升级LLM分支到2.3.0-MindSpeed-RL-AtomGit

Mcli/train_dapo.py+2-1

	@@ -248,7 +248,8 @@ def get_megatron_module():
248	from megatron.training.training import setup_model_and_optimizer	248	from megatron.training.training import setup_model_and_optimizer
249	from megatron.core.enums import ModelType	249	from megatron.core.enums import ModelType
250	from megatron.core.distributed import finalize_model_grads	250	from megatron.core.distributed import finalize_model_grads
251	- from mindspeed.utils import set_actual_seq_len, set_position_ids, get_actual_seq_len	251	+ from mindspeed.utils import set_position_ids
		252	+ from mindspeed.core.context_parallel.get_batch_utils import set_actual_seq_len, get_actual_seq_len
252	from megatron.core.optimizer.distrib_optimizer import DistributedOptimizer	253	from megatron.core.optimizer.distrib_optimizer import DistributedOptimizer
253	from megatron.core.optimizer.optimizer import Float16OptimizerWithFloat16Params	254	from megatron.core.optimizer.optimizer import Float16OptimizerWithFloat16Params
254		255

Mcli/train_dpo.py+3-3

	@@ -202,11 +202,11 @@ def initialize_megatron(
202	load_args_from_checkpoint(args)	202	load_args_from_checkpoint(args)
203		203
204	_ori_vari_length = None	204	_ori_vari_length = None
205	- _ori_vari_length = args.variable_seq_lengths	205	+ _ori_vari_length = args.no_pad_to_seq_lengths
206	- args.variable_seq_lengths = False	206	+ args.no_pad_to_seq_lengths = False
207	validate_args(args, args_defaults)	207	validate_args(args, args_defaults)
208	if _ori_vari_length is not None:	208	if _ori_vari_length is not None:
209	- args.variable_seq_lengths = _ori_vari_length	209	+ args.no_pad_to_seq_lengths = _ori_vari_length
210		210
211	set_global_variables(args)	211	set_global_variables(args)
212		212

Mcli/train_grpo.py+2-1

	@@ -260,7 +260,8 @@ def get_megatron_module():
260	from megatron.training.training import setup_model_and_optimizer	260	from megatron.training.training import setup_model_and_optimizer
261	from megatron.core.enums import ModelType	261	from megatron.core.enums import ModelType
262	from megatron.core.distributed import finalize_model_grads	262	from megatron.core.distributed import finalize_model_grads
263	- from mindspeed.utils import set_actual_seq_len, set_position_ids, get_actual_seq_len	263	+ from mindspeed.utils import set_position_ids
		264	+ from mindspeed.core.context_parallel.get_batch_utils import set_actual_seq_len, get_actual_seq_len
264	from megatron.core.optimizer.distrib_optimizer import DistributedOptimizer	265	from megatron.core.optimizer.distrib_optimizer import DistributedOptimizer
265	from megatron.core.optimizer.optimizer import Float16OptimizerWithFloat16Params	266	from megatron.core.optimizer.optimizer import Float16OptimizerWithFloat16Params
266		267

Mcli/train_ppo.py+2-1

	@@ -291,7 +291,8 @@ def get_megatron_module():
291	from megatron.training.training import setup_model_and_optimizer	291	from megatron.training.training import setup_model_and_optimizer
292	from megatron.core.enums import ModelType	292	from megatron.core.enums import ModelType
293	from megatron.core.distributed import finalize_model_grads	293	from megatron.core.distributed import finalize_model_grads
294	- from mindspeed.utils import set_actual_seq_len, set_position_ids, get_actual_seq_len	294	+ from mindspeed.utils import set_position_ids
		295	+ from mindspeed.core.context_parallel.get_batch_utils import set_actual_seq_len, get_actual_seq_len
295	from megatron.core.optimizer.distrib_optimizer import DistributedOptimizer	296	from megatron.core.optimizer.distrib_optimizer import DistributedOptimizer
296	from megatron.core.optimizer.optimizer import Float16OptimizerWithFloat16Params	297	from megatron.core.optimizer.optimizer import Float16OptimizerWithFloat16Params
297		298

