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

fix bfloat16 crash on T4 for FORCE_FLOAT32 models and disable trust_remote_code auto-enable for native t5 models

Browse source
This commit is contained in:
Roland Tannous 2026-04-06 17:28:55 +00:00
commit e8328c11de
2 changed files with 7 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

8
studio/backend/core/training/worker.py
View file

@ -402,9 +402,9 @@ def run_training_process(
# ── 1a. Auto-enable trust_remote_code for unsloth/* transformers 5.x models ──
# Some newer architectures (e.g. NemotronH) have config parsing bugs in
# transformers that require trust_remote_code=True as a workaround.
# Only auto-enable for unsloth/* prefixed models (trusted source).
# Exclude Gemma 4 since it is a native transformers 5.5 model and
# trust_remote_code=True would bypass the compiler (disabling fused CE).
# Only auto-enable for models that genuinely need it (set in YAML defaults).
# Native transformers 5.x models (Qwen3.5, Gemma 4, etc.) do NOT need it
# and enabling it can bypass the compiler (disabling fused CE).
from utils.transformers_version import get_transformers_tier
_lowered = model_name.lower()
@ -412,7 +412,7 @@ def run_training_process(
if (
_tier != "default"
and _lowered.startswith("unsloth/")
and _tier != "550" # Gemma 4 is native t5.5 — trust_remote_code bypasses compiler
and _tier not in ("530", "550") # Native t5 models don't need trust_remote_code
and not config.get("trust_remote_code", False)
):
config["trust_remote_code"] = True

4
unsloth/models/loader.py
View file

@ -1357,7 +1357,9 @@ class FastModel(FastBaseModel):
or disable_name.lower() in model_types_all
) and ((dtype == torch.float16) or not SUPPORTS_BFLOAT16):
os.environ["UNSLOTH_FORCE_FLOAT32"] = "1"
dtype = torch.bfloat16 # Change to bfloat16 loading
# Use bfloat16 storage where supported; fall back to float32 on
# older GPUs (e.g. T4) that lack native bfloat16 support.
dtype = torch.bfloat16 if SUPPORTS_BFLOAT16 else torch.float32
break
# Apply gradient checkpointing with smart heuristics
use_gradient_checkpointing = apply_unsloth_gradient_checkpointing(