CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-30 08:42:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

convert : allow sharding reflinked models

Browse Source
This commit is contained in:
Francis Couture-Harpin
2025-09-04 19:08:09 -04:00
parent fb879b40c0
commit cec3449507
1 changed files with 9 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -184,11 +184,15 @@ class GGUFWriter:
if self.use_reflinks: if self.use_reflinks:
# reflinks require alignment to the filesystem blocks # reflinks require alignment to the filesystem blocks
block_size = os.stat(self.path.parent).st_blksize block_size = os.stat(self.path.parent).st_blksize
# necessary to get an appropriate data start offset # necessary to get an appropriate data start offset when padding for reflinks;
# when padding for reflinks; # using the real alignment (8 bytes, from safetensors) would result in a unusable base data offset
# using the real alignment (8 bytes, from safetensors) self.data_alignment = block_size
# would result in a unusable base data offset # for all shards to allow reading them on their own
self.add_custom_alignment(block_size) for i, kv in enumerate(self.kv_data):
# insert at the start of the key-values
if Keys.General.ALIGNMENT in kv:
del kv[Keys.General.ALIGNMENT]
self.kv_data[i] = { Keys.General.ALIGNMENT: GGUFValue(block_size, GGUFValueType.UINT32), **kv }
def print_plan(self) -> list[Path]: def print_plan(self) -> list[Path]:
logger.info("Writing the following files:") logger.info("Writing the following files:")