convert : workaround conversion for p-e-w/gpt-oss-20b-heretic-ara-v4 (#31860)

* fix(conversion-base): changed tokenizer init only for gpt-oss checkpoints, no content modification

* fix(conversion-gpt_oss): normalized 4 tensor suffixes, no content modification

* fix: whitespaces

* revert fix(conversion-base) per #31860

* updated conversion/gpt_oss.py alt->erro reframe

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
This commit is contained in:
LinesandSigbjørn Skjæret authored and GitHub committed 2026-11-10 13:00:03 +01:00
1 parent ac931b5bcb
commit 4ed88da2ea
1 file changed
+10
+10
View File
@@ -85,6 +85,16 @@ class GptOssModel(TextModel):
def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
name, gen = item
# Normalize erroneous GPT-OSS MXFP4 tensor naming.
# https://github.com/p-e-w/heretic/pull/211#issuecomment-5094370390
if "mlp.experts." in name:
for proj in ("down_proj", "gate_up_proj"):
name = name.replace(
f"{proj}_blocks_blocks", f"{proj}_blocks"
).replace(
f"{proj}_blocks_scales", f"{proj}_scales"
)
if "sinks" in name:
name += ".weight"