Mconfigs/dapo_qwen25_32b_A2_20k.yaml+1-4

	@@ -3,9 +3,6 @@ defaults:
3	- qwen25_32b	3	- qwen25_32b
4		4
5	megatron_training:	5	megatron_training:
6	- use_ascend_coc: True
7	- coc_fused_kernel: True
8	- coc_parallel_num: 2
9	model: qwen25_32b	6	model: qwen25_32b
10	use_fused_rmsnorm: true	7	use_fused_rmsnorm: true
11	use_mcore_models: true	8	use_mcore_models: true
	@@ -29,7 +26,7 @@ megatron_training:
29	hidden_dropout: 0.0	26	hidden_dropout: 0.0
30	distributed_backend: nccl	27	distributed_backend: nccl
31	no_shared_storage: true	28	no_shared_storage: true
32	- variable_seq_lengths: true	29	+ no_pad_to_seq_lengths: true
33	dataset_additional_keys: ['labels',]	30	dataset_additional_keys: ['labels',]
34	data_path: ./data	31	data_path: ./data
35	split: 100,0,0	32	split: 100,0,0

Mconfigs/dapo_qwen25_32b_A3_32k.yaml+1-4

	@@ -3,9 +3,6 @@ defaults:
3	- qwen25_32b	3	- qwen25_32b
4		4
5	megatron_training:	5	megatron_training:
6	- use_ascend_coc: True
7	- coc_fused_kernel: True
8	- coc_parallel_num: 2
9	model: qwen25_32b	6	model: qwen25_32b
10	use_fused_rmsnorm: true	7	use_fused_rmsnorm: true
11	use_mcore_models: true	8	use_mcore_models: true
	@@ -29,7 +26,7 @@ megatron_training:
29	hidden_dropout: 0.0	26	hidden_dropout: 0.0
30	distributed_backend: nccl	27	distributed_backend: nccl
31	no_shared_storage: true	28	no_shared_storage: true
32	- variable_seq_lengths: true	29	+ no_pad_to_seq_lengths: true
33	dataset_additional_keys: ['labels',]	30	dataset_additional_keys: ['labels',]
34	data_path: ./data	31	data_path: ./data
35	split: 100,0,0	32	split: 100,0,0

Mconfigs/dapo_qwen3_30b_a3b_A3.yaml+1-1

	@@ -27,7 +27,7 @@ megatron_training:
27	hidden_dropout: 0.0	27	hidden_dropout: 0.0
28	distributed_backend: nccl	28	distributed_backend: nccl
29	no_shared_storage: true	29	no_shared_storage: true
30	- variable_seq_lengths: true	30	+ no_pad_to_seq_lengths: true
31	dataset_additional_keys: ['labels',]	31	dataset_additional_keys: ['labels',]
32	data_path: ./data	32	data_path: ./data
33	split: 100,0,0	33	split: 100,0,0

Mconfigs/dapo_qwen3_32b_A3.yaml+1-1

	@@ -27,7 +27,7 @@ megatron_training:
27	hidden_dropout: 0.0	27	hidden_dropout: 0.0
28	distributed_backend: nccl	28	distributed_backend: nccl
29	no_shared_storage: true	29	no_shared_storage: true
30	- variable_seq_lengths: true	30	+ no_pad_to_seq_lengths: true
31	dataset_additional_keys: ['labels',]	31	dataset_additional_keys: ['labels',]
32	data_path: ./data	32	data_path: ./data
33	split: 100,0,0	33	split: 100,0,0

Mconfigs/dpo_qwen3_30b_a3b_A3.yaml+1-1

	@@ -7,7 +7,7 @@ megatron_training:
7	finetune: true	7	finetune: true
8	stage: dpo	8	stage: dpo
9	is_pairwise_dataset: true	9	is_pairwise_dataset: true
10	- variable_seq_lengths: true	10	+ no_pad_to_seq_lengths: true
11	tokenizer_not_use_fast: true	11	tokenizer_not_use_fast: true
12	dpo_loss_type: sigmoid	12	dpo_loss_type: sigmoid
13	# gpt_args:	13	# gpt_args:

