已关闭
[ignore] [test-ci] Good case #71
hanboyou创建于 2025年12月16日关闭于 2025年12月24日
[ignore] [test-ci] Good case #71
已关闭
共 3 个文件变更+9-9
| @@ -5,7 +5,7 @@ | |||
| 5 | - [2025/10] DeepSeek-R1、Qwen2.5模型样例首次上线。 | 5 | - [2025/10] DeepSeek-R1、Qwen2.5模型样例首次上线。 |
| 6 | 6 | ||
| 7 | ## 🎉概述 | 7 | ## 🎉概述 |
| 8 | -cann-recipes-train仓库旨在针对LLM与多模态模型训练业务中的典型模型、算法，提供基于CANN平台的优化样例，方便开发者简单、快速、高效地使用CANN平台进行模型训练。 | 8 | +cann-recipes-train仓库旨在针对LLM与多模态模型训练业务中的典型模型、算法，提供基于CANN平台的优化样例，方便开发者简单、快速、高效地使用CANN平台进行模型训练。j |
| 9 | 9 | ||
| 10 | 10 | ||
| 11 | ## ✨实践列表 | 11 | ## ✨实践列表 |
| @@ -159,7 +159,7 @@ bash ray_start_npu.sh TRAIN_SCRIPT ENV_SCRIPT | |||
| 159 | ### 文件说明 | 159 | ### 文件说明 |
| 160 | |上级目录|文件路径|说明| | 160 | |上级目录|文件路径|说明| |
| 161 | |-------|--------|--------| | 161 | |-------|--------|--------| |
| 162 | -|megatron|[0001-megatron-bugfix-state_ten-verification.patch](patches/megatron/0001-megatron-bugfix-state_ten-verification.patch)|在处理优化器状态时新增空值判断，避免因空值导致的运行异常| | 162 | +|megatron|[0001-megatron-bugfix-state_ten_verification.patch](patches/megatron/0001-megatron-bugfix-state_ten_verification.patch)|在处理优化器状态时新增空值判断，避免因空值导致的运行异常| |
| 163 | |megatron|[0002-megatron-feature-enable_hdp.patch](patches/megatron/0002-megatron-feature-enable_hdp.patch)|在ROPE中增加HDP相关处理逻辑，`USE_HDP`开启时，使能HDP功能| | 163 | |megatron|[0002-megatron-feature-enable_hdp.patch](patches/megatron/0002-megatron-feature-enable_hdp.patch)|在ROPE中增加HDP相关处理逻辑，`USE_HDP`开启时，使能HDP功能| |
| 164 | |mindspeed|[0001-mindspeed-bugfix-builder.patch](patches/mindspeed/0001-mindspeed-bugfix-builder.patch)|兼容openeuler24.03版本下编译头文件缺失| | 164 | |mindspeed|[0001-mindspeed-bugfix-builder.patch](patches/mindspeed/0001-mindspeed-bugfix-builder.patch)|兼容openeuler24.03版本下编译头文件缺失| |
| 165 | |mindspeed|[0002-mindspeed-feature-enable_hdp.patch](patches/mindspeed/0002-mindspeed-feature-enable_hdp.patch)|在Ring Attention中增加HDP相关处理逻辑，`USE_HDP`开启时，使能HDP功能| | 165 | |mindspeed|[0002-mindspeed-feature-enable_hdp.patch](patches/mindspeed/0002-mindspeed-feature-enable_hdp.patch)|在Ring Attention中增加HDP相关处理逻辑，`USE_HDP`开启时，使能HDP功能| |
| @@ -173,7 +173,7 @@ bash ray_start_npu.sh TRAIN_SCRIPT ENV_SCRIPT | |||
| 173 | |verl|[0008-verl-bugfix-enable_compile.patch](patches/verl/0008-verl-bugfix-enable_compile.patch)|NPU上MindSpeed训练框架会无效化torch.compile规避训练侧的compile失败，在推理时开启compile| | 173 | |verl|[0008-verl-bugfix-enable_compile.patch](patches/verl/0008-verl-bugfix-enable_compile.patch)|NPU上MindSpeed训练框架会无效化torch.compile规避训练侧的compile失败，在推理时开启compile| |
| 174 | |verl|[0009-verl-feature-support_EPLB.patch](patches/verl/0009-verl-feature-support_EPLB.patch)|`VLLM_ENABLE_EPLB`开启时，使能推理的EPLB| | 174 | |verl|[0009-verl-feature-support_EPLB.patch](patches/verl/0009-verl-feature-support_EPLB.patch)|`VLLM_ENABLE_EPLB`开启时，使能推理的EPLB| |
| 175 | |verl|[0010-verl-feature-enable_hdp.patch](patches/verl/0010-verl-feature-enable_hdp.patch)|`USE_HDP`开启时，使能HDP功能| | 175 | |verl|[0010-verl-feature-enable_hdp.patch](patches/verl/0010-verl-feature-enable_hdp.patch)|`USE_HDP`开启时，使能HDP功能| |
