Skip to content

⚔️ Fix bf16 fp16 config conflict issue #3598

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 13 commits into from
Jun 20, 2025
Merged

⚔️ Fix bf16 fp16 config conflict issue #3598

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
13 commits
Select commit Hold shift + click to select a range
240daff
fix bf16 and fp16 config conflict issue
yao-matrix Jun 16, 2025
c9ac4f3
add more
yao-matrix Jun 16, 2025
b2f5d52
fix
yao-matrix Jun 16, 2025
76abbe7
Merge branch 'main' into bf16-fp16-config
yao-matrix Jun 18, 2025
164a339
Merge branch 'main' into bf16-fp16-config
qgallouedec Jun 19, 2025
d8152f2
Merge branch 'main' into bf16-fp16-config
qgallouedec Jun 20, 2025
1ae429c
more informative
qgallouedec Jun 20, 2025
49a8b31
style
qgallouedec Jun 20, 2025
bcf8a39
style
qgallouedec Jun 20, 2025
7c4b80a
update
qgallouedec Jun 20, 2025
561aa66
Merge branch 'main' into bf16-fp16-config
qgallouedec Jun 20, 2025
a12a9ff
Merge branch 'main' into bf16-fp16-config
qgallouedec Jun 20, 2025
45ecc05
harmonize
qgallouedec Jun 20, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 12 additions & 10 deletions trl/trainer/bco_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -83,19 +83,16 @@ class BCOConfig(TrainingArguments):
logging_steps: float = field(
default=10,
metadata={
"help": (
"Log every X updates steps. Should be an integer or a float in range `[0,1)`. "
"If smaller than 1, will be interpreted as ratio of total training steps."
)
"help": "Log every X updates steps. Should be an integer or a float in range `[0,1)`. If smaller than 1, "
"will be interpreted as ratio of total training steps."
},
)
bf16: bool = field(
default=True,
bf16: Optional[bool] = field(
default=None,
metadata={
"help": (
"Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or using CPU (use_cpu) or Ascend NPU. This is an experimental API and it may change."
)
"help": "Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or Intel XPU or using CPU (use_cpu) or Ascend NPU. If not set, it defaults to `True` if "
"`fp16` is not set."
},
)

Expand Down Expand Up @@ -202,3 +199,8 @@ class BCOConfig(TrainingArguments):
default=10.0,
metadata={"help": "Maximum value of the density ratio. The estimated density ratio is clamped to this value."},
)

def __post_init__(self):
self.bf16 = not (self.fp16) if self.bf16 is None else self.bf16

super().__post_init__()
22 changes: 12 additions & 10 deletions trl/trainer/cpo_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -90,19 +90,16 @@ class CPOConfig(TrainingArguments):
logging_steps: float = field(
default=10,
metadata={
"help": (
"Log every X updates steps. Should be an integer or a float in range `[0,1)`. "
"If smaller than 1, will be interpreted as ratio of total training steps."
)
"help": "Log every X updates steps. Should be an integer or a float in range `[0,1)`. If smaller than 1, "
"will be interpreted as ratio of total training steps."
},
)
bf16: bool = field(
default=True,
bf16: Optional[bool] = field(
default=None,
metadata={
"help": (
"Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or using CPU (use_cpu) or Ascend NPU. This is an experimental API and it may change."
)
"help": "Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or Intel XPU or using CPU (use_cpu) or Ascend NPU. If not set, it defaults to `True` if "
"`fp16` is not set."
},
)

Expand Down Expand Up @@ -188,3 +185,8 @@ class CPOConfig(TrainingArguments):
default=None,
metadata={"help": "Number of processes to use for processing the dataset."},
)

def __post_init__(self):
self.bf16 = not (self.fp16) if self.bf16 is None else self.bf16

super().__post_init__()
22 changes: 12 additions & 10 deletions trl/trainer/dpo_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -189,19 +189,16 @@ class DPOConfig(TrainingArguments):
logging_steps: float = field(
default=10,
metadata={
"help": (
"Log every X updates steps. Should be an integer or a float in range `[0,1)`. "
"If smaller than 1, will be interpreted as ratio of total training steps."
)
"help": "Log every X updates steps. Should be an integer or a float in range `[0,1)`. If smaller than 1, "
"will be interpreted as ratio of total training steps."
},
)
bf16: bool = field(
default=True,
bf16: Optional[bool] = field(
default=None,
metadata={
"help": (
"Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or using CPU (use_cpu) or Ascend NPU. This is an experimental API and it may change."
)
"help": "Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or Intel XPU or using CPU (use_cpu) or Ascend NPU. If not set, it defaults to `True` if "
"`fp16` is not set."
},
)

