Fix conversion of unnormalized BF16->BF16 weights

convert_hf_to_gguf.py

@@ -316,7 +316,7 @@ def prepare_tensors(self):
                 if self.ftype != gguf.LlamaFileType.ALL_F32 and extra_f16 and not extra_f32:
                     if self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
                         data = gguf.quantize_bf16(data)
-                        assert data.dtype == np.int16
+                        assert data.dtype == np.uint16
                         data_qtype = gguf.GGMLQuantizationType.BF16
 
                     elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0 and gguf.can_quantize_to_q8_0(data):

ggml/include/ggml.h

@@ -349,6 +349,7 @@ extern "C" {
     GGML_API ggml_bf16_t ggml_fp32_to_bf16(float);
     GGML_API float       ggml_bf16_to_fp32(ggml_bf16_t);  // consider just doing << 16
     GGML_API void        ggml_bf16_to_fp32_row(const ggml_bf16_t *, float *, int64_t);
+    GGML_API void        ggml_fp32_to_bf16_row_reference(const float *, ggml_bf16_t *, int64_t);
     GGML_API void        ggml_fp32_to_bf16_row(const float *, ggml_bf16_t *, int64_t);
 
     struct ggml_object;

ggml/src/ggml-impl.h

@@ -80,8 +80,9 @@ static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
 /**
  * Converts float32 to brain16.
  *
- * This function is binary identical to AMD Zen4 VCVTNEPS2BF16.
- * Subnormals shall be flushed to zero, and NANs will be quiet.
+ * This is binary identical with Google Brain float conversion.
+ * Floats shall round to nearest even, and NANs shall be quiet.
+ * Subnormals aren't flushed to zero, except perhaps when used.
  * This code should vectorize nicely if using modern compilers.
  */
 static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
@@ -95,10 +96,6 @@ static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
         h.bits = (u.i >> 16) | 64; /* force to quiet */
         return h;
     }
-    if (!(u.i & 0x7f800000)) { /* subnormal */
-        h.bits = (u.i & 0x80000000) >> 16; /* flush to zero */
-        return h;
-    }
     h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
     return h;
 }