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

Fix full-finetuning precision fallback for fp32 models

Browse source
This commit is contained in:
Daniel Han 2026-02-25 15:50:11 +00:00
commit 997bde3ead
2 changed files with 29 additions and 13 deletions
Download patch file Download diff file Expand all files Collapse all files

5
unsloth/models/loader.py
View file

@ -1200,8 +1200,9 @@ class FastModel(FastBaseModel):
redirector = contextlib.redirect_stdout(open(os.devnull, "w"))
model_types = ["siglip"] + model_types
# Set forced float32 env flag
os.environ["UNSLOTH_FORCE_FLOAT32"] = "0"
# Preserve explicit user override if set prior to model loading.
user_forced_float32 = os.environ.get("UNSLOTH_FORCE_FLOAT32", "0") == "1"
os.environ["UNSLOTH_FORCE_FLOAT32"] = "1" if user_forced_float32 else "0"
do_forced_float32 = False
for model_type_arch in model_types:
if model_type_arch != "siglip":

37
unsloth/models/rl.py
View file

@ -715,19 +715,22 @@ def _patch_trl_rl_trainers(trainer_file = "grpo_trainer"):
"if type(use_bf16) is not bool: use_bf16 = False\n"
"use_fp16 = getattr(args, 'fp16', False)\n"
"if type(use_fp16) is not bool: use_fp16 = False\n"
"force_float32 = False\n"
"full_finetuning = os.environ.get('UNSLOTH_ENABLE_FULL_FINETUNING', '0') == '1'\n"
"if not full_finetuning and (os.environ.get('UNSLOTH_FORCE_FLOAT32', '0') == '1'):\n"
"force_float32 = os.environ.get('UNSLOTH_FORCE_FLOAT32', '0') == '1'\n"
"if force_float32:\n"
" print('Unsloth: Switching to float32 training since model cannot work with float16')\n"
" force_float32 = True\n"
"mixed_precision_dtype = os.environ.get('UNSLOTH_MIXED_PRECISION', 'float32')\n"
"dtype = getattr(model.config, 'dtype', None) or getattr(model.config, 'torch_dtype', None)\n"
"if dtype is None: dtype = model.get_input_embeddings().weight.dtype\n"
"from unsloth_zoo.utils import _get_dtype\n"
"dtype = _get_dtype(dtype)\n"
"float16 = dtype == torch.float16\n"
"if not force_float32 and (float16 and use_bf16): raise TypeError('Unsloth: Model is in float16 precision but you want to use bfloat16 precision. Set fp16 to `True` and bf16 to `False`')\n"
"if not force_float32 and (not float16 and use_fp16): raise TypeError('Unsloth: Model is in bfloat16 precision but you want to use float16 precision. Set fp16 to `False` and bf16 to `True`')\n"
"is_float16 = dtype == torch.float16\n"
"is_bfloat16 = dtype == torch.bfloat16\n"
"is_float32 = dtype == torch.float32\n"
"if not force_float32 and (is_float16 and use_bf16): raise TypeError('Unsloth: Model is in float16 precision but you want to use bfloat16 precision. Set fp16 to `True` and bf16 to `False`')\n"
"if not force_float32 and (is_bfloat16 and use_fp16): raise TypeError('Unsloth: Model is in bfloat16 precision but you want to use float16 precision. Set fp16 to `False` and bf16 to `True`')\n"
"if not force_float32 and (is_float32 and use_fp16):\n"
" print('Unsloth: Model is float32 but fp16 was requested. Switching to float32 training.')\n"
" force_float32 = True\n"
"if force_float32:\n"
" # Forced float32 training\n"
" args.fp16 = False\n"
@ -737,10 +740,22 @@ def _patch_trl_rl_trainers(trainer_file = "grpo_trainer"):
" # args.mixed_precision is a new argument which needs to be set now\n"
"elif (not use_bf16 and not use_fp16) and mixed_precision_dtype == 'float32':\n"
" # Mixed precision training\n"
" args.fp16 = float16\n"
" args.bf16 = not float16\n"
" os.environ['ACCELERATE_MIXED_PRECISION'] = 'fp16' if float16 else 'bf16'\n"
" if hasattr(args, 'mixed_precision'): args.mixed_precision = 'fp16' if float16 else 'bf16'\n"
" if is_float16:\n"
" args.fp16 = True\n"
" args.bf16 = False\n"
" os.environ['ACCELERATE_MIXED_PRECISION'] = 'fp16'\n"
" if hasattr(args, 'mixed_precision'): args.mixed_precision = 'fp16'\n"
" elif is_bfloat16:\n"
" args.fp16 = False\n"
" args.bf16 = True\n"
" os.environ['ACCELERATE_MIXED_PRECISION'] = 'bf16'\n"
" if hasattr(args, 'mixed_precision'): args.mixed_precision = 'bf16'\n"
" else:\n"
" # Float32 params should not force bf16/fp16 mixed precision.\n"
" args.fp16 = False\n"
" args.bf16 = False\n"
" os.environ['ACCELERATE_MIXED_PRECISION'] = 'no'\n"
" if hasattr(args, 'mixed_precision'): args.mixed_precision = 'no'\n"
" # args.mixed_precision is a new argument which needs to be set now\n"
"elif mixed_precision_dtype == 'bfloat16':\n"
" # Both False since bfloat16 full finetuning doesn't do any autocasting.\n"