || sd_ctx->sd->version == VERSION_SD3_5_2B) {
1533+
ggml_set_f32(init_latent, 0.0609f);
15161534
} else if (sd_ctx->sd->version == VERSION_FLUX_DEV || sd_ctx->sd->version == VERSION_FLUX_SCHNELL) {
15171535
ggml_set_f32(init_latent, 0.1159f);
1518-
} else {
1519-
ggml_set_f32(init_latent, 0.f);
1520-
}
1536+
} else {
1537+
ggml_set_f32(init_latent, 0.f);
1538+
}
15211539

15221540
*result_images = generate_image(sd_ctx,
15231541
work_ctx,
@@ -1543,30 +1561,42 @@ SDError txt2img(sd_ctx_t* sd_ctx,
15431561
skip_layer_start,
15441562
skip_layer_end);
15451563

1546-
if (*result_images == NULL) {
1547-
ggml_free(work_ctx);
1548-
return SD_ERROR_PROCESSING;
1549-
}
1564+
if (*result_images == NULL) {
1565+
ggml_free(work_ctx);
1566+
ggml_error_jmp_set = 0;
1567+
return SD_ERROR_PROCESSING;
1568+
}
15501569

1551-
*result_count = batch_count;
1570+
*result_count = batch_count;
15521571

1553-
size_t t1 = ggml_time_ms();
1572+
size_t t1 = ggml_time_ms();
15541573

1555-
LOG_INFO("txt2img completed in %.2fs", (t1 - t0) * 1.0f / 1000);
1574+
LOG_INFO("txt2img completed in %.2fs", (t1 - t0) * 1.0f / 1000);
15561575

1557-
ggml_free(work_ctx);
1558-
return SD_SUCCESS;
1559-
} catch (const std::exception& e) {
1560-
LOG_ERROR("Exception in txt2img: %s", e.what());
1561-
if (work_ctx) {
15621576
ggml_free(work_ctx);
1577+
ggml_error_jmp_set = 0;
1578+
return SD_SUCCESS;
1579+
} catch (const std::exception& e) {
1580+
LOG_ERROR("Exception in txt2img: %s", e.what());
1581+
if (work_ctx) {
1582+
ggml_free(work_ctx);
1583+
}
1584+
ggml_error_jmp_set = 0;
1585+
return SD_ERROR_PROCESSING;
1586+
} catch (...) {
1587+
LOG_ERROR("Unknown exception in txt2img");
1588+
if (work_ctx) {
1589+
ggml_free(work_ctx);
1590+
}
1591+
ggml_error_jmp_set = 0;
1592+
return SD_ERROR_PROCESSING;
15631593
}
1564-
return SD_ERROR_PROCESSING;
1565-
} catch (...) {
1566-
LOG_ERROR("Unknown exception in txt2img");
1594+
} else {
1595+
LOG_ERROR("GGML error in txt2img: %s", ggml_get_error_message());
15671596
if (work_ctx) {
15681597
ggml_free(work_ctx);
15691598
}
1599+
ggml_error_jmp_set = 0;
15701600
return SD_ERROR_PROCESSING;
15711601
}
15721602
}

0 commit comments

Comments
 (0)