已关闭
[ignore] [ci-test] Bad patch 3 Conflicting patches #80
hanboyou创建于 2025年12月22日关闭于 2025年12月24日
[ignore] [ci-test] Bad patch 3 Conflicting patches #80
已关闭
共 4 个文件变更+36-38
| @@ -1,5 +1,11 @@ | |||
| 1 | # Qwen3系列模型 RL训练优化实践样例 | 1 | # Qwen3系列模型 RL训练优化实践样例 |
| 2 | - | 2 | +[2025/12/21 00:33:59.173 GMT+08:00] - [Missing reference] Patch '0001-feature-model_converter.patch' is not mentioned in README.md. |
| 3 | +[2025/12/21 00:33:59.173 GMT+08:00] - [Missing reference] Patch '0006-vllm_ascend-feature-chunk_moe_eager.patch' is not mentioned in README.md. | ||
| 4 | +[2025/12/21 00:33:59.173 GMT+08:00] - [Missing reference] Patch '0011-verl-feature-enable_rollout_rebalance.patch' is not mentioned in README.md. | ||
| 5 | +[2025/12/21 00:33:59.173 GMT+08:00] - [Missing reference] Patch '0001-mindspeed-bugfix-bad_context.patch' is not mentioned in README.md. | ||
| 6 | +[2025/12/21 00:33:59.173 GMT+08:00] - [Missing reference] Patch '0001-megatron-bugfix-ssssss.patch' is not mentioned in README.md. | ||
| 7 | +llm_rl/qwen3/patches/megatron/0001-megatron-bugfix-bad_path.patch | ||
| 8 | +0001-megatron-bugfix-state_ten_verification.patch | ||
| 3 | ## 概述 | 9 | ## 概述 |
| 4 | 本样例针对Qwen3-235B-A22B和Qwen3-32B模型，基于[veRL开源框架](https://github.com/volcengine/verl)，使用veRL原生支持的MindSpeed和vLLM-Ascend框架，完成RL训练全流程的优化适配。 | 10 | 本样例针对Qwen3-235B-A22B和Qwen3-32B模型，基于[veRL开源框架](https://github.com/volcengine/verl)，使用veRL原生支持的MindSpeed和vLLM-Ascend框架，完成RL训练全流程的优化适配。 |
| 5 | 11 | ||
| @@ -293,3 +299,5 @@ bash ray_start_npu.sh TRAIN_SCRIPT ENV_SCRIPT | |||
| 293 | 299 | ||
| 294 | 300 | ||
| 295 | 后续步骤可参考[基于Dockerfile构建环境](#基于dockerfile构建环境) `2. 源码准备及安装所需的python依赖` 和 `3. 使能patch修改`。 | 301 | 后续步骤可参考[基于Dockerfile构建环境](#基于dockerfile构建环境) `2. 源码准备及安装所需的python依赖` 和 `3. 使能patch修改`。 |
| 302 | +0001-megatron-bugfix-conflicting_patch1.patch | ||
| 303 | +0002-megatron-bugfix-conflicting_patch2.patch | ||
Rllm_rl/qwen3/patches/megatron/0001-megatron-bugfix-state_ten_verification.patch→llm_rl/qwen3/patches/megatron/0001-megatron-bugfix-conflicting_patch1.patch+0-0
文件重命名但无更改。
| @@ -0,0 +1,27 @@ | |||
| 1 | +From 0faba7fc5b17bcc5ae428f3abb00495764dd2007 Mon Sep 17 00:00:00 2001 | ||
| 2 | +From: caojingyi <caojingyi@noreply.gitcode.com> | ||
| 3 | +Date: Mon, 17 Nov 2025 20:43:10 +0800 | ||
| 4 | +Subject: [PATCH 1/1] Update megatron: state_ten verification | ||
| 5 | +Add state_ten value check in the DistributedOptimizer class to prevent runtime exceptions caused by None, | ||
| 6 | +thereby improving code robustness. | ||
| 7 | + | ||
| 8 | +--- | ||
| 9 | + llm_rl/qwen3/megatron/core/optimizer/distrib_optimizer.py | 2 ++ | ||
| 10 | + 1 file changed, 2 insertions(+) | ||
