已合并
mindspeed训练后端特性使用指南 #749
sz创建于 2025年11月19日
mindspeed训练后端特性使用指南 #749
已合并
共 2 个文件变更+39-0
| @@ -2,6 +2,8 @@ approvers: | |||
| 2 | - fengliangjun | 2 | - fengliangjun |
| 3 | - zhoubeirong | 3 | - zhoubeirong |
| 4 | - elvinp | 4 | - elvinp |
| 5 | +- wucong0509 | ||
| 6 | +- zengshu19991010 | ||
| 5 | reviewers: | 7 | reviewers: |
| 6 | - fengliangjun | 8 | - fengliangjun |
| 7 | - QsingHuan | 9 | - QsingHuan |
| @@ -0,0 +1,37 @@ | |||
| 1 | +# Verl 使用 MindSpeed 训练后端 | ||
| 2 | + | ||
| 3 | +## 使能 MindSpeed 后端 | ||
| 4 | + | ||
| 5 | +确认模型对应的 `strategy` 配置为 `megatron`，例如 `actor_rollout_ref.actor.strategy=megatron`，可以在 shell 脚本中或者 config 配置文档中设置。 | ||
| 6 | + | ||
| 7 | +MindSpeed 自定义入参可通过 `override_transformer_config` 参数传入，例如对 `actor` 模型开启 FA 特性可使用 `+actor_rollout_ref.actor.megatron.override_transformer_config.use_flash_attn=True` | ||
| 8 | + | ||
| 9 | +## 特性支持列表 | ||
| 10 | +| 特性名称 | 配置参数 | 状态 | | ||
| 11 | +|-------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------|---------| | ||
| 12 | +| TP | actor_rollout_ref.actor.megatron.tensor_model_parallel_size<br>actor_rollout_ref.ref.megatron.tensor_model_parallel_size | Preview | | ||
| 13 | +| PP | actor_rollout_ref.actor.megatron.pipeline_model_parallel_size<br>actor_rollout_ref.ref.megatron.pipeline_model_parallel_size | Preview | | ||
| 14 | +| EP | actor_rollout_ref.actor.megatron.expert_model_parallel_size<br>actor_rollout_ref.ref.megatron.expert_model_parallel_size | Preview | | ||
| 15 | +| ETP | actor_rollout_ref.actor.megatron.expert_tensor_parallel_size<br>actor_rollout_ref.ref.megatron.expert_tensor_parallel_size | Preview | | ||
| 16 | +| SP | actor_rollout_ref.actor.megatron.override_transformer_config.sequence_parallel | Preview | | ||
| 17 | +| 分布式优化器 | actor_rollout_ref.actor.megatron.override_transformer_config.use_distributed_optimizer | Preview | | ||
| 18 | +|swap-optimizer |actor_rollout_ref.actor.megatron.override_transformer_config.swap_optimizer | Preview | | ||
| 19 | +|自定义空操作层 |actor_rollout_ref.actor.megatron.override_transformer_config.noop_layers<br>actor_rollout_ref.actor.megatron.override_transformer_config.num_layers | Preview | | ||
| 20 | +| 重计算 | actor_rollout_ref.actor.megatron.override_transformer_config.recompute_method<br>actor_rollout_ref.actor.megatron.override_transformer_config.recompute_granularity<br>actor_rollout_ref.actor.megatron.override_transformer_config.recompute_num_layers | Preview | | ||
| 21 | +| CP | actor_rollout_ref.actor.megatron.context_parallel_size<br>actor_rollout_ref.actor.megatron.override_transformer_config.context_parallel_size | Preview | | ||
| 22 | +| MoE alltoall dispatcher | actor_rollout_ref.actor.megatron.override_transformer_config.moe_permutation_async_comm | Preview | | ||
| 23 | +|Multi-head Latent Attention| actor_rollout_ref.actor.megatron.override_transformer_config.multi_head_latent_attention | Preview | | ||
| 24 | + | ||
| 25 | +## 融合算子支持列表 | ||
| 26 | +| 特性名称 | 配置参数 | 状态 | | ||
| 27 | +|-------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------|---------| | ||
| 28 | +| fusion attention| actor_rollout_ref.actor.megatron.override_transformer_config.use_flash_attn | Preview | | ||
| 29 | +| rms norm | actor_rollout_ref.actor.megatron.override_transformer_config.use_fused_rmsnorm | Preview | | ||
| 30 | +| swiglu | actor_rollout_ref.actor.megatron.override_transformer_config.use_fused_swiglu | Preview | | ||
| 31 | +| permute_unpermute |actor_rollout_ref.actor.megatron.override_transformer_config.fused_permute_and_unpermute | Preview | | ||
| 32 | +|gemm | actor_rollout_ref.actor.megatron.override_transformer_config.moe_grouped_gemm | Preview | | ||
| 33 | +|mc2 |actor_rollout_ref.actor.megatron.override_transformer_config.use_ascend_mc2 | Preview | | ||
| 34 | + | ||
| 35 | + | ||
| 36 | + | ||
| 37 | +注："Preview"状态表示预览非正式发布版本，"Released"状态表示正式发布版本，"Dev"状态表示正在开发中。 | ||