| 176 | -|verl|[0011-verl-feature-enable-rollout-rebalance.patch](patches/verl/0011-verl-feature-enable-rollout-rebalance.patch)|`ROLLOUT_REBALANCE_ENABLE`开启时，使能Rollout Rebalance功能，详细说明可参考[RL On-Policy 推理场景的序列级均衡调度引擎](patches/verl/features/rollout_optimize/README.md)| | 176 | +|verl|[0011-verl-feature-enable_rollout_rebalance.patch](patches/verl/0011-verl-feature-enable_rollout_rebalance.patch)|`ROLLOUT_REBALANCE_ENABLE`开启时，使能Rollout Rebalance功能，详细说明可参考[RL On-Policy 推理场景的序列级均衡调度引擎](patches/verl/features/rollout_optimize/README.md)| |
| 177 | |verl|[0012-verl-feature-enabled_sam_spec_decode.patch](patches/verl/0012-verl-feature-enabled_sam_spec_decode.patch)|SAM投机推理适配verl框架：允许通过脚本配置项开关SAM投机推理并配置相关参数| | 177 | |verl|[0012-verl-feature-enabled_sam_spec_decode.patch](patches/verl/0012-verl-feature-enabled_sam_spec_decode.patch)|SAM投机推理适配verl框架：允许通过脚本配置项开关SAM投机推理并配置相关参数| |
| 178 | |verl|[0013-verl-bugfix-dataProto_concat.patch](patches/verl/0013-verl-bugfix-dataProto_concat.patch)|合并DataProto数据时，避免因不同节点的`data['timing']['generate_sequences']`存在细微差异导致报错| | 178 | |verl|[0013-verl-bugfix-dataProto_concat.patch](patches/verl/0013-verl-bugfix-dataProto_concat.patch)|合并DataProto数据时，避免因不同节点的`data['timing']['generate_sequences']`存在细微差异导致报错| |
| 179 | |verl|[0014-verl-feature-dapo_data_rebalance.patch](patches/verl/0014-verl-feature-dapo_data_rebalance.patch)|`data_rebalance` DAPO算法适配| | 179 | |verl|[0014-verl-feature-dapo_data_rebalance.patch](patches/verl/0014-verl-feature-dapo_data_rebalance.patch)|`data_rebalance` DAPO算法适配| |
| @@ -186,13 +186,13 @@ bash ray_start_npu.sh TRAIN_SCRIPT ENV_SCRIPT | |||
| 186 | |vllm_ascend|[0003-vllm_ascend-feature-enable_zero_tp_to_ep.patch](patches/vllm_ascend/0003-vllm_ascend-feature-enable_zero_tp_to_ep.patch)|零冗余TP转EP通信方案，将o_proj的AllReduce算子替换为ReduceScatter算子，减少冗余通信| | 186 | |vllm_ascend|[0003-vllm_ascend-feature-enable_zero_tp_to_ep.patch](patches/vllm_ascend/0003-vllm_ascend-feature-enable_zero_tp_to_ep.patch)|零冗余TP转EP通信方案，将o_proj的AllReduce算子替换为ReduceScatter算子，减少冗余通信| |
| 187 | |vllm_ascend|[0004-vllm_ascend-feature-dummy_run_load_balance.patch](patches/vllm_ascend/0004-vllm_ascend-feature-dummy_run_load_balance.patch)| 在dummy_run阶段强制负载均衡，优化内存分配| | 187 | |vllm_ascend|[0004-vllm_ascend-feature-dummy_run_load_balance.patch](patches/vllm_ascend/0004-vllm_ascend-feature-dummy_run_load_balance.patch)| 在dummy_run阶段强制负载均衡，优化内存分配| |
| 188 | |vllm_ascend|[0005-vllm_ascend-feature-support_EPLB.patch](patches/vllm_ascend/0005-vllm_ascend-feature-support_EPLB.patch) | `VLLM_ENABLE_EPLB`开启时，使能推理的EPLB| | 188 | |vllm_ascend|[0005-vllm_ascend-feature-support_EPLB.patch](patches/vllm_ascend/0005-vllm_ascend-feature-support_EPLB.patch) | `VLLM_ENABLE_EPLB`开启时，使能推理的EPLB| |
| 189 | -|vllm_ascend|[0006-vllm_ascend-feature-chunk-moe-eager.patch](patches/vllm_ascend/0006-vllm_ascend-feature-chunk-moe-eager.patch) | 针对MoE计算场景分块处理优化，解决prefill阶段可能引起的峰值内存过高，单算子模式实现| | 189 | +|vllm_ascend|[0006-vllm_ascend-feature-chunk_moe_eager.patch](patches/vllm_ascend/0006-vllm_ascend-feature-chunk_moe_eager.patch) | 针对MoE计算场景分块处理优化，解决prefill阶段可能引起的峰值内存过高，单算子模式实现| |
| 190 | |vllm_ascend|[0007-vllm_ascend-bugfix-disabled_fia.patch](patches/vllm_ascend/0007-vllm_ascend-bugfix-disabled-fia.patch)|取消vllm_ascend的特殊逻辑，投机推理的算子从FusedInferAttention算子回调为PagedAttention算子，避免性能劣化| | 190 | |vllm_ascend|[0007-vllm_ascend-bugfix-disabled_fia.patch](patches/vllm_ascend/0007-vllm_ascend-bugfix-disabled-fia.patch)|取消vllm_ascend的特殊逻辑，投机推理的算子从FusedInferAttention算子回调为PagedAttention算子，避免性能劣化| |
