Add B/F16 to get_ftype()

2025-11-29 23:52:51 +00:00 · 2025-11-29 23:52:51 +00:00 · b97cda6289
parent 229109f329
commit b97cda6289
1 changed files with 6 additions and 1 deletions
--- a/tools/quantize/quantize.cpp
+++ b/tools/quantize/quantize.cpp
@ -512,7 +512,12 @@ static const char * get_ftype(const float bpw) {
        {4.5000, "Q4_K"},
        {5.5000, "Q5_K"},
        {6.5625, "Q6_K"},
-        {8.5000, "Q8_0"}
+        {8.5000, "Q8_0"},
+#ifdef GGML_USE_METAL
+        {16.0000, "F16"}
+#else
+        {16.0000, "BF16"}
+#endif
    };

    return quant_bpw.lower_bound(bpw)->second;