Skip to content

Gemma3n fix #338

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Oct 27, 2025
Merged

Gemma3n fix #338

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
move PYTORCH_CUDA_ALLOC_CONF into zoo
  • Loading branch information
@mmathew23
mmathew23 committed Oct 23, 2025
commit 1efecc4936e61579c9b309f24cd586e3de3e716a
51 changes: 51 additions & 0 deletions unsloth_zoo/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,6 +18,8 @@

import os
import warnings
import re
import torch
# Hugging Face Hub faster downloads
if "HF_HUB_ENABLE_HF_TRANSFER" not in os.environ:
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
Expand Down Expand Up @@ -80,6 +82,55 @@
ALLOW_PREQUANTIZED_MODELS,
)

# Reduce VRAM usage by reducing fragmentation
# And optimize pinning of memory
# TODO(billishyahao): need to add hip related optimization...
if (DEVICE_TYPE in ("cuda", "hip")) and (os.environ.get("UNSLOTH_VLLM_STANDBY", "0")=="0"):
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = \
"expandable_segments:True,"\
"roundup_power2_divisions:[32:256,64:128,256:64,>:32]"
os.environ["PYTORCH_HIP_ALLOC_CONF"] = "expandable_segments:True"
elif (DEVICE_TYPE in ("cuda", "hip")) and (os.environ.get("UNSLOTH_VLLM_STANDBY", "0")=="1") and \
("expandable_segments:True" in os.environ.get("PYTORCH_CUDA_ALLOC_CONF", "")):
warnings.warn(
"Unsloth: `UNSLOTH_VLLM_STANDBY` is on, but requires `expandable_segments` to be off.\n"\
"We will remove `expandable_segments`.",
stacklevel = 2,
)
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = re.sub(
r"expandable\_segments\:True\,?",
"",
os.environ["PYTORCH_CUDA_ALLOC_CONF"],
)
os.environ["PYTORCH_HIP_ALLOC_CONF"] = re.sub(
r"expandable\_segments\:True\,?",
"",
os.environ["PYTORCH_HIP_ALLOC_CONF"],
)
pass
# We support Pytorch 2
# Fixes https://github.com/unslothai/unsloth/issues/38
torch_version = str(re.match(r"[0-9\.]{3,}", str(torch.__version__)).group(0)).split(".")
major_torch, minor_torch = torch_version[0], torch_version[1]
major_torch, minor_torch = int(major_torch), int(minor_torch)
if (major_torch < 2):
raise ImportError("Unsloth only supports Pytorch 2 for now. Please update your Pytorch to 2.1.\n"\
"We have some installation instructions on our Github page.")
elif (major_torch == 2) and (minor_torch < 2):
# Disable expandable_segments
del os.environ["PYTORCH_CUDA_ALLOC_CONF"]
del os.environ["PYTORCH_HIP_ALLOC_CONF"]
pass

# CCE fails on Torch 2.8 and above
# OutOfResources: out of resource: shared memory, Required: 98304, Hardware limit: 65536. Reducing block sizes or `num_stages`
if (major_torch >= 2 and minor_torch >= 8) or (major_torch > 2):
os.environ["UNSLOTH_ENABLE_CCE"] = "0"
elif DEVICE_TYPE == "hip":
# CCE also fails in HIP / AMD
os.environ["UNSLOTH_ENABLE_CCE"] = "0"


if not ("UNSLOTH_IS_PRESENT" in os.environ):
raise ImportError("Please install Unsloth via `pip install unsloth`!")
pass
Expand Down