| 11 | + | ||
| 12 | +diff --git a/llm_rl/qwen3/megatron/core/optimizer/distrib_optimizer.py b/llm_rl/qwen3/megatron/core/optimizer/distrib_optimizer.py | ||
| 13 | +index a00d1a3..62fe20f 100644 | ||
| 14 | +--- a/llm_rl/qwen3/megatron/core/optimizer/distrib_optimizer.py | ||
| 15 | ++++ b/llm_rl/qwen3/megatron/core/optimizer/distrib_optimizer.py | ||
| 16 | +@@ -1319,6 +1319,8 @@ class DistributedOptimizer(MixedPrecisionOptimizer): | ||
| 17 | + # The optimizer state of STEP is handled | ||
| 18 | + # specifically and is read from param_groups. | ||
| 19 | + continue | ||
| 20 | ++ if state_ten is None: | ||
| 21 | ++ return [] | ||
| 22 | + replace_kwargs = dict( | ||
| 23 | + key=f'{prefix}.{state_key}.{sharded_metadata.key}', | ||
| 24 | + data=state_ten, | ||
| 25 | +-- | ||
| 26 | +2.50.1.windows.1 | ||
| 27 | + | ||
| @@ -1,37 +0,0 @@ | |||
| 1 | -From 233fadd6ac6259494429fe104f66c49a90ddaaf9 Mon Sep 17 00:00:00 2001 | ||
| 2 | -From: qushiyi <qushiyi@huawei.com> | ||
| 3 | -Date: Wed, 26 Nov 2025 18:19:26 +0800 | ||
| 4 | -Subject: [PATCH] megatron-feature-enable_hdp | ||
| 5 | - | ||
| 6 | ---- | ||
| 7 | - .../megatron/core/models/common/embeddings/rope_utils.py | 7 +++++++ | ||
| 8 | - 1 file changed, 7 insertions(+) | ||
| 9 | - | ||
| 10 | -diff --git a/llm_rl/qwen3/megatron/core/models/common/embeddings/rope_utils.py b/llm_rl/qwen3/megatron/core/models/common/embeddings/rope_utils.py | ||
| 11 | -index 08c9457..b38a6ca 100644 | ||
| 12 | ---- a/llm_rl/qwen3/megatron/core/models/common/embeddings/rope_utils.py | ||
| 13 | -+++ b/llm_rl/qwen3/megatron/core/models/common/embeddings/rope_utils.py | ||
| 14 | -@@ -43,6 +43,8 @@ try: | ||
| 15 | - except ImportError: | ||
| 16 | - apply_rotary_emb_flash = None | ||
| 17 | - | ||
| 18 | -+from patches.verl.utils.hybrid_data_parallel.utils import set_batch_hdp_group, get_batch_hdp_group | ||
| 19 | -+ | ||
| 20 | - | ||
| 21 | - __all__ = ['apply_rotary_emb_flash'] | ||
| 22 | - | ||
| 23 | -@@ -159,6 +161,11 @@ def _apply_rotary_pos_emb_thd( | ||
| 24 | - | ||
| 25 | - cp_size = parallel_state.get_context_parallel_world_size() | ||
| 26 | - cp_rank = parallel_state.get_context_parallel_rank() | ||
| 27 | -+ batch_hdp_group = get_batch_hdp_group() | ||
| 28 | -+ if batch_hdp_group is not None: | ||
| 29 | -+ hdp_group = next(group for group in batch_hdp_group if cp_rank in group) | ||
| 30 | -+ cp_size = len(hdp_group) | ||
| 31 | -+ cp_rank = hdp_group.index(cp_rank) | ||
| 32 | - cu_seqlens = cu_seqlens // cp_size | ||
| 33 | - seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist() | ||
| 34 | - | ||
| 35 | --- | ||
| 36 | -2.50.1.windows.1 | ||
| 37 | - | ||