Skip to content

Feat/add tp pp fc #2768

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
zjjlivein merged 2 commits into from
Oct 21, 2025
Merged

Feat/add tp pp fc #2768

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
b87026d
Add model test tp pp fc case
huanghengheng Oct 21, 2025
e0eca2a
black the py file
huanghengheng Oct 21, 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
88 changes: 86 additions & 2 deletions scripts/regression/test_dpo.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -129,7 +129,7 @@ def test_dpo_full(self):
self.dpotrain_tester.assert_result(training_p.returncode, training_p.stdout)

# test training loss
EXCEPTED_LOSS = 0.474259
EXCEPTED_LOSS = 0.474242
self.dpotrain_tester.assert_loss(training_p.stdout, EXCEPTED_LOSS)

# test model generate
Expand Down Expand Up @@ -163,7 +163,7 @@ def test_dpo_lora(self):
self.dpotrain_tester.assert_result(training_p.returncode, training_p.stdout)

# test training loss
EXCEPTED_LOSS = 0.474163
EXCEPTED_LOSS = 0.474235
self.dpotrain_tester.assert_loss(training_p.stdout, EXCEPTED_LOSS)

# test lora merge
Expand All @@ -189,3 +189,87 @@ def test_dpo_lora(self):
[[22407, 120525, 77505, 113631, 47887, 134141, 122487, 61092, 40897, 40601]]
)
self.dpotrain_tester.create_and_check_model_generate(lora_merge_output_dir, EXPECTED_RESULT)

def test_dpo_full_tp_pp(self):
output_dir = os.path.join(OUTPUT_DIR, "dpo_full_tp_pp")
update_args = {
"model_name_or_path": MODEL_NAME_OR_PATH,
"train_dataset_path": "./tests/fixtures/dummy/ernie/dpo-train.jsonl",
"eval_dataset_path": "./tests/fixtures/dummy/ernie/dpo-train.jsonl",
"output_dir": output_dir,
}
config_path = os.path.join(CONFIG_PATH, "full_tp_pp.yaml")
updated_config_path = self.dpotrain_tester.update_training_args(config_path, output_dir, update_args)
train_path = os.path.join(TRAIN_PATH, "alignment/dpo/run_dpo.py")
cmd = [
"python",
"-u",
"-m",
"paddle.distributed.launch",
"--devices",
"0,1,2,3",
train_path,
updated_config_path,
]
training_p = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)

# test training result
self.dpotrain_tester.assert_result(training_p.returncode, training_p.stdout)

# test training loss
EXCEPTED_LOSS = 0.495105
self.dpotrain_tester.assert_loss(training_p.stdout, EXCEPTED_LOSS)

# test model generate
EXPECTED_RESULT = paddle.to_tensor([[22407, 90612, 90612, 90612, 90612, 90612, 90612, 90612, 90612, 90612]])
self.dpotrain_tester.create_and_check_model_generate(output_dir, EXPECTED_RESULT)

def test_dpo_lora_tp_pp(self):
output_dir = os.path.join(OUTPUT_DIR, "dpo_lora_tp_pp")
update_args = {
"model_name_or_path": MODEL_NAME_OR_PATH,
"train_dataset_path": "./tests/fixtures/dummy/ernie/dpo-train.jsonl",
"eval_dataset_path": "./tests/fixtures/dummy/ernie/dpo-train.jsonl",
"output_dir": output_dir,
}
config_path = os.path.join(CONFIG_PATH, "lora_tp_pp.yaml")
updated_config_path = self.dpotrain_tester.update_training_args(config_path, output_dir, update_args)
train_path = os.path.join(TRAIN_PATH, "alignment/dpo/run_dpo.py")
cmd = [
"python",
"-u",
train_path,
updated_config_path,
]
training_p = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)