Expand Down Expand Up @@ -439,3 +436,8 @@ class DPOConfig(TrainingArguments):
"Comet during evaluation."
},
)

def __post_init__(self):
self.bf16 = not (self.fp16) if self.bf16 is None else self.bf16

super().__post_init__()
19 changes: 9 additions & 10 deletions trl/trainer/grpo_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -212,19 +212,16 @@ class GRPOConfig(TrainingArguments):
logging_steps: float = field(
default=10,
metadata={
"help": (
"Log every X updates steps. Should be an integer or a float in range `[0,1)`. "
"If smaller than 1, will be interpreted as ratio of total training steps."
)
"help": "Log every X updates steps. Should be an integer or a float in range `[0,1)`. If smaller than 1, "
"will be interpreted as ratio of total training steps."
},
)
bf16: bool = field(
default=True,
bf16: Optional[bool] = field(
default=None,
metadata={
"help": (
"Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or using CPU (use_cpu) or Ascend NPU. This is an experimental API and it may change."
)
"help": "Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or Intel XPU or using CPU (use_cpu) or Ascend NPU. If not set, it defaults to `True` if "
"`fp16` is not set."
},
)

Expand Down Expand Up @@ -530,6 +527,8 @@ class GRPOConfig(TrainingArguments):
)

def __post_init__(self):
self.bf16 = not (self.fp16) if self.bf16 is None else self.bf16

super().__post_init__()

num_processes = self.world_size
Expand Down
19 changes: 9 additions & 10 deletions trl/trainer/iterative_sft_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,19 +54,16 @@ class may differ from those in [`~transformers.TrainingArguments`].
logging_steps: float = field(
default=10,
metadata={
"help": (
"Log every X updates steps. Should be an integer or a float in range `[0,1)`. "
"If smaller than 1, will be interpreted as ratio of total training steps."
)
"help": "Log every X updates steps. Should be an integer or a float in range `[0,1)`. If smaller than 1, "
"will be interpreted as ratio of total training steps."
},
)
bf16: bool = field(
default=True,
bf16: Optional[bool] = field(
default=None,
metadata={
"help": (
"Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or using CPU (use_cpu) or Ascend NPU. This is an experimental API and it may change."
)
"help": "Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or Intel XPU or using CPU (use_cpu) or Ascend NPU. If not set, it defaults to `True` if "
"`fp16` is not set."
},
)

Expand Down Expand Up @@ -96,6 +93,8 @@ class may differ from those in [`~transformers.TrainingArguments`].
)

def __post_init__(self):
self.bf16 = not (self.fp16) if self.bf16 is None else self.bf16

super().__post_init__()

if self.truncation_mode not in ["keep_end", "keep_start"]:
Expand Down
22 changes: 12 additions & 10 deletions trl/trainer/kto_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -97,19 +97,16 @@ class KTOConfig(TrainingArguments):
logging_steps: float = field(
default=10,
metadata={
"help": (
"Log every X updates steps. Should be an integer or a float in range `[0,1)`. "
"If smaller than 1, will be interpreted as ratio of total training steps."
)
"help": "Log every X updates steps. Should be an integer or a float in range `[0,1)`. If smaller than 1, "
"will be interpreted as ratio of total training steps."
},
)
bf16: bool = field(
default=True,
bf16: Optional[bool] = field(
default=None,
metadata={
"help": (
"Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or using CPU (use_cpu) or Ascend NPU. This is an experimental API and it may change."
)
"help": "Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or Intel XPU or using CPU (use_cpu) or Ascend NPU. If not set, it defaults to `True` if "
"`fp16` is not set."
},
)

Expand Down Expand Up @@ -231,3 +228,8 @@ class KTOConfig(TrainingArguments):
"`use_liger_loss` is `True`."
},
)

def __post_init__(self):
self.bf16 = not (self.fp16) if self.bf16 is None else self.bf16

super().__post_init__()
20 changes: 10 additions & 10 deletions trl/trainer/online_dpo_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -81,19 +81,16 @@ class may differ from those in [`~transformers.TrainingArguments`].
logging_steps: float = field(
default=10,
metadata={
"help": (
"Log every X updates steps. Should be an integer or a float in range `[0,1)`. "
"If smaller than 1, will be interpreted as ratio of total training steps."
)
"help": "Log every X updates steps. Should be an integer or a float in range `[0,1)`. If smaller than 1, "
"will be interpreted as ratio of total training steps."
},
)
bf16: bool = field(
default=True,
bf16: Optional[bool] = field(
default=None,
metadata={
"help": (
"Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or using CPU (use_cpu) or Ascend NPU. This is an experimental API and it may change."
)
"help": "Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or Intel XPU or using CPU (use_cpu) or Ascend NPU. If not set, it defaults to `True` if "
"`fp16` is not set."
},
)