| 191 | |vllm_ascend|[0008-vllm_ascend-feature-bs_threshold_for_spec_decode.patch](patches/vllm_ascend/0008-vllm_ascend-feature-bs_threshold_for_spec_decode.patch)|增加投机推理特性自动开关，解决投机推理特性在batch_size过高时性能劣化的问题| | 191 | |vllm_ascend|[0008-vllm_ascend-feature-bs_threshold_for_spec_decode.patch](patches/vllm_ascend/0008-vllm_ascend-feature-bs_threshold_for_spec_decode.patch)|增加投机推理特性自动开关，解决投机推理特性在batch_size过高时性能劣化的问题| |
| 192 | |vllm_ascend|[0009-vllm_ascend-feature-rewrote_rejection_sampler.patch](patches/vllm_ascend/0009-vllm_ascend-feature-rewrote-rejection-sampler.patch)|重写vllm_ascend的rejectionsampler实现，优化性能| | 192 | |vllm_ascend|[0009-vllm_ascend-feature-rewrote_rejection_sampler.patch](patches/vllm_ascend/0009-vllm_ascend-feature-rewrote-rejection-sampler.patch)|重写vllm_ascend的rejectionsampler实现，优化性能| |
| 193 | |vllm_ascend|[0010-vllm_ascend-feature-enabled_sam_spec_decode.patch](patches/vllm_ascend/0010-vllm_ascend-feature-enabled_sam_spec_decode.patch)|SAM投机推理适配vllm_ascend框架：适配vllm框架的改动| | 193 | |vllm_ascend|[0010-vllm_ascend-feature-enabled_sam_spec_decode.patch](patches/vllm_ascend/0010-vllm_ascend-feature-enabled_sam_spec_decode.patch)|SAM投机推理适配vllm_ascend框架：适配vllm框架的改动| |
| 194 | |vllm_ascend|[spec_decode/sam_proposer.py](patches/vllm_ascend/spec_decode/sam_proposer.py)|SAM投机推理适配vllm_ascend框架：实现`SAMProposer`类，作为vllm调用SAM投机推理能力的接口| | 194 | |vllm_ascend|[spec_decode/sam_proposer.py](patches/vllm_ascend/spec_decode/sam_proposer.py)|SAM投机推理适配vllm_ascend框架：实现`SAMProposer`类，作为vllm调用SAM投机推理能力的接口| |
| 195 | -|patches|[0001-feature-model-converter.patch](patches/0001-feature-model-converter.patch) | 新增`USE_ALLTOALL_OVERLAP`开启时hf2mcore权重转换逻辑| | 195 | +|patches|[0001-feature-model_converter.patch](patches/0001-feature-model_converter.patch) | 新增`USE_ALLTOALL_OVERLAP`开启时hf2mcore权重转换逻辑| |
| 196 | 196 | ||
| 197 | ### 手动准备环境 | 197 | ### 手动准备环境 |
| 198 | 198 | ||
| @@ -6,13 +6,13 @@ Subject: [PATCH] Fix DataProto concat bug | |||
| 6 | DataProto concat bug when timing generate_sequences is sightly different | 6 | DataProto concat bug when timing generate_sequences is sightly different |
| 7 | https://github.com/volcengine/verl/pull/4030 | 7 | https://github.com/volcengine/verl/pull/4030 |
| 8 | --- | 8 | --- |
| 9 | - rl_train/qwen3/verl/protocol.py | 25 ++++++++++++++++++++++++- | 9 | + /qwen3/verl/protocol.py | 25 ++++++++++++++++++++++++- |
| 10 | 1 file changed, 24 insertions(+), 1 deletion(-) | 10 | 1 file changed, 24 insertions(+), 1 deletion(-) |
| 11 | 11 | ||
| 12 | -diff --git a/rl_train/qwen3/verl/protocol.py b/rl_train/qwen3/verl/protocol.py | 12 | +diff --git a/llm_rl/qwen3/verl/protocol.py b/llm_rl/qwen3/verl/protocol.py |
| 13 | index e0b1aff..fda1e7c 100644 | 13 | index e0b1aff..fda1e7c 100644 |
| 14 | ---- a/rl_train/qwen3/verl/protocol.py | 14 | +--- a/llm_rl/qwen3/verl/protocol.py |
| 15 | -+++ b/rl_train/qwen3/verl/protocol.py | 15 | ++++ b/llm_rl/qwen3/verl/protocol.py |
| 16 | class DataProto: | 16 | class DataProto: |
| 17 | for key, val in non_tensor_batch.items(): | 17 | for key, val in non_tensor_batch.items(): |
| 18 | non_tensor_batch[key] = np.concatenate(val, axis=0) | 18 | non_tensor_batch[key] = np.concatenate(val, axis=0) |