Fix conversion of unnormalized BF16->BF16 weights (ggerganov#7843)

* add truncate_bf16

* truncate intermediate fp32 if converting bf16 to bf16

* fix masking in __compute_fp32_to_bf16

* np.int16 no longer used

* missing cast and additional numpy 2.x fix

* ggml-impl : do not flush bf16 subnormals to zero

* ggml : add reference fp32 to bf16 conversion

The fast version is no longer equivalent for all platforms
because of the handling of subnormal values.

* gguf-py : remove flush to zero for bf16 subnormals

* gguf-py : remove float32 truncation to bf16

Rounding achieves the same thing in the cases where this was used.

* missed prototype update in merge

* merge cleanup

---------

Co-authored-by: Francis Couture-Harpin <git@compilade.net>

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_ref(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;
 }

ggml/src/ggml.c

@@ -480,9 +480,16 @@ void ggml_bf16_to_fp32_row(const ggml_bf16_t * x, float * y, int64_t n) {
     }
 }
 
+void ggml_fp32_to_bf16_row_ref(const float * x, ggml_bf16_t * y, int64_t n) {
+    for (int i = 0; i < n; i++) {
+        y[i] = ggml_compute_fp32_to_bf16(x[i]);
+    }
+}
+
 void ggml_fp32_to_bf16_row(const float * x, ggml_bf16_t * y, int64_t n) {
   int i = 0;
 #if defined(__AVX512BF16__)
+  // subnormals are flushed to zero on this platform
   for (; i + 32 <= n; i += 32) {
         _mm512_storeu_si512(
             (__m512i *)(y + i),
@@ -962,7 +969,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .is_quantized             = false,
         .to_float                 = (ggml_to_float_t) ggml_bf16_to_fp32_row,
         .from_float               = (ggml_from_float_t) ggml_fp32_to_bf16_row,
-        .from_float_ref           = (ggml_from_float_t) ggml_fp32_to_bf16_row,
+        .from_float_ref           = (ggml_from_float_t) ggml_fp32_to_bf16_row_ref,
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_bf16,
         .vec_dot_type             = GGML_TYPE_BF16,
         .nrows                    = 1,
@@ -20650,7 +20657,7 @@ size_t ggml_quantize_chunk(
         case GGML_TYPE_BF16:
             {
                 size_t elemsize = sizeof(ggml_bf16_t);
-                ggml_fp32_to_bf16_row(src + start, (ggml_bf16_t *)dst + start, n);
+                ggml_fp32_to_bf16_row_ref(src + start, (ggml_bf16_t *)dst + start, n);
                 result = n * elemsize;
             } break;
         case GGML_TYPE_F32:

gguf-py/gguf/quants.py

@@ -25,14 +25,12 @@ def quant_shape_from_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizati
 
 # same as ggml_compute_fp32_to_bf16 in ggml-impl.h
 def __compute_fp32_to_bf16(n: np.ndarray) -> np.ndarray:
-    n = n.astype(np.float32, copy=False).view(np.int32)
+    n = n.astype(np.float32, copy=False).view(np.uint32)
     # force nan to quiet
-    n = np.where((n & 0x7fffffff) > 0x7f800000, (n & 0xffff0000) | (64 << 16), n)
-    # flush subnormals to zero
-    n = np.where((n & 0x7f800000) == 0, n & 0x80000000, n)
+    n = np.where((n & 0x7fffffff) > 0x7f800000, (n & np.uint32(0xffff0000)) | np.uint32(64 << 16), n)
     # round to nearest even
-    n = (n + (0x7fff + ((n >> 16) & 1))) >> 16
-    return n.astype(np.int16)
+    n = (np.uint64(n) + (0x7fff + ((n >> 16) & 1))) >> 16
+    return n.astype(np.uint16)
 
 
 # This is faster than np.vectorize and np.apply_along_axis because it works on more than one row at a time
@@ -49,10 +47,10 @@ def __apply_over_grouped_rows(func: Callable[[np.ndarray], np.ndarray], arr: np.
 
 
 def __quantize_bf16_array(n: np.ndarray) -> np.ndarray:
-    return __apply_over_grouped_rows(__compute_fp32_to_bf16, arr=n, otype=np.int16, oshape=n.shape)
+    return __apply_over_grouped_rows(__compute_fp32_to_bf16, arr=n, otype=np.uint16, oshape=n.shape)
 
 
-__quantize_bf16_lazy = LazyNumpyTensor._wrap_fn(__quantize_bf16_array, meta_noop=np.int16)
+__quantize_bf16_lazy = LazyNumpyTensor._wrap_fn(__quantize_bf16_array, meta_noop=np.uint16)
 
 
 def quantize_bf16(n: np.ndarray):