Skip to content

Fix conversion between F64 and F32 #726

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 2 commits into
base: master
Choose a base branch
from
+53 −0
Open

Fix conversion between F64 and F32 #726

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
5a508b0
improve f64 support (for convert mostly)
stduhpf Jul 8, 2025
0aa6ca7
f64<->quant
stduhpf Jul 8, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
53 changes: 53 additions & 0 deletions model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -835,6 +835,12 @@ void convert_tensor(void* src,
} else if (src_type == GGML_TYPE_F32) {
if (dst_type == GGML_TYPE_F16) {
ggml_fp32_to_fp16_row((float*)src, (ggml_fp16_t*)dst, n);
} else if (dst_type == GGML_TYPE_F64) {
double* ddst = (double*)dst;
float* fsrc = (float*)src;
for (int64_t i = 0; i < n; i++) {
ddst[i] = (double)(fsrc[i]);
}
} else {
std::vector<float> imatrix(n_per_row, 1.0f); // dummy importance matrix
const float* im = imatrix.data();
Expand All @@ -843,6 +849,12 @@ void convert_tensor(void* src,
} else if (dst_type == GGML_TYPE_F32) {
if (src_type == GGML_TYPE_F16) {
ggml_fp16_to_fp32_row((ggml_fp16_t*)src, (float*)dst, n);
} else if (src_type == GGML_TYPE_F64) {
float* fdst = (float*)dst;
double* dsrc = (double*)src;
for (int64_t i = 0; i < n; i++) {
fdst[i] = (float)(dsrc[i]);
}
} else {
auto qtype = ggml_get_type_traits(src_type);
if (qtype->to_float == NULL) {
Expand All @@ -851,6 +863,47 @@ void convert_tensor(void* src,
}
qtype->to_float(src, (float*)dst, n);
}
} else if (src_type == GGML_TYPE_F64) {
if (dst_type == GGML_TYPE_F16) {
// ggml_fp32_to_fp16_row((float*)src, (ggml_fp16_t*)dst, n);
ggml_fp16_t* fdst = (ggml_fp16_t*)dst;
double* dsrc = (double*)src;
for (int64_t i = 0; i < n; i++) {
fdst[i] = ggml_fp32_to_fp16((float)dsrc[i]);
}
} else {
std::vector<float> imatrix(n_per_row, 1.0f); // dummy importance matrix
const float* im = imatrix.data();
float* fsrc = (float*)malloc(n * sizeof(float));
double* dsrc = (double*)src;
for (int64_t i = 0; i < n; i++) {
fsrc[i] = (float)(dsrc[i]);
}
ggml_quantize_chunk(dst_type, fsrc, dst, 0, nrows, n_per_row, im);
free(fsrc);
}
} else if (dst_type == GGML_TYPE_F64) {
if (src_type == GGML_TYPE_F16) {
// ggml_fp16_to_fp32_row((ggml_fp16_t*)src, (float*)dst, n);
double* ddst = (double*)dst;
ggml_fp16_t* fsrc = (ggml_fp16_t*)src;
for (int64_t i = 0; i < n; i++) {
ddst[i] = (double)ggml_fp16_to_fp32(fsrc[i]);
}
} else {
auto qtype = ggml_get_type_traits(src_type);
if (qtype->to_float == NULL) {
throw std::runtime_error(format("type %s unsupported for integer quantization: no dequantization available",
ggml_type_name(src_type)));
}
float* fdst = (float*)malloc(n * sizeof(float));
qtype->to_float(src, fdst, n);
double* ddst = (double*)dst;
for (int64_t i = 0; i < n; i++) {
ddst[i] = (double)(fdst[i]);
}
free(fdst);
}
} else {
// src_type == GGML_TYPE_F16 => dst_type is quantized
// src_type is quantized => dst_type == GGML_TYPE_F16 or dst_type is quantized
Expand Down
Loading