Mconfigs/grpo_deepseek_r1_671b_A2.yaml+2-2

	@@ -26,14 +26,14 @@ megatron_training:
26	hidden_dropout: 0.0	26	hidden_dropout: 0.0
27	distributed_backend: nccl	27	distributed_backend: nccl
28	no_shared_storage: true	28	no_shared_storage: true
29	- variable_seq_lengths: true	29	+ no_pad_to_seq_lengths: true
30	dataset_additional_keys: ['labels',]	30	dataset_additional_keys: ['labels',]
31	no_shuffle: false	31	no_shuffle: false
32	data_path: ./deepscaler	32	data_path: ./deepscaler
33	split: 100,0,0	33	split: 100,0,0
34	swap_optimizer: true	34	swap_optimizer: true
35	moe_alltoall_overlap_comm: true	35	moe_alltoall_overlap_comm: true
36	- reset_position_ids: true	36	+ reset_attention_mask: true
37		37
38	actor_config:	38	actor_config:
39	model: deepseekv3_671b	39	model: deepseekv3_671b

Mconfigs/grpo_deepseek_r1_671b_A3.yaml+2-2

	@@ -27,7 +27,7 @@ megatron_training:
27	hidden_dropout: 0.0	27	hidden_dropout: 0.0
28	distributed_backend: nccl	28	distributed_backend: nccl
29	no_shared_storage: true	29	no_shared_storage: true
30	- variable_seq_lengths: true	30	+ no_pad_to_seq_lengths: true
31	dataset_additional_keys: ['labels']	31	dataset_additional_keys: ['labels']
32	data_path: ./deepseek/deepscaler	32	data_path: ./deepseek/deepscaler
33	split: 100,0,0	33	split: 100,0,0
	@@ -36,7 +36,7 @@ megatron_training:
36	seed: 1	36	seed: 1
37	swap_optimizer: true	37	swap_optimizer: true
38	moe_alltoall_overlap_comm: true	38	moe_alltoall_overlap_comm: true
39	- reset_position_ids: true	39	+ reset_attention_mask: true
40		40
41	actor_config:	41	actor_config:
42	model: deepseekv3_671b	42	model: deepseekv3_671b

Mconfigs/grpo_deepseek_r1_671b_A3_eplb.yaml+2-2

	@@ -27,7 +27,7 @@ megatron_training:
27	hidden_dropout: 0.0	27	hidden_dropout: 0.0
28	distributed_backend: nccl	28	distributed_backend: nccl
29	no_shared_storage: true	29	no_shared_storage: true
30	- variable_seq_lengths: true	30	+ no_pad_to_seq_lengths: true
31	dataset_additional_keys: ['labels']	31	dataset_additional_keys: ['labels']
32	data_path: ./deepseek/deepscaler	32	data_path: ./deepseek/deepscaler
33	split: 100,0,0	33	split: 100,0,0
	@@ -36,7 +36,7 @@ megatron_training:
36	seed: 1	36	seed: 1
37	swap_optimizer: true	37	swap_optimizer: true
38	moe_alltoall_overlap_comm: true	38	moe_alltoall_overlap_comm: true
39	- reset_position_ids: true	39	+ reset_attention_mask: true
40		40
41	actor_config:	41	actor_config:
42	model: deepseekv3_671b	42	model: deepseekv3_671b

Mconfigs/grpo_qwen25_32b_A2.yaml+2-5

	@@ -3,9 +3,6 @@ defaults:
3	- qwen25_32b	3	- qwen25_32b
4		4
5	megatron_training:	5	megatron_training:
6	- use_ascend_coc: True
7	- coc_fused_kernel: True
8	- coc_parallel_num: 2
9	model: qwen25_32b	6	model: qwen25_32b
10	use_fused_rmsnorm: true	7	use_fused_rmsnorm: true
11	use_mcore_models: true	8	use_mcore_models: true
	@@ -29,11 +26,11 @@ megatron_training:
29	hidden_dropout: 0.0	26	hidden_dropout: 0.0
30	distributed_backend: nccl	27	distributed_backend: nccl
31	no_shared_storage: true	28	no_shared_storage: true
32	- variable_seq_lengths: true	29	+ no_pad_to_seq_lengths: true
33	dataset_additional_keys: ['labels']	30	dataset_additional_keys: ['labels']
34	data_path: ./data	31	data_path: ./data
35	split: 100,0,0	32	split: 100,0,0
36	- reset_position_ids: true	33	+ reset_attention_mask: true
37	seed: 1	34	seed: 1
38		35
39	actor_config:	36	actor_config:

