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

Fixes

Browse source
This commit is contained in:
Daniel Han 2025-10-26 22:39:38 -07:00
commit c51468c40c
3 changed files with 7 additions and 9 deletions
Download patch file Download diff file Expand all files Collapse all files

4
unsloth/import_fixes.py
View file

@ -150,8 +150,8 @@ def patch_ipykernel_hf_xet():
Version(importlib_version("ipykernel")) == Version("7.0.0")
):
print(
"#### Unsloth: `hf_xet==1.1.10` and `ipykernel>6.30.1` breaks progress bars. Disabling for now in XET.\n"\
"#### Unsloth: To re-enable progress bars, please downgrade to `ipykernel==6.30.1` or wait for a fix to\n"\
"#### Unsloth: `hf_xet==1.1.10` and `ipykernel==7.0.0` breaks progress bars. Disabling for now in XET.\n"\
"#### Unsloth: To re-enable progress bars, please upgrade to `ipykernel>7.0.0` or wait for a fix to\n"\
"https://github.com/huggingface/xet-core/issues/526"
)
from huggingface_hub.utils import disable_progress_bars

8
unsloth/models/llama.py
View file

@ -1203,12 +1203,8 @@ def CausalLM_fast_forward(fast_forward_inference):
else:
RETURN_LOGITS = os.environ.get("UNSLOTH_RETURN_LOGITS", "0") == "1"
# < 1024 Normal Unsloth uses less VRAM!
if DEVICE_TYPE == "hip":
# [TODO] AMD GPUs fail on chunked_cross_entropy loss!
# RuntimeError: Triton Error [HIP]: Code: 1, Messsage: invalid argument
RETURN_LOGITS = False
elif bsz*q_len <= 1024:
# Uses 800MB more VRAM it seems than fused CE Loss
if bsz * q_len <= 1024 and not RETURN_LOGITS:
# Use unsloth_fused_ce_loss which actually calculates the best chunk size to reduce VRAM usage
RETURN_LOGITS = False
if not RETURN_LOGITS and labels is not None:

4
unsloth/models/mistral.py
View file

@ -298,7 +298,9 @@ def MistralForCausalLM_fast_forward(
else:
RETURN_LOGITS = os.environ.get("UNSLOTH_RETURN_LOGITS", "0") == "1"
# < 1024 Normal Unsloth uses less VRAM!
if bsz * q_len <= 1024: RETURN_LOGITS = True
if bsz * q_len <= 1024 and not RETURN_LOGITS:
# Use unsloth_fused_ce_loss which actually calculates the best chunk size to reduce VRAM usage
RETURN_LOGITS = False
if not RETURN_LOGITS and labels is not None:
n_items = kwargs.get("num_items_in_batch", None) or kwargs.get("n_items", None)