feat: do not convert bf16 to f32 (#1055)

This commit is contained in:
leejet
2025-12-06 23:55:51 +08:00
committed by GitHub
parent 689e44c9a8
commit bfbb929790
4 changed files with 6 additions and 33 deletions

View File

@@ -168,7 +168,6 @@ struct TensorStorage {
std::string name;
ggml_type type = GGML_TYPE_F32;
ggml_type expected_type = GGML_TYPE_COUNT;
bool is_bf16 = false;
bool is_f8_e4m3 = false;
bool is_f8_e5m2 = false;
bool is_f64 = false;
@@ -202,7 +201,7 @@ struct TensorStorage {
}
int64_t nbytes_to_read() const {
if (is_bf16 || is_f8_e4m3 || is_f8_e5m2) {
if (is_f8_e4m3 || is_f8_e5m2) {
return nbytes() / 2;
} else if (is_f64 || is_i64) {
return nbytes() * 2;
@@ -250,9 +249,7 @@ struct TensorStorage {
std::string to_string() const {
std::stringstream ss;
const char* type_name = ggml_type_name(type);
if (is_bf16) {
type_name = "bf16";
} else if (is_f8_e4m3) {
if (is_f8_e4m3) {
type_name = "f8_e4m3";
} else if (is_f8_e5m2) {
type_name = "f8_e5m2";