Mconfigs/grpo_qwen25_32b_A3.yaml+2-2

	@@ -26,14 +26,14 @@ megatron_training:
26	hidden_dropout: 0.0	26	hidden_dropout: 0.0
27	distributed_backend: nccl	27	distributed_backend: nccl
28	no_shared_storage: true	28	no_shared_storage: true
29	- variable_seq_lengths: true	29	+ no_pad_to_seq_lengths: true
30	dataset_additional_keys: ['labels']	30	dataset_additional_keys: ['labels']
31	data_path: ./data	31	data_path: ./data
32	split: 100,0,0	32	split: 100,0,0
33	no_shuffle: false	33	no_shuffle: false
34	full_shuffle_instruction_dataset: false	34	full_shuffle_instruction_dataset: false
35	seed: 1	35	seed: 1
36	- reset_position_ids: true	36	+ reset_attention_mask: true
37		37
38	actor_config:	38	actor_config:
39	model: qwen25_32b	39	model: qwen25_32b

Mconfigs/grpo_qwen25_7b_A3.yaml+0-3

	@@ -3,9 +3,6 @@ defaults:
3	- qwen25_7b	3	- qwen25_7b
4		4
5	megatron_training:	5	megatron_training:
6	- use_ascend_coc: True
7	- coc_fused_kernel: True
8	- coc_parallel_num: 2
9	model: qwen25_7b	6	model: qwen25_7b
10	use_fused_rmsnorm: true	7	use_fused_rmsnorm: true
11	use_mcore_models: true	8	use_mcore_models: true

Mconfigs/grpo_qwen3_235b_a22b_A2.yaml+2-2

	@@ -29,12 +29,12 @@ megatron_training:
29	hidden_dropout: 0.0	29	hidden_dropout: 0.0
30	distributed_backend: nccl	30	distributed_backend: nccl
31	no_shared_storage: true	31	no_shared_storage: true
32	- variable_seq_lengths: true	32	+ no_pad_to_seq_lengths: true
33	dataset_additional_keys: ['labels',]	33	dataset_additional_keys: ['labels',]
34	data_path: ./data	34	data_path: ./data
35	split: 100,0,0	35	split: 100,0,0
36	swap_optimizer: true	36	swap_optimizer: true
37	- reset_position_ids: true	37	+ reset_attention_mask: true
38		38
39	actor_config:	39	actor_config:
40	model: qwen3_235b_a22b	40	model: qwen3_235b_a22b

Mconfigs/ppo_qwen25_32b_A3.yaml+1-1

	@@ -27,7 +27,7 @@ megatron_training:
27	hidden_dropout: 0.0	27	hidden_dropout: 0.0
28	distributed_backend: nccl	28	distributed_backend: nccl
29	no_shared_storage: true	29	no_shared_storage: true
30	- variable_seq_lengths: true	30	+ no_pad_to_seq_lengths: true
31	dataset_additional_keys: ['labels']	31	dataset_additional_keys: ['labels']
32	data_path: ./data	32	data_path: ./data
33	split: 100,0,0	33	split: 100,0,0

