Description
This issue summarizes documentation improvements identified from a reader's perspective. The issues are organized into three categories: missing sections, confusing content, and areas needing more examples.
Missing Sections
-
Tutorial/End-to-End Examples
- Missing practical application examples (distributed training, model inference, etc.)
- No complete walkthrough of real-world usage scenarios
-
Performance Benchmark Section
- No performance comparison data (HCCL vs NCCL, YR transport vs regular transport)
- Missing latency/throughput benchmarks for different operations
-
Migration Guide
- No guide for migrating from other backends (e.g., NCCL) to HCCL
- Missing transition path for existing Ray users
-
Version Compatibility Matrix
- Unclear version correspondence between ray-ascend ↔ Ray ↔ CANN ↔ torch-npu
- "Same version as ray-ascend" is ambiguous
-
Cluster Deployment Guide
- Missing multi-node environment setup and configuration
- No instructions for cluster networking configuration
-
Quickstart with Working Example
- The current Quick Start only has a pip install command
- Needs a minimal runnable example to verify installation
Confusing Content
| Location |
Issue |
user_guide/index.md |
"Ray: Same version as ray-ascend" — unclear what this means exactly, how version numbers correspond |
installation.md |
Only provides Docker pull commands, no instructions on how to start container, mount data, or configure devices |
hccl_collective.md |
"Using Ray's Collective API" example is incomplete, missing code at # Create and setup group... |
best_practices.md |
"CPU tensors use a packed binary format" — what is this format? Do users need to understand it? |
| Various pages |
Project name inconsistency: "Ray-Ascend" vs "ray-ascend" |
installation.md |
No explanation of YR_DS_WORKER_HOST/PORT default values or cluster configuration |
installation.md |
No post-installation verification step or script |
Areas Needing More Examples
| Feature |
Missing Examples |
| HCCL |
Complete examples for allgather, reduce, reducescatter, send/recv, barrier |
| HCCL |
Examples for different reduce operations (PRODUCT, MAX, MIN) |
| YR Transport |
Multi-node transport examples (currently only single machine) |
| YR Transport |
Mixed CPU/NPU tensor transport examples |
| Combined Usage |
Complete application example using both HCCL and YR Transport together |
Additional Suggestions
- Chinese Documentation - Consider providing bilingual (English/Chinese) documentation since many target users are Chinese speakers
- Glossary - Add more detailed explanations for terms like HCCL, YuanRong, CANN when first mentioned
- Terminology Consistency - Standardize on either "Ray-Ascend" or "ray-ascend" throughout documentation
Related Link
Description
This issue summarizes documentation improvements identified from a reader's perspective. The issues are organized into three categories: missing sections, confusing content, and areas needing more examples.
Missing Sections
Tutorial/End-to-End Examples
Performance Benchmark Section
Migration Guide
Version Compatibility Matrix
Cluster Deployment Guide
Quickstart with Working Example
Confusing Content
user_guide/index.mdinstallation.mdhccl_collective.md# Create and setup group...best_practices.mdinstallation.mdinstallation.mdAreas Needing More Examples
allgather,reduce,reducescatter,send/recv,barrierAdditional Suggestions
Related Link