DarthAffe/stable-diffusion.cpp
1
0 Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-11-11 03:24:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

feat: restore legacy fp8 handling when building with upstream ggml (#2001)

Browse Source
This commit is contained in:
Wagner Bruna authored and GitHub committed 2026-09-21 00:48:30 +08:00
1 parent 137f7409bb
commit 008ca5b492
8 files changed
+106 -17
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/model_io/safetensors_io.cpp
+18 -9
View File
@@ -86,7 +86,12 @@ static ggml_type safetensors_dtype_to_ggml_type(const std::string& dtype) {
ttype = GGML_TYPE_F32;
} else if (dtype == "F64") {
ttype = GGML_TYPE_F32;
#ifndef SD_USE_UPSTREAM_GGML
#ifdef SD_USE_UPSTREAM_GGML
} else if (dtype == "F8_E4M3") {
ttype = GGML_TYPE_F16;
} else if (dtype == "F8_E5M2") {
ttype = GGML_TYPE_F16;
#else
} else if (dtype == "F8_E4M3") {
ttype = GGML_TYPE_F8_E4M3;
} else if (dtype == "F8_E5M2") {
@@ -287,12 +292,6 @@ bool read_safetensors_file(const std::string& file_path,
continue;
}
#ifdef SD_USE_UPSTREAM_GGML
if (dtype == "F8_E4M3" || dtype == "F8_E5M2") {
set_error(error, "FP8 is not supported by this ggml build (tensor '" + name + "')");
return false;
}
#endif
ggml_type type = safetensors_dtype_to_ggml_type(dtype);
if (type == GGML_TYPE_COUNT) {
set_error(error, "unsupported dtype '" + dtype + "' (tensor '" + name + "')");
@@ -376,10 +375,20 @@ bool read_safetensors_file(const std::string& file_path,
bool tensor_size_ok;
if (dtype == "F8_E4M3") {
tensor_storage.is_f8_e4m3 = true;
tensor_size_ok = (tensor_storage.nbytes() == tensor_data_size);
#ifdef SD_USE_UPSTREAM_GGML
// f8 -> f16
tensor_size_ok = (tensor_storage.nbytes() == tensor_data_size * 2);
#else
tensor_size_ok = (tensor_storage.nbytes() == tensor_data_size);
#endif
} else if (dtype == "F8_E5M2") {
tensor_storage.is_f8_e5m2 = true;
tensor_size_ok = (tensor_storage.nbytes() == tensor_data_size);
#ifdef SD_USE_UPSTREAM_GGML
// f8 -> f16
tensor_size_ok = (tensor_storage.nbytes() == tensor_data_size * 2);
#else
tensor_size_ok = (tensor_storage.nbytes() == tensor_data_size);
#endif
} else if (dtype == "F64") {
tensor_storage.is_f64 = true;
// f64 -> f32
src/model_io/tensor_storage.h
+4
View File
@@ -58,6 +58,10 @@ struct TensorStorage {
int64_t nbytes_to_read() const {
if (is_f64 || is_i64) {
return nbytes() * 2;
#ifdef SD_USE_UPSTREAM_GGML
} else if (is_f8_e4m3 || is_f8_e5m2) {
return nbytes() / 2;
#endif
} else {
return nbytes();
}