Mdocs/features/context_parallel.md+7-7

	@@ -21,14 +21,15 @@
21	actor_config:	21	actor_config:
22	context_parallel_size: 2	22	context_parallel_size: 2
23	context_parallel_algo: ulysses_cp_algo	23	context_parallel_algo: ulysses_cp_algo
		24	+ attention_mask_type: general
24		25
25	# 与remove_padding特性一起使用	26	# 与remove_padding特性一起使用
26	rl_config:	27	rl_config:
27	use_remove_padding: true	28	use_remove_padding: true
28		29
29	megatron_training:	30	megatron_training:
30	- reset_position_ids: true	31	+ reset_attention_mask: true
31	- variable_seq_lengths: true	32	+ no_pad_to_seq_lengths: true
32	```	33	```
33		34
34	对于直接偏好对齐（DPO）算法，通过如下配置可以使能：	35	对于直接偏好对齐（DPO）算法，通过如下配置可以使能：
	@@ -36,7 +37,7 @@ megatron_training:
36	```yaml	37	```yaml
37	# 填写在megatron_training	38	# 填写在megatron_training
38	megatron_training:	39	megatron_training:
39	- variable_seq_lengths: true	40	+ no_pad_to_seq_lengths: true
40	context_parallel_size: 2	41	context_parallel_size: 2
41	context_parallel_algo: ulysses_cp_algo	42	context_parallel_algo: ulysses_cp_algo
42	```	43	```
	@@ -46,7 +47,7 @@ megatron_training:
46	`context_parallel_algo` 表示选用的长序列并行方法，如果不配置此参数，默认取ulysses_cp_algo	47	`context_parallel_algo` 表示选用的长序列并行方法，如果不配置此参数，默认取ulysses_cp_algo
47		48
48	特别的，此特性与remove_padding一起使用时，配置说明如下：	49	特别的，此特性与remove_padding一起使用时，配置说明如下：
49	-直接叠加使能remove_padding的配置即可：将use_remove_padding、reset_position_ids、variable_seq_lengths都设置为true。	50	+直接叠加使能remove_padding的配置即可：将use_remove_padding、reset_attention_mask、no_pad_to_seq_lengths都设置为true。
50		51
51	### Ring Attention	52	### Ring Attention
52	```yaml	53	```yaml
	@@ -60,8 +61,7 @@ rl_config:
60	use_remove_padding: true	61	use_remove_padding: true
61		62
62	megatron_training:	63	megatron_training:
63	- reset_position_ids: true	64	+ no_pad_to_seq_lengths: true
64	- variable_seq_lengths: true
65	reset_attention_mask: true	65	reset_attention_mask: true
66	```	66	```
67		67
	@@ -76,4 +76,4 @@ megatron_training:
76		76
77	特别的，此特性与remove_padding一起使用时，配置说明如下：	77	特别的，此特性与remove_padding一起使用时，配置说明如下：
78		78
79	-除了需要将use_remove_padding、reset_position_ids、variable_seq_lengths都设置为true之外（使能remove_padding），还需要将reset_attention_mask设置为true; 否则这些配置都不要配，默认false。	79	+除了需要将use_remove_padding、reset_attention_mask、no_pad_to_seq_lengths都设置为true之外（使能remove_padding），还需要将reset_attention_mask设置为true; 否则这些配置都不要配，默认false。

Mdocs/features/remove_padding.md+2-2

	@@ -34,8 +34,8 @@
34		34
35	```yaml	35	```yaml
36	megatron_training:	36	megatron_training:
37	- variable_seq_lengths: true	37	+ no_pad_to_seq_lengths: true
38	- reset_position_ids: true	38	+ reset_attention_mask: true
39		39
40	rl_config:	40	rl_config:
41	use_remove_padding: true	41	use_remove_padding: true

Mmindspeed_rl/config_cls/megatron_config.py+3-2

	@@ -111,7 +111,7 @@ class MegatronConfig(BaseConfig):
111	is_instruction_dataset: Whether the dataset is instruction-based (default: False)	111	is_instruction_dataset: Whether the dataset is instruction-based (default: False)
112	is_pairwise_dataset: Whether the dataset is pairwise format that has a chosen sequence and rejected	112	is_pairwise_dataset: Whether the dataset is pairwise format that has a chosen sequence and rejected
113	sequence, which usually used in reinforce learning (default: False)	113	sequence, which usually used in reinforce learning (default: False)
114	- variable_seq_lengths: Whether to use variable sequence lengths (default: False)	114	+ variable_seq_length: Whether to use variable sequence lengths (default: False)
115	no_shuffle: Whether to shuffle the dataset (default: False)	115	no_shuffle: Whether to shuffle the dataset (default: False)
116	stage: Stage of the model (default: None)	116	stage: Stage of the model (default: None)
117	sequence_parallel: Whether to use sequence parallelism (default: False)	117	sequence_parallel: Whether to use sequence parallelism (default: False)
	@@ -311,7 +311,8 @@ class MegatronConfig(BaseConfig):
311	self.no_save_rng = None	311	self.no_save_rng = None
312	self.is_instruction_dataset = False	312	self.is_instruction_dataset = False
313	self.is_pairwise_dataset = False	313	self.is_pairwise_dataset = False
314	- self.variable_seq_lengths = False	314	+ self.variable_seq_length = False
		315	+ self.no_pad_to_seq_lengths = False
315	self.no_shuffle = False	316	self.no_shuffle = False
316	self.stage = None	317	self.stage = None
317	self.sequence_parallel = False	318	self.sequence_parallel = False

