Skip to content

[Feature] Update loss_scale method call to pass through inputs.extra_kwargs #6156

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

Closed
wants to merge 3 commits into from
+247 −107
Closed

[Feature] Update loss_scale method call to pass through inputs.extra_kwargs #6156

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
9e2d004
feat:Update loss_scale method call to pass through inputs.extra_kwargs
CJack812 Oct 15, 2025
63af43a
style update
CJack812 Oct 15, 2025
fadd51a
style update
CJack812 Oct 15, 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
8 changes: 4 additions & 4 deletions examples/train/rft/rft.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,7 +22,7 @@ def do_sample(model: str, model_type: str, dataset: List[str], iter: int):
for device in range(device_count):
sample_cmd = (f'{conda_prefix} USE_OPENCOMPASS_EVALUATOR=True CUDA_VISIBLE_DEVICES={device} swift sample '
f'--model {model} --model_type {model_type} '
f'--dataset {" ".join(dataset)} '
f'--dataset {' '.join(dataset)} '
f'--data_range {device} {device_count} '
f'--max_length 2048 '
f'--system "You are a math model, you should **think step by step** carefully, '
Expand Down Expand Up @@ -61,7 +61,7 @@ def do_sample(model: str, model_type: str, dataset: List[str], iter: int):
sample_cmd = (
f'{conda_prefix} USE_OPENCOMPASS_EVALUATOR=True CUDA_VISIBLE_DEVICES={device} swift sample '
f'--model {model} --model_type {model_type} ' # change to --resume_from_checkpoint to use the latest optimizer state # noqa
f'--dataset {" ".join(dataset)} '
f'--dataset {' '.join(dataset)} '
f'--data_range {device} {device_count} '
f'--max_length 2048 '
f'--system "You are a math model, you should **think step by step** carefully, '
Expand Down Expand Up @@ -91,7 +91,7 @@ def do_sample(model: str, model_type: str, dataset: List[str], iter: int):
for proc, handler in enumerate(handlers):
handler.wait()
assert os.path.exists(os.path.join('sample_output', f'iter_{iter}_proc_{proc}_sampling.jsonl')), (
f'{os.path.join("sample_output", f"iter_{iter}_proc_{proc}_sampling.jsonl")} not exists, '
f'{os.path.join('sample_output', f"iter_{iter}_proc_{proc}_sampling.jsonl")} not exists, '
'please check the sample logs to get the detail error.')
datasets.append(os.path.join('sample_output', f'iter_{iter}_proc_{proc}_sampling.jsonl'))
print(f'Sampling done, files:{datasets}', flush=True)
Expand All @@ -110,7 +110,7 @@ def do_train(model: str, model_type: str, datasets: List[str], iter, cmd='sft'):
ga = 128 // get_device_count() // 2
train_cmd = (f'{conda_prefix} {gpu_prefix} swift {cmd} '
f'--model {model} --model_type {model_type} '
f'--dataset {" ".join(datasets)} '
f'--dataset {' '.join(datasets)} '
f'--max_length 2048 '
f'--num_train_epochs 1 '
f'--load_args false '
Expand Down
10 changes: 5 additions & 5 deletions scripts/benchmark/exp_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -122,7 +122,7 @@ def run(self, exp: Experiment):
exp.runtime = runtime
envs = deepcopy(runtime.get('env', {}))
envs.update(os.environ)
logger.info(f'Running cmd: {runtime["running_cmd"]}, env: {runtime.get("env", {})}')
logger.info(f'Running cmd: {runtime['running_cmd']}, env: {runtime.get('env', {})}')
os.makedirs('exp', exist_ok=True)
log_file = os.path.join('exp', f'{exp.name}.eval.log')
exp.handler = subprocess.Popen(runtime['running_cmd'] + f' > {log_file} 2>&1', env=envs, shell=True)
Expand All @@ -140,7 +140,7 @@ def run(self, exp: Experiment):
exp.runtime = runtime
envs = deepcopy(runtime.get('env', {}))
envs.update(os.environ)
logger.info(f'Running cmd: {runtime["running_cmd"]}, env: {runtime.get("env", {})}')
logger.info(f'Running cmd: {runtime['running_cmd']}, env: {runtime.get('env', {})}')
os.makedirs('exp', exist_ok=True)
log_file = os.path.join('exp', f'{exp.name}.{exp.cmd}.log')
exp.handler = subprocess.Popen(runtime['running_cmd'] + f' > {log_file} 2>&1', env=envs, shell=True)
Expand All @@ -162,10 +162,10 @@ def _build_eval_cmd(self, exp: Experiment):
if best_model_checkpoint is not None:
if not os.path.exists(os.path.join(best_model_checkpoint, 'args.json')):
cmd = f'swift eval --ckpt_dir {best_model_checkpoint} ' \
+ f'--infer_backend pt --train_type full --eval_dataset {" ".join(eval_dataset)}'
+ f'--infer_backend pt --train_type full --eval_dataset {' '.join(eval_dataset)}'
else:
cmd = f'swift eval --model {exp.args.get("model")} --infer_backend pt ' \
f'--eval_dataset {" ".join(eval_dataset)}'
cmd = f'swift eval --model {exp.args.get('model')} --infer_backend pt ' \
f'--eval_dataset {' '.join(eval_dataset)}'