Expand Down Expand Up @@ -180,6 +177,9 @@ class may differ from those in [`~transformers.TrainingArguments`].
)

def __post_init__(self):
self.bf16 = not (self.fp16) if self.bf16 is None else self.bf16

super().__post_init__()

if hasattr(self.beta, "__len__") and len(self.beta) == 1:
self.beta = self.beta[0]
22 changes: 12 additions & 10 deletions trl/trainer/orpo_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -75,19 +75,16 @@ class ORPOConfig(TrainingArguments):
logging_steps: float = field(
default=10,
metadata={
"help": (
"Log every X updates steps. Should be an integer or a float in range `[0,1)`. "
"If smaller than 1, will be interpreted as ratio of total training steps."
)
"help": "Log every X updates steps. Should be an integer or a float in range `[0,1)`. If smaller than 1, "
"will be interpreted as ratio of total training steps."
},
)
bf16: bool = field(
default=True,
bf16: Optional[bool] = field(
default=None,
metadata={
"help": (
"Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or using CPU (use_cpu) or Ascend NPU. This is an experimental API and it may change."
)
"help": "Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or Intel XPU or using CPU (use_cpu) or Ascend NPU. If not set, it defaults to `True` if "
"`fp16` is not set."
},
)

Expand Down Expand Up @@ -159,3 +156,8 @@ class ORPOConfig(TrainingArguments):
default=None,
metadata={"help": "Number of processes to use for processing the dataset."},
)

def __post_init__(self):
self.bf16 = not (self.fp16) if self.bf16 is None else self.bf16

super().__post_init__()
22 changes: 12 additions & 10 deletions trl/trainer/prm_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -56,19 +56,16 @@ class PRMConfig(TrainingArguments):
logging_steps: float = field(
default=10,
metadata={
"help": (
"Log every X updates steps. Should be an integer or a float in range `[0,1)`. "
"If smaller than 1, will be interpreted as ratio of total training steps."
)
"help": "Log every X updates steps. Should be an integer or a float in range `[0,1)`. If smaller than 1, "
"will be interpreted as ratio of total training steps."
},
)
bf16: bool = field(
default=True,
bf16: Optional[bool] = field(
default=None,
metadata={
"help": (
"Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or using CPU (use_cpu) or Ascend NPU. This is an experimental API and it may change."
)
"help": "Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or Intel XPU or using CPU (use_cpu) or Ascend NPU. If not set, it defaults to `True` if "
"`fp16` is not set."
},
)
average_tokens_across_devices: bool = field(
Expand Down Expand Up @@ -110,3 +107,8 @@ class PRMConfig(TrainingArguments):
default=None,
metadata={"help": "Number of processes to use for processing the dataset."},
)

def __post_init__(self):
self.bf16 = not (self.fp16) if self.bf16 is None else self.bf16

super().__post_init__()
22 changes: 12 additions & 10 deletions trl/trainer/reward_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -51,19 +51,16 @@ class may differ from those in [`~transformers.TrainingArguments`].
logging_steps: float = field(
default=10,
metadata={
"help": (
"Log every X updates steps. Should be an integer or a float in range `[0,1)`. "
"If smaller than 1, will be interpreted as ratio of total training steps."
)
"help": "Log every X updates steps. Should be an integer or a float in range `[0,1)`. If smaller than 1, "
"will be interpreted as ratio of total training steps."
},
)
bf16: bool = field(
default=True,
bf16: Optional[bool] = field(
default=None,
metadata={
"help": (
"Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or using CPU (use_cpu) or Ascend NPU. This is an experimental API and it may change."
)
"help": "Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or Intel XPU or using CPU (use_cpu) or Ascend NPU. If not set, it defaults to `True` if "
"`fp16` is not set."
},
)
average_tokens_across_devices: bool = field(
Expand Down Expand Up @@ -103,3 +100,8 @@ class may differ from those in [`~transformers.TrainingArguments`].
"if the dataset is pretokenized."
},
)

def __post_init__(self):
self.bf16 = not (self.fp16) if self.bf16 is None else self.bf16

super().__post_init__()
19 changes: 9 additions & 10 deletions trl/trainer/sft_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -104,19 +104,16 @@ class SFTConfig(TrainingArguments):
logging_steps: float = field(
default=10,
metadata={
"help": (
"Log every X updates steps. Should be an integer or a float in range `[0,1)`. "
"If smaller than 1, will be interpreted as ratio of total training steps."
)
"help": "Log every X updates steps. Should be an integer or a float in range `[0,1)`. If smaller than 1, "
"will be interpreted as ratio of total training steps."
},
)
bf16: bool = field(
default=True,
bf16: Optional[bool] = field(
default=None,
metadata={
"help": (
"Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or using CPU (use_cpu) or Ascend NPU. This is an experimental API and it may change."
)
"help": "Whether to use bf16 (mixed) precision instead of 32-bit. Requires Ampere or higher NVIDIA "
"architecture or Intel XPU or using CPU (use_cpu) or Ascend NPU. If not set, it defaults to `True` if "
"`fp16` is not set."
},
)
average_tokens_across_devices: bool = field(
Expand Down Expand Up @@ -252,6 +249,8 @@ class SFTConfig(TrainingArguments):
)

def __post_init__(self):
self.bf16 = not (self.fp16) if self.bf16 is None else self.bf16

super().__post_init__()

if self.max_seq_length is not None:
Expand Down
Loading
Loading