Mmindspeed_rl/config_cls/validate_config.py+6-10

	@@ -55,26 +55,22 @@ def validate_rl_args(
55	if actor_config.context_parallel_size > 1 and actor_config.context_parallel_algo is not None:	55	if actor_config.context_parallel_size > 1 and actor_config.context_parallel_algo is not None:
56	if actor_config.context_parallel_algo not in ["ulysses_cp_algo", "megatron_cp_algo"]:	56	if actor_config.context_parallel_algo not in ["ulysses_cp_algo", "megatron_cp_algo"]:
57	raise ValueError("Now just support ulysses CP and megatron cp(ring)")	57	raise ValueError("Now just support ulysses CP and megatron cp(ring)")
58	- if actor_config.attention_mask_type not in ["causal"]:	58	+
59	- raise ValueError("Now just support causal attention_mask_type")
60	if actor_config.context_parallel_algo == "megatron_cp_algo" and actor_config.context_parallel_size > 1 and rl_config.use_remove_padding:	59	if actor_config.context_parallel_algo == "megatron_cp_algo" and actor_config.context_parallel_size > 1 and rl_config.use_remove_padding:
61	if not actor_config.reset_attention_mask:	60	if not actor_config.reset_attention_mask:
62	raise ValueError("when use ring cp and remove_padding, reset_attention_mask must be true")	61	raise ValueError("when use ring cp and remove_padding, reset_attention_mask must be true")
63	- if actor_config.context_parallel_size <= 1 or actor_config.context_parallel_algo != "megatron_cp_algo" or not rl_config.use_remove_padding:
64	- if actor_config.reset_attention_mask:
65	- raise ValueError("Just when use ring cp >=2 with remove_padding, reset_attention_mask must be true; otherwise should be false")
66		62
67	# 校验移除填充特性相关配置	63	# 校验移除填充特性相关配置
68	if rl_config.use_remove_padding:	64	if rl_config.use_remove_padding:
69	- if actor_config.pipeline_model_parallel_size > 1 and not actor_config.variable_seq_lengths:	65	+ if actor_config.pipeline_model_parallel_size > 1 and not actor_config.no_pad_to_seq_lengths:
70	raise ValueError(	66	raise ValueError(
71	- "'use_remove_padding' feature requires 'variable_seq_lengths=True' when using pipeline parallelism!"	67	+ "'use_remove_padding' feature requires 'no_pad_to_seq_lengths=True' when using pipeline parallelism!"
72	- "If you want to use context parallelism under this premise and encounter the mindspeed_llm validation error about variable_seq_lengths, "	68	+ "If you want to use context parallelism under this premise and encounter the mindspeed_llm validation error about no_pad_to_seq_lengths, "
73	"you just need to delete the validation code of mindspeed_llm, and it will not cause problems.")	69	"you just need to delete the validation code of mindspeed_llm, and it will not cause problems.")
74		70
75	- if not actor_config.reset_position_ids:	71	+ if not actor_config.reset_attention_mask:
76	raise ValueError(	72	raise ValueError(
77	- "'use_remove_padding' feature requires 'reset_position_ids=True'! ")	73	+ "'use_remove_padding' feature requires 'reset_attention_mask=True'! ")
78		74
79	if rl_config.is_multimodal:	75	if rl_config.is_multimodal:
80	raise ValueError(	76	raise ValueError(