return {
'running_cmd': cmd,
Expand Down
44 changes: 22 additions & 22 deletions scripts/benchmark/generate_report.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,23 +69,23 @@ def tuner_hyper_params(self):
return ''
if args['sft_type'] in ('lora', 'adalora', 'longlora'):
if 'lora_rank' in args:
hyper_params += f'rank={args["lora_rank"]}/' \
f'target={args["lora_target_modules"]}/' \
f'alpha={args["lora_alpha"]}/' \
f'lr_ratio={args.get("lora_lr_ratio", None)}/' \
f'use_rslora={args.get("use_rslora", False)}/' \
f'use_dora={args.get("use_dora", False)}'
hyper_params += f'rank={args['lora_rank']}/' \
f'target={args['lora_target_modules']}/' \
f'alpha={args['lora_alpha']}/' \
f'lr_ratio={args.get('lora_lr_ratio', None)}/' \
f'use_rslora={args.get('use_rslora', False)}/' \
f'use_dora={args.get('use_dora', False)}'
else:
hyper_params = ''
if args['sft_type'] == 'full':
if 'use_galore' in args and args['use_galore'] == 'true':
hyper_params += f'galore_rank={args["galore_rank"]}/' \
f'galore_per_parameter={args["galore_optim_per_parameter"]}/' \
f'galore_with_embedding={args["galore_with_embedding"]}/'
hyper_params += f'galore_rank={args['galore_rank']}/' \
f'galore_per_parameter={args['galore_optim_per_parameter']}/' \
f'galore_with_embedding={args['galore_with_embedding']}/'
if args['sft_type'] == 'llamapro':
hyper_params += f'num_blocks={args["llamapro_num_new_blocks"]}/'
hyper_params += f'num_blocks={args['llamapro_num_new_blocks']}/'
if 'neftune_noise_alpha' in args and args['neftune_noise_alpha']:
hyper_params += f'neftune_noise_alpha={args["neftune_noise_alpha"]}/'
hyper_params += f'neftune_noise_alpha={args['neftune_noise_alpha']}/'

if hyper_params.endswith('/'):
hyper_params = hyper_params[:-1]
Expand All @@ -95,8 +95,8 @@ def tuner_hyper_params(self):
def hyper_parameters(self):
if 'learning_rate' not in self.args:
return ''
return f'lr={self.args["learning_rate"]}/' \
f'epoch={self.args["num_train_epochs"]}'
return f'lr={self.args['learning_rate']}/' \
f'epoch={self.args['num_train_epochs']}'

@property
def train_speed(self):
Expand Down Expand Up @@ -190,10 +190,10 @@ def generate_sft_report(outputs: List[ModelOutput]):
ceval_acc = '' if not ceval_acc else f'**{ceval_acc:.3f}**'

line = f'|{output.name}|' \
f'{output.args["model_type"]}|' \
f'{output.args.get("dataset")}|' \
f'{output.args.get("train_dataset_mix_ratio", 0.)}|' \
f'{output.args.get("sft_type")}|' \
f'{output.args['model_type']}|' \
f'{output.args.get('dataset')}|' \
f'{output.args.get('train_dataset_mix_ratio', 0.)}|' \
f'{output.args.get('sft_type')}|' \
f'{output.tuner_hyper_params}|' \
f'{output.num_trainable_parameters}({output.trainable_parameters_percentage})|' \
f'{use_flash_attn}|' \
Expand Down Expand Up @@ -267,14 +267,14 @@ def generate_export_report(outputs: List[ModelOutput]):
ceval_acc = '' if not ceval_acc else f'**{ceval_acc:.3f}**'

if output.train_dataset_info:
dataset_info = f'{output.args["dataset"]}/{output.train_dataset_info}'
dataset_info = f'{output.args['dataset']}/{output.train_dataset_info}'
else:
dataset_info = f'{output.args["dataset"]}'
dataset_info = f'{output.args['dataset']}'
line = f'|{output.name}|' \
f'{output.args["model_type"]}|' \
f'{output.args['model_type']}|' \
f'{dataset_info}|' \
f'{output.args["quant_method"]}|' \
f'{output.args["quant_bits"]}|' \
f'{output.args['quant_method']}|' \
f'{output.args['quant_bits']}|' \
f'{infer_speed}|' \
f'{gsm8k_acc}|' \
f'{arc_acc}|' \
Expand Down
Loading
Loading