root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 02:44:36 +00:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

gguf-py : add support for I8, I16 and I32 (#6045)

Browse Source 
* Refactor dtype handling to be extensible

This code is equivalent as before, but now it is prepared to easily add
more NumPy dtypes.

* Add support for I8, I16 and I32

These types are allowed in the GGUF specification.

* Add support for I8, I16 and I32 to gguf_writer

* Add support for I8, I16, I32 to gguf_reader
This commit is contained in:
Ondřej Čertík 2024-03-14 04:40:14 -06:00 committed by GitHub
parent 3fe8d7a17f
commit 3ca23481dd
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 27 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
gguf-py/gguf/constants.py
View File

@ -661,6 +661,9 @@ class GGMLQuantizationType(IntEnum):
IQ3_S = 21
IQ2_S = 22
IQ4_XS = 23
I8 = 24
I16 = 25
I32 = 26
class GGUFEndian(IntEnum):
@ -727,6 +730,9 @@ GGML_QUANT_SIZES = {
GGMLQuantizationType.IQ3_S: (256, 2 + QK_K // 4 + QK_K // 8 + QK_K // 32 + 4),
GGMLQuantizationType.IQ2_S: (256, 2 + QK_K // 4 + QK_K // 16),
GGMLQuantizationType.IQ4_XS: (256, 2 + 2 + QK_K // 2 + QK_K // 64),
GGMLQuantizationType.I8: (1, 1),
GGMLQuantizationType.I16: (1, 2),
GGMLQuantizationType.I32: (1, 4),
}

9
gguf-py/gguf/gguf_reader.py
View File

@ -248,6 +248,15 @@ class GGUFReader:
elif ggml_type == GGMLQuantizationType.F16:
item_count = n_elems
item_type = np.float16
elif ggml_type == GGMLQuantizationType.I8:
item_count = n_elems
item_type = np.int8
elif ggml_type == GGMLQuantizationType.I16:
item_count = n_elems
item_type = np.int16
elif ggml_type == GGMLQuantizationType.I32:
item_count = n_elems
item_type = np.int32
else:
item_count = n_bytes
item_type = np.uint8

16
gguf-py/gguf/gguf_writer.py
View File

@ -196,9 +196,6 @@ class GGUFWriter:
if self.state is not WriterState.EMPTY:
raise ValueError(f'Expected output file to be empty, got {self.state}')
if raw_dtype is None and tensor_dtype not in (np.float32, np.float16):
raise ValueError("Only F32 and F16 tensors are supported for now")
encoded_name = name.encode("utf8")
self.ti_data += self._pack("Q", len(encoded_name))
self.ti_data += encoded_name
@ -207,7 +204,18 @@ class GGUFWriter:
for i in range(n_dims):
self.ti_data += self._pack("Q", tensor_shape[n_dims - 1 - i])
if raw_dtype is None:
dtype = GGMLQuantizationType.F32 if tensor_dtype == np.float32 else GGMLQuantizationType.F16
if tensor_shape == np.float32:
dtype = GGMLQuantizationType.F32
elif tensor_dtype == np.float16:
dtype = GGMLQuantizationType.F16
elif tensor_dtype == np.int8:
dtype = GGMLQuantizationType.I8
elif tensor_dtype == np.int16:
dtype = GGMLQuantizationType.I16
elif tensor_dtype == np.int32:
dtype = GGMLQuantizationType.I32
else:
raise ValueError("Only F32, F16, I8, I16, I32 tensors are supported for now")
else:
dtype = raw_dtype
self.ti_data += self._pack("I", dtype)