# test training result
self.dpotrain_tester.assert_result(training_p.returncode, training_p.stdout)

# test training loss
EXCEPTED_LOSS = 0.495105
self.dpotrain_tester.assert_loss(training_p.stdout, EXCEPTED_LOSS)

# test lora merge
lora_merge_output_dir = os.path.join(output_dir, "lora_tp_pp_merge")
lora_merge_path = os.path.join(TRAIN_PATH, "tools/mergekit.py")

lora_merge_cmd = [
"python",
"-u",
lora_merge_path,
"--lora_model_path",
output_dir,
"--model_name_or_path",
MODEL_NAME_OR_PATH,
"--output_path",
lora_merge_output_dir,
]
lora_merge_p = subprocess.run(lora_merge_cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)
self.dpotrain_tester.assert_result(lora_merge_p.returncode, lora_merge_p.stdout)

# test lora_merge_model generate
EXPECTED_RESULT = paddle.to_tensor(
[[22407, 120525, 77505, 113631, 47887, 134141, 122487, 61092, 40897, 40601]]
)
self.dpotrain_tester.create_and_check_model_generate(lora_merge_output_dir, EXPECTED_RESULT)
146 changes: 144 additions & 2 deletions scripts/regression/test_sft.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,6 +60,8 @@ def assert_loss(self, output, base_loss):
avg_loss = round(sum_loss, 6)
else:
avg_loss = 0
print(f"Current loss : {avg_loss}")
print(f"Base loss : {base_loss}")
self.assertTrue(abs(avg_loss - base_loss) <= 0.0001, f"loss: {avg_loss}, base_loss: {base_loss}, exist diff!")

def assert_result(self, ret_code, log_output):
Expand Down Expand Up @@ -130,7 +132,7 @@ def test_sft_full(self):
self.sfttrain_tester.assert_result(training_p.returncode, training_p.stdout)

# test training loss
EXCEPTED_LOSS = 11.945683
EXCEPTED_LOSS = 11.945673
self.sfttrain_tester.assert_loss(training_p.stdout, EXCEPTED_LOSS)

# test model resume
Expand Down Expand Up @@ -179,7 +181,7 @@ def test_sft_lora(self):
self.sfttrain_tester.assert_result(training_p.returncode, training_p.stdout)

# test training loss
EXCEPTED_LOSS = 11.956834
EXCEPTED_LOSS = 11.956829
self.sfttrain_tester.assert_loss(training_p.stdout, EXCEPTED_LOSS)

# test lora merge
Expand Down Expand Up @@ -209,3 +211,143 @@ def test_sft_lora(self):
[[22407, 120525, 77505, 113631, 47887, 134141, 122487, 61092, 40897, 40601]]
)
self.sfttrain_tester.create_and_check_model_generate(lora_merge_output_dir, EXPECTED_RESULT)

def test_sft_full_tp_pp(self):
output_dir = os.path.join(OUTPUT_DIR, "sft_full_tp_pp")
update_args = {
"model_name_or_path": MODEL_NAME_OR_PATH,
"train_dataset_path": "./tests/fixtures/dummy/ernie/sft-train.jsonl",
"eval_dataset_path": "./tests/fixtures/dummy/ernie/sft-train.jsonl",
"output_dir": output_dir,
}
config_path = os.path.join(CONFIG_PATH, "full_tp_pp.yaml")
updated_config_path = self.sfttrain_tester.update_training_args(config_path, output_dir, update_args)
train_path = os.path.join(TRAIN_PATH, "run_finetune.py")
# n_gpus = paddle.device.cuda.device_count()
# devices = ",".join(str(i) for i in range(n_gpus))
cmd = [
"python",
"-u",
"-m",
"paddle.distributed.launch",
"--devices",
"0,1,2,3",
train_path,
updated_config_path,
]
print(f"cmd {cmd}")
training_p = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)

