CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-04 09:32:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

convert : better logging of partially reflinkable tensors

Browse Source
This commit is contained in:
Francis Couture-Harpin
2025-09-09 11:00:00 -04:00
parent 2499e47cfd
commit 34f37c283b
2 changed files with 12 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
gguf-py/gguf/gguf_writer.py
View File

@@ -201,7 +201,7 @@ class GGUFWriter:
assert len(filenames) == len(self.tensors) assert len(filenames) == len(self.tensors)
for name, tensors in zip(filenames, self.tensors): for name, tensors in zip(filenames, self.tensors):
total_size = sum(ti.nbytes for ti in tensors.values()) total_size = sum(ti.nbytes for ti in tensors.values())
reflinkable_size = count_reflinkable_size(ti.tensor for ti in tensors.values()) if self.use_reflinks else 0 reflinkable_size = count_reflinkable_size((name, ti.tensor) for name, ti in tensors.items()) if self.use_reflinks else 0
logger.info(f"{name}: n_tensors = {len(tensors)}, total_size = {GGUFWriter.format_n_bytes_to_str(total_size)}{', reflinked = ' + GGUFWriter.format_n_bytes_to_str(total_size - reflinkable_size) if self.use_reflinks else ''}") logger.info(f"{name}: n_tensors = {len(tensors)}, total_size = {GGUFWriter.format_n_bytes_to_str(total_size)}{', reflinked = ' + GGUFWriter.format_n_bytes_to_str(total_size - reflinkable_size) if self.use_reflinks else ''}")
if self.dry_run: if self.dry_run:

17
gguf-py/gguf/lazy.py
View File

@@ -278,17 +278,22 @@ def best_extra_offset(t: np.ndarray | LazyNumpyTensor | None, current_offset: in
return best_offset return best_offset
def count_reflinkable_size(tensors: Iterable[np.ndarray | LazyNumpyTensor | None]) -> int: def count_reflinkable_size(tensors: Iterable[tuple[str, np.ndarray | LazyNumpyTensor | None]]) -> int:
if not hasattr(os, "copy_file_range"): if not hasattr(os, "copy_file_range"):
return 0 return 0
size = 0 size = 0
for t in tensors: for name, t in tensors:
if isinstance(t, LazyNumpyTensor) and len(t._ranges) > 0: if isinstance(t, LazyNumpyTensor) and len(t._ranges) > 0:
align_offset = best_extra_offset(t, 0) align_offset = best_extra_offset(t, 0)
misaligned = 0
for range in t._ranges: for range in t._ranges:
if range.block_size > 0 and range.offset % range.block_size == align_offset: if range.block_size > 0:
if range.offset % range.block_size == align_offset:
size += range.size size += range.size
else:
misaligned += 1
if misaligned > 0:
logger.debug(f"{name} misaligned for reflinking, fallback to copy for {misaligned} of {len(t._ranges)} parts")
return size return size
@@ -317,7 +322,7 @@ def copy_tensor_ranges(t: LazyNumpyTensor, fout: BufferedWriter):
has_copy_file_range = hasattr(os, "copy_file_range") has_copy_file_range = hasattr(os, "copy_file_range")
for i, r in enumerate(ranges): for r in ranges:
src = src_files[r.filename] src = src_files[r.filename]
if has_copy_file_range: if has_copy_file_range:
if r.block_size > 0 and (r.offset % r.block_size) == (start_offset % r.block_size): if r.block_size > 0 and (r.offset % r.block_size) == (start_offset % r.block_size):
@@ -354,8 +359,6 @@ def copy_tensor_ranges(t: LazyNumpyTensor, fout: BufferedWriter):
os.copy_file_range(src.fileno(), fout.fileno(), size, offset_src, dst_offset) os.copy_file_range(src.fileno(), fout.fileno(), size, offset_src, dst_offset)
dst_offset += r.size - extra_size dst_offset += r.size - extra_size
else: else:
if r.block_size > 0:
logger.debug(f"misaligned for reflinking, falling back to copy ({i}/{len(ranges)})")
# not trying to use reflinks, but still using os.copy_file_range for speed # not trying to use reflinks, but still using os.copy_file_range for speed
os.copy_file_range(src.fileno(), fout.fileno(), r.size, r.offset, dst_offset) os.copy_file_range(src.fileno(), fout.fileno(), r.size, r.offset, dst_offset)
dst_offset += r.size dst_offset += r.size