2.2.0资料修改-MindSpeed-RL-AtomGit

已合并

2.2.0资料修改 #766

NIE SHIYU创建于 2025年11月28日从xiazhahe/MindSpeed-RL： 2.2.0 合入到 Ascend/MindSpeed-RL： 2.2.0

2.2.0资料修改 #766

已合并

NIE SHIYU创建于 2025年11月28日

从xiazhahe/MindSpeed-RL 2.2.0合入到Ascend/MindSpeed-RL 2.2.0

共 13 个文件变更+10-15

Mcli/train_dapo.py+0-1

	@@ -1,5 +1,4 @@
1	# Copyright (c) 2025, HUAWEI CORPORATION. All rights reserved.	1	# Copyright (c) 2025, HUAWEI CORPORATION. All rights reserved.
2	-# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
3	"""	2	"""
4	Note that we don't combine the main with trainer as trainer is used by other main.	3	Note that we don't combine the main with trainer as trainer is used by other main.
5	"""	4	"""

Mcli/train_grpo.py+0-1

	@@ -1,5 +1,4 @@
1	# Copyright (c) 2025, HUAWEI CORPORATION. All rights reserved.	1	# Copyright (c) 2025, HUAWEI CORPORATION. All rights reserved.
2	-# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
3	"""	2	"""
4	Note that we don't combine the main with trainer as trainer is used by other main.	3	Note that we don't combine the main with trainer as trainer is used by other main.
5	"""	4	"""

Mcli/train_ppo.py+0-1

	@@ -1,5 +1,4 @@
1	# Copyright (c) 2025, HUAWEI CORPORATION. All rights reserved.	1	# Copyright (c) 2025, HUAWEI CORPORATION. All rights reserved.
2	-# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
3	"""	2	"""
4	Note that we don't combine the main with trainer as trainer is used by other main.	3	Note that we don't combine the main with trainer as trainer is used by other main.
5	"""	4	"""

Mdocs/algorithms/ppo.md+3-4

	@@ -204,23 +204,22 @@ rl_config:
204	\| `e2e_tps` \| 端到端的tokens/p/s指标 \|	204	\| `e2e_tps` \| 端到端的tokens/p/s指标 \|
205	\| `update_tps` \| 训练的tokens/p/s指标 \|	205	\| `update_tps` \| 训练的tokens/p/s指标 \|
206	\| `vllm_tps` \| 推理的tokens/p/s指标 \|	206	\| `vllm_tps` \| 推理的tokens/p/s指标 \|
207	-
208	* e2e_tps计算方式	207	* e2e_tps计算方式
209		208
210	$$	209	$$
211	-(\text{response_length_mean} + \text{prompt_length_mean}) \times \text{global_batch_size} \times \text{n_samples_per_prompt} / \text{world_size} \ / \text{time_all}	210	+(\text{response\_length\_mean} + \text{prompt\_length\_mean}) \times \text{global\_batch\_size} \times \text{n\_samples\_per\_prompt} / \text{world\_size} \ / \text{time\_all}
212	$$	211	$$
213		212
214	* update_tps计算方式	213	* update_tps计算方式
215		214
216	$$	215	$$
217	-(\text{response_length_mean} + \text{prompt_length_mean}) \times \text{global_batch_size} \times \text{n_samples_per_prompt} / \text{world_size} \ / \text{time_update}	216	+(\text{response\_length\_mean} + \text{prompt\_length\_mean}) \times \text{global\_batch\_size} \times \text{n\_samples\_per\_prompt} / \text{world\_size} \ / \text{time\_update}
218	$$	217	$$
219		218
220	* vllm_tps计算方式	219	* vllm_tps计算方式
221		220
222	$$	221	$$
223	-(\text{response_length_mean} + \text{prompt_length_mean}) \times \text{global_batch_size} \times \text{n_samples_per_prompt} / \text{world_size} \ / \text{time_rollout}	222	+(\text{response\_length\_mean} + \text{prompt\_length\_mean}) \times \text{global\_batch\_size} \times \text{n\_samples\_per\_prompt} / \text{world\_size} \ / \text{time\_rollout}
224	$$	223	$$
225		224
226	注：以上计算公式中 ` time_all`、`time_update`、`time_rollout`、`response_length_mean` 和 `prompt_length_mean` 即分别对应于指标说明里的`timing/all`、`timing/update`、`timing/rollout`、`response_length/mean`和`prompt_length/mean`，此处名字修改是为了区别于公式里的`/`计算符号；	225	注：以上计算公式中 ` time_all`、`time_update`、`time_rollout`、`response_length_mean` 和 `prompt_length_mean` 即分别对应于指标说明里的`timing/all`、`timing/update`、`timing/rollout`、`response_length/mean`和`prompt_length/mean`，此处名字修改是为了区别于公式里的`/`计算符号；

