happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

gguf-py : skip endian-conversion of MXFP4 data (#17523) 

* gguf_convert_endian.py: skip MXFP4 data

* Use gguf.constants.GGML_QUANT_SIZES to determine block sizes
This commit is contained in:
Aleksei Nikiforov 2025-11-27 11:35:38 +01:00 committed by GitHub
parent b78db3bd50
commit 4fcd87cf7c
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 12 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

30
gguf-py/gguf/scripts/gguf_convert_endian.py
View File

@ -19,6 +19,11 @@ import gguf
logger = logging.getLogger("gguf-convert-endian") logger = logging.getLogger("gguf-convert-endian")
def byteswap_noop(tensor, block_offs):
# this function is used when byteswapping is not needed
pass
def byteswap_q4_0(tensor, block_offs): def byteswap_q4_0(tensor, block_offs):
# Each block_q4_0 consists of an f16 delta (scaling factor) followed by 16 int8 quantizations. # Each block_q4_0 consists of an f16 delta (scaling factor) followed by 16 int8 quantizations.
@ -55,22 +60,11 @@ def byteswap_q6_k(tensor, block_offs):
byteswap_tensors = { byteswap_tensors = {
gguf.GGMLQuantizationType.Q4_0: { gguf.GGMLQuantizationType.Q4_0: byteswap_q4_0,
"block_size": 18, # 18 bytes = <f16 delta scaling factor> + 16 * <int8 quant> gguf.GGMLQuantizationType.Q8_0: byteswap_q8_0,
"byteswap_func": byteswap_q4_0, gguf.GGMLQuantizationType.Q4_K: byteswap_q4_k,
}, gguf.GGMLQuantizationType.Q6_K: byteswap_q6_k,
gguf.GGMLQuantizationType.Q8_0: { gguf.GGMLQuantizationType.MXFP4: byteswap_noop,
"block_size": 34, # 34 bytes = <f16 delta scaling factor> + 32 * <int8 quant>
"byteswap_func": byteswap_q8_0,
},
gguf.GGMLQuantizationType.Q4_K: {
"block_size": 144, # 144 bytes = 2 * <f16 delta scaling factor> + 140 * <int8 quant>
"byteswap_func": byteswap_q4_k,
},
gguf.GGMLQuantizationType.Q6_K: {
"block_size": 210, # 210 bytes = <f16 delta scaling factor> + 208 * <int8 quant>
"byteswap_func": byteswap_q6_k,
},
} }
@ -135,8 +129,8 @@ def convert_byteorder(reader: gguf.GGUFReader, args: argparse.Namespace) -> None
tensor.data.resize(newshape) tensor.data.resize(newshape)
block_size = byteswap_tensors[tensor.tensor_type]["block_size"] block_size = gguf.constants.GGML_QUANT_SIZES[tensor.tensor_type][1]
byteswap_func = byteswap_tensors[tensor.tensor_type]["byteswap_func"] byteswap_func = byteswap_tensors[tensor.tensor_type]
n_blocks = len(tensor.data) // block_size n_blocks = len(tensor.data) // block_size
for block_num in (inner_pbar := tqdm(range(n_blocks), desc="Byte-swapping Blocks", leave=False)): for block_num in (inner_pbar := tqdm(range(n_blocks), desc="Byte-swapping Blocks", leave=False)):