Mmindspeed_rl/models/base/base_training_engine.py+1-1

	@@ -316,7 +316,7 @@ class BaseTrainingEngine(ABC):
316	cu_seqlens_padded_ring = cu_seqlens_padded	316	cu_seqlens_padded_ring = cu_seqlens_padded
317	if self.megatron_config.attention_mask_type == 'causal':	317	if self.megatron_config.attention_mask_type == 'causal':
318	cu_seqlens_padded_ring = (cu_seqlens_padded / get_ring_degree(self.megatron_config)).to(torch.int)	318	cu_seqlens_padded_ring = (cu_seqlens_padded / get_ring_degree(self.megatron_config)).to(torch.int)
319	- self.set_actual_seq_len(cu_seqlens_padded_ring.tolist())	319	+ self.set_actual_seq_len(cu_seqlens_padded_ring)
320		320
321	if cp_size > 1:	321	if cp_size > 1:
322	input_ids, position_ids, batch, index = self._get_batch_data_with_cp(batch, input_ids, position_ids, labels)	322	input_ids, position_ids, batch, index = self._get_batch_data_with_cp(batch, input_ids, position_ids, labels)

Mtests/st/configs/test_grpo_trainer_qwen25_7b_integrated.yaml+2-5

	@@ -8,15 +8,12 @@ megatron_training:
8	use_mcore_models: true	8	use_mcore_models: true
9	sequence_parallel: true	9	sequence_parallel: true
10	use_flash_attn: true	10	use_flash_attn: true
11	- reset_position_ids: true	11	+ reset_attention_mask: true
12	no_masked_softmax_fusion: true	12	no_masked_softmax_fusion: true
13	attention_softmax_in_fp32: true	13	attention_softmax_in_fp32: true
14	no_gradient_accumulation_fusion: true	14	no_gradient_accumulation_fusion: true
15	use_fused_swiglu: true	15	use_fused_swiglu: true
16	use_fused_rotary_pos_emb: true	16	use_fused_rotary_pos_emb: true
17	- use_ascend_coc: True
18	- coc_fused_kernel: True
19	- coc_parallel_num: 2
20	bf16: true	17	bf16: true
21	use_distributed_optimizer: true	18	use_distributed_optimizer: true
22	tokenizer_type: PretrainedFromHF	19	tokenizer_type: PretrainedFromHF
	@@ -31,7 +28,7 @@ megatron_training:
31	hidden_dropout: 0.0	28	hidden_dropout: 0.0
32	distributed_backend: nccl	29	distributed_backend: nccl
33	no_shared_storage: true	30	no_shared_storage: true
34	- variable_seq_lengths: true	31	+ no_pad_to_seq_lengths: true
35	dataset_additional_keys: ['labels',]	32	dataset_additional_keys: ['labels',]
36	data_path: /data/for_dt/datasets/pe-nlp/data	33	data_path: /data/for_dt/datasets/pe-nlp/data
37	split: 100,0,0	34	split: 100,0,0

Mtests/st/resharding/test_resharding.py+1-1

	@@ -69,7 +69,7 @@ def make_megatron_config(args):
69	'tokenizer_type': args.tokenizer_type, 'tokenizer_name_or_path': args.tokenizer_path, 'global_batch_size': 2,	69	'tokenizer_type': args.tokenizer_type, 'tokenizer_name_or_path': args.tokenizer_path, 'global_batch_size': 2,
70	'seq_length': 512, 'save_interval': 10000,	70	'seq_length': 512, 'save_interval': 10000,
71	'train_iters': 1000, 'distributed_backend': 'nccl', 'no_shared_storage': True,	71	'train_iters': 1000, 'distributed_backend': 'nccl', 'no_shared_storage': True,
72	- 'variable_seq_lengths': True, 'micro_batch_size': 1, 'tensor_model_parallel_size': args.train_tp,	72	+ 'no_pad_to_seq_lengths': True, 'micro_batch_size': 1, 'tensor_model_parallel_size': args.train_tp,
73	'pipeline_model_parallel_size': args.train_pp, 'lr': '1e-7', 'lr_decay_style': 'constant',	73	'pipeline_model_parallel_size': args.train_pp, 'lr': '1e-7', 'lr_decay_style': 'constant',
74	'min_lr': 0.0, 'weight_decay': 0.0, 'lr_warmup_fraction': 0.0, 'clip_grad': 10000.0,	74	'min_lr': 0.0, 'weight_decay': 0.0, 'lr_warmup_fraction': 0.0, 'clip_grad': 10000.0,
75	'adam_beta1': 0.9, 'adam_beta2': 0.999, 'initial_loss_scale': 4096, 'finetune': True,	75	'adam_beta1': 0.9, 'adam_beta2': 0.999, 'initial_loss_scale': 4096, 'finetune': True,