# test training result
self.sfttrain_tester.assert_result(training_p.returncode, training_p.stdout)

# test training loss
EXCEPTED_LOSS = 11.956903
self.sfttrain_tester.assert_loss(training_p.stdout, EXCEPTED_LOSS)

# test model resume
# EXCEPTED_LOSS = 9.550503
# self.sfttrain_tester.assert_loss(reusme_p.stdout, EXCEPTED_LOSS)
# test model generate
EXPECTED_RESULT = paddle.to_tensor([[22407, 90612, 90612, 90612, 90612, 90612, 90612, 90612, 90612, 90612]])
self.sfttrain_tester.create_and_check_model_generate(output_dir, EXPECTED_RESULT)

def test_sft_lora_tp_pp(self):
output_dir = os.path.join(OUTPUT_DIR, "sft_lora_tp_pp")
update_args = {
"model_name_or_path": MODEL_NAME_OR_PATH,
"train_dataset_path": "./tests/fixtures/dummy/ernie/sft-train.jsonl",
"eval_dataset_path": "./tests/fixtures/dummy/ernie/sft-train.jsonl",
"output_dir": output_dir,
}
config_path = os.path.join(CONFIG_PATH, "lora_tp_pp.yaml")
updated_config_path = self.sfttrain_tester.update_training_args(config_path, output_dir, update_args)
train_path = os.path.join(TRAIN_PATH, "run_finetune.py")
cmd = [
"python",
"-u",
"-m",
"paddle.distributed.launch",
"--devices",
"0,1,2,3",
train_path,
updated_config_path,
]
training_p = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)

# test training result
self.sfttrain_tester.assert_result(training_p.returncode, training_p.stdout)

# test training loss
EXCEPTED_LOSS = 11.956903
self.sfttrain_tester.assert_loss(training_p.stdout, EXCEPTED_LOSS)

# test lora merge
lora_merge_output_dir = os.path.join(output_dir, "lora_tp_pp_merge")
lora_merge_path = os.path.join(TRAIN_PATH, "tools/mergekit.py")

lora_merge_cmd = [
"python",
"-u",
"-m",
"paddle.distributed.launch",
"--devices",
"0,1,2,3",
lora_merge_path,
"--lora_model_path",
output_dir,
"--model_name_or_path",
MODEL_NAME_OR_PATH,
"--output_path",
lora_merge_output_dir,
]
lora_merge_p = subprocess.run(lora_merge_cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)
self.sfttrain_tester.assert_result(lora_merge_p.returncode, lora_merge_p.stdout)

# test lora_merge_model generate
EXPECTED_RESULT = paddle.to_tensor(
[[22407, 120525, 77505, 113631, 47887, 134141, 122487, 61092, 40897, 40601]]
)
self.sfttrain_tester.create_and_check_model_generate(lora_merge_output_dir, EXPECTED_RESULT)

def test_sft_full_function_call(self):
output_dir = os.path.join(OUTPUT_DIR, "sft_full_function_call")
update_args = {
"model_name_or_path": MODEL_NAME_OR_PATH,
"train_dataset_path": "./tests/fixtures/dummy/function-call/function-call-train.jsonl",
"eval_dataset_path": "./tests/fixtures/dummy/function-call/function-call-eval.jsonl",
"output_dir": output_dir,
}
config_path = os.path.join(CONFIG_PATH, "full_function_call.yaml")
updated_config_path = self.sfttrain_tester.update_training_args(config_path, output_dir, update_args)
train_path = os.path.join(TRAIN_PATH, "run_finetune.py")
cmd = [
"python",
"-u",
"-m",
"paddle.distributed.launch",
"--devices",
"0,1,2,3",
train_path,
updated_config_path,
]
print(f"cmd {cmd}")
training_p = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)

# test training result
self.sfttrain_tester.assert_result(training_p.returncode, training_p.stdout)

# test training loss
EXCEPTED_LOSS = 11.769741
self.sfttrain_tester.assert_loss(training_p.stdout, EXCEPTED_LOSS)

# test model resume
# reusme_p = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)
# self.sfttrain_tester.assert_result(reusme_p.returncode, reusme_p.stdout)

# EXCEPTED_LOSS = 9.550503
# self.sfttrain_tester.assert_loss(reusme_p.stdout, EXCEPTED_LOSS)

# test model generate
EXPECTED_RESULT = paddle.to_tensor([[22407, 90612, 90612, 90612, 90612, 90612, 90612, 90612, 90612, 90612]])
self.sfttrain_tester.create_and_check_model_generate(output_dir, EXPECTED_RESULT)
Loading