aws / aws/sagemaker-python-sdk
Add ml.p5e.48xlarge to EFA instance lists in sagemaker-train and sagemaker-core
- 主要语言
- Python
- 星标
- 2.3k
- 派生
- 1.3k
- 平均合并
- 1 天 22 小时
- 30 天内合并 PR
- 35
描述
# Add ml.p5e.48xlarge to EFA instance lists in sagemaker-train and sagemaker-core
## Description
The `SM_EFA_NCCL_INSTANCES` and `SM_EFA_RDMA_INSTANCES` lists in the sagemaker-python-sdk are missing `ml.p5e.48xlarge`, causing NCCL hangs during distributed training initialization on P5e instances when using the SDK's container drivers.
Additionally, `ml.p5.48xlarge` is missing from `SM_EFA_RDMA_INSTANCES` (it's only in `SM_EFA_NCCL_INSTANCES`).
## Current State
```python
SM_EFA_NCCL_INSTANCES = [
"ml.g4dn.8xlarge",
"ml.g4dn.12xlarge",
"ml.g5.48xlarge",
"ml.p3dn.24xlarge",
"ml.p4d.24xlarge",
"ml.p4de.24xlarge",
"ml.p5.48xlarge",
"ml.trn1.32xlarge",
]
SM_EFA_RDMA_INSTANCES = [
"ml.p4d.24xlarge",
"ml.p4de.24xlarge",
"ml.trn1.32xlarge",
]
```
## Expected State
```python
SM_EFA_NCCL_INSTANCES = [
"ml.g4dn.8xlarge",
"ml.g4dn.12xlarge",
"ml.g5.48xlarge",
"ml.p3dn.24xlarge",
"ml.p4d.24xlarge",
"ml.p4de.24xlarge",
"ml.p5.48xlarge",
"ml.p5e.48xlarge", # ADD
"ml.trn1.32xlarge",
]
SM_EFA_RDMA_INSTANCES = [
"ml.p4d.24xlarge",
"ml.p4de.24xlarge",
"ml.p5.48xlarge", # ADD
"ml.p5e.48xlarge", # ADD
"ml.trn1.32xlarge",
]
```
## Impact
Without these entries, the SDK's container drivers don't set the required EFA environment variables (`FI_PROVIDER=efa`, `FI_EFA_USE_DEVICE_RDMA=1`, `RDMAV_FORK_SAFE=1`) for P5e instances, causing NCCL to hang during collective initialization in multi-node distributed training.
## Related
- sagemaker-training-toolkit issue: https://github.com/aws/sagemaker-training-toolkit/issues/240
- sagemaker-training-toolkit PR: https://github.com/aws/sagemaker-training-toolkit/pull/241
- P5e instances use EFA with RDMA support, same as P4d/P4de/P5
## Questions
1. Is there a specific process for testing EFA/instance-specific changes on actual hardware before merging?
2. Should integration tests be added for P5e EFA configuration, or are unit tests sufficient?
贡献指南
调研方向
首先定位 sagemaker-train 和 sagemaker-core 组件中的 SM_EFA_NCCL_INSTANCES 和 SM_EFA_RDMA_INSTANCES 定义。将两个列表与此 issue 中的预期状态进行比较,然后检查相关的单元测试覆盖率;完成的标准是两个 P5 实例都出现在相应的 EFA 列表中。
由索引模型根据 Issue 内容生成。
评估
- 技术栈
- aws, python
- 领域
- cloud, machine-learning
- Issue 类型
- 缺陷
- 难度
- 2/5
- 预计耗时
- 1-3 小时
- 活跃度
- 停滞
- 描述清晰度
- 描述清楚
- 新手友好度
- 58/100