Mexamples/grpo/grpo_trainer_qwen25_7b.sh+4-0

	@@ -14,6 +14,10 @@ mkdir logs
14		14
15	export TASK_QUEUE_ENABLE=2	15	export TASK_QUEUE_ENABLE=2
16	export HCCL_IF_BASE_PORT=24703	16	export HCCL_IF_BASE_PORT=24703
		17	+# 请确认该文件存在(可通过 find /usr -name libjemalloc.so.2 确认)
		18	+export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2
		19	+export MALLOC_MMAP_THRESHOLD_=512768
		20	+export LCAL_COMM_ID=127.0.0.1:27001
17		21
18	NNODES=1	22	NNODES=1
19	NPUS_PER_NODE=16	23	NPUS_PER_NODE=16

Mexamples/ppo/ppo_trainer_qwen25_32b.sh+2-1

	@@ -12,7 +12,8 @@ mkdir logs
12		12
13	export TASK_QUEUE_ENABLE=2	13	export TASK_QUEUE_ENABLE=2
14	export HCCL_IF_BASE_PORT=24703	14	export HCCL_IF_BASE_PORT=24703
15	-export LD_PRELOAD=/usr/local/lib/libjemalloc.so.2	15	+# 请确认该文件存在(可通过 find /usr -name libjemalloc.so.2 确认)
		16	+export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2
16	export MALLOC_MMAP_THRESHOLD_=512768	17	export MALLOC_MMAP_THRESHOLD_=512768
17	export LCAL_COMM_ID=127.0.0.1:27001	18	export LCAL_COMM_ID=127.0.0.1:27001
18		19

Mmindspeed_rl/datasets/build_dataset.py+0-1

	@@ -1,5 +1,4 @@
1	# coding=utf-8	1	# coding=utf-8
2	-# Copyright (c) 2020; NVIDIA CORPORATION. All rights reserved.
3	# Copyright (c) 2025, HUAWEI CORPORATION. All rights reserved.	2	# Copyright (c) 2025, HUAWEI CORPORATION. All rights reserved.
4	#	3	#
5	# Licensed under the Apache License, Version 2.0 (the "License");	4	# Licensed under the Apache License, Version 2.0 (the "License");

Mmindspeed_rl/datasets/data_samplers.py+0-1

	@@ -1,5 +1,4 @@
1	# coding=utf-8	1	# coding=utf-8
2	-# Copyright (c) 2020; NVIDIA CORPORATION. All rights reserved.
3	# Copyright (c) 2025, HUAWEI CORPORATION. All rights reserved.	2	# Copyright (c) 2025, HUAWEI CORPORATION. All rights reserved.
4		3
5	class PretrainingSampler:	4	class PretrainingSampler:

Mmindspeed_rl/datasets/utils.py+0-1

	@@ -1,5 +1,4 @@
1	# coding=utf-8	1	# coding=utf-8
2	-# Copyright (c) 2020; NVIDIA CORPORATION. All rights reserved.
3	# Copyright (c) 2025, HUAWEI CORPORATION. All rights reserved.	2	# Copyright (c) 2025, HUAWEI CORPORATION. All rights reserved.
4		3
5	import os	4	import os

Mmindspeed_rl/trainer/utils/training.py+0-1

	@@ -1,4 +1,3 @@
1	-# Copyright (c) 2022; NVIDIA CORPORATION. All rights reserved.
2	# Copyright (c) 2025, HUAWEI CORPORATION. All rights reserved.	1	# Copyright (c) 2025, HUAWEI CORPORATION. All rights reserved.
3	import time	2	import time
4		3

Mmindspeed_rl/utils/utils.py+0-1

	@@ -1,5 +1,4 @@
1	# Copyright (c) 2025, HUAWEI CORPORATION. All rights reserved.	1	# Copyright (c) 2025, HUAWEI CORPORATION. All rights reserved.
2	-# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
3		2
4	import os	3	import os
5	import re	4	import re

Mtests/st/resharding/test_resharding.py+0-1

	@@ -1,7 +1,6 @@
1	#!/usr/bin/env python	1	#!/usr/bin/env python
2	# -- coding: utf-8 --	2	# -- coding: utf-8 --
3	# Copyright (c) Huawei Technologies Co., Ltd.2023-2025. All rights reserved.	3	# Copyright (c) Huawei Technologies Co., Ltd.2023-2025. All rights reserved.
4	-# Copyright (c) 2025, NVIDIA CORPORATION. All rights reversed.
5	import sys	4	import sys
6	from unittest.mock import MagicMock	5	from unittest.mock import MagicMock
7	from datetime import timedelta	6	from datetime import timedelta

Mtests/ut/features/reuse_fp32_param/test_reuse_fp32_param.py+1-1

	@@ -1,4 +1,4 @@
1	-# Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved.	1	+# Copyright (c) 2025, HUAWEI CORPORATION. All rights reserved.
2		2
3	from functools import partial	3	from functools import partial
4	import copy	4	import copy