You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
[optim] Optimize the exception message for Yuanrong when storage strategy does not support a certain type of data (#120)
## Main changes
- Modify the exception message to make it more understandable.
- Add a solution for the "Cannot retrieve stored data" error to FAQ.
- Add two environmental variables `DS_D2H_MEMCPY_POLICY` and
`DS_H2D_MEMCPY_POLICY` to datasystem workers
---------
Signed-off-by: dpj135 <958208521@qq.com>
Copy file name to clipboardExpand all lines: docs/storage_backends/openyuanrong_datasystem.md
+16-1Lines changed: 16 additions & 1 deletion
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -420,7 +420,7 @@ ps aux | grep datasystem_worker
420
420
dscli stop --worker_address <IP>:<PORT>
421
421
422
422
# Force cleanup (use with caution)
423
-
pkill -f datasystem_worker
423
+
pkill -9 -f datasystem_worker
424
424
```
425
425
426
426
### Multi-Process Initialization
@@ -463,6 +463,21 @@ RuntimeError: code: [Out of memory], msg: [Shared memory no space in arena: ...]
463
463
464
464
Solution: Increase `--shared_memory_size_mb` in `worker_args`, or reduce the data volume being cached.
465
465
466
+
### "Cannot retrieve stored data" Error on get/clear
467
+
468
+
If you encounter an error like:
469
+
```
470
+
ValueError: Cannot retrieve stored data because the backend that originally stored it is unavailable in the current process or node. Please check that the configuration and NPU resource availability are consistent across all processes and nodes.
471
+
```
472
+
473
+
This occurs when `kv_batch_get` cannot find the storage backend that originally handled the data. The most common cause is a mismatch between the process that originally `put` the data and the process performing `get`, such as:
474
+
475
+
- Different `enable_yr_npu_transport` settings across processes or nodes (e.g., `true` vs `false`).
476
+
- NPU hardware or CANN/torch-npu unavailable on the `get` process or node, even though the configuration is identical.
477
+
- When running inside Ray actors, the actor may not be assigned NPU resources (e.g., missing `"NPU": 1` in `.options(resources=...)`), preventing the NPU transport backend from initializing.
478
+
479
+
Solution: Ensure that all processes and nodes use the same TransferQueue configuration and have consistent NPU resource availability. When using Ray actors, make sure NPU resources are properly allocated via `.options(resources={"NPU": 1})`.
0 commit comments