ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Drop redundant transformer_quant re-normalization in the dense-quant gate

Browse source
This commit is contained in:
oobabooga 2026-07-06 10:08:19 -03:00
commit 35dccf2da6
1 changed files with 4 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

8
studio/backend/core/inference/diffusion.py
View file

@ -891,8 +891,8 @@ class DiffusionBackend:
@staticmethod
def _safetensors_param_count(path: Path) -> int:
"""Total tensor elements in a safetensors file, read from its JSON header (an
8-byte little-endian length prefix then the header) without touching tensor data."""
"""Total tensor elements in a safetensors file, read from its JSON header without
touching the tensor data. 0 on any read/parse failure."""
try:
with open(path, "rb") as fh:
header_len = int.from_bytes(fh.read(8), "little")
@ -1062,7 +1062,7 @@ class DiffusionBackend:
if dense_quant_requested and dense_fits:
scheme = select_transformer_quant_scheme(
target,
normalize_transformer_quant(transformer_quant),
transformer_quant, # normalized above
family = getattr(fam, "name", None),
)
prequant = (
@ -1135,7 +1135,7 @@ class DiffusionBackend:
logger.warning(
"diffusion.transformer_quant: %s requested but the dense fast path needs "
"a resident load that doesn't fit here; loading GGUF without dense quant",
normalize_transformer_quant(transformer_quant),
transformer_quant, # normalized above
)
if pipe is None: