root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-24 10:24:35 +00:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

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

Browse Source 
* 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>
This commit is contained in:
Sigbjørn Skjæret 2024-08-02 21:11:39 +02:00 committed by GitHub
parent e09a800f9a
commit b72c20b85c
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
5 changed files with 20 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
convert_hf_to_gguf.py
View File

@ -316,7 +316,7 @@ class Model:
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):

1
ggml/include/ggml.h
View File

@ -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;

9
ggml/src/ggml-impl.h
View File

@ -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;
}

11
ggml/src/ggml.c
View File

@ -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:

14
gguf-py/gguf/quants.py
View File

@ -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):