-
Notifications
You must be signed in to change notification settings - Fork 3
Open
Description
New Features
- [Roadmap] StreamingDataLoader for task-separated RL post-training TransferQueue/TransferQueue#85
- Implement
check_data_consumption_status&check_data_production_statusforTransferQueueController[Feat]: add check_data_production_status and check_consumption_status and support Polling get metadata TransferQueue/TransferQueue#157
Backend Related
- Support Ray RDT backend [Feat]: Add RayStorageClient to support the RDT feature of Ray TransferQueue/TransferQueue#167
- Support MoonCakeStore backend [Feat] Support Mooncake Store backend TransferQueue/TransferQueue#162
Adaption Works
- Support verl multi-modal scenario https://github.com/TransferQueue/verl/tree/agentloop_tq
- Merge with verl main code entry
- Support verl router replay https://github.com/TransferQueue/verl/tree/new_router_replay
Performance
- Optimize data spliting & gathering opterations (
get_transfer_data,StorageUnitData.put_data/get_data, etc.) @jianjunzhong [Refactor] Simplify the put/get logic of AsyncSimpleStorageManager TransferQueue/TransferQueue#151 - Optimize the socket event loop by splitting into to multiple threads, one for each type of request
- Strengthen
TransferQueueControllerconcurrency - Improve zero-copy performance [BREAKING][Perf] Optimize TransferQueue memory usage TransferQueue/TransferQueue#163 [4/4][serialization] Perf: support non-jagged nested tensor for zero-copy serialization in SimpleUnit backend TransferQueue/TransferQueue#121
User Interface
- Unified environmental variable management
Debug
- Performance metric [Perf] Add performance monitor tool & optimize log output TransferQueue/TransferQueue#132
- Command line debug interface
- Cleanup redundant logs [Misc] Clean up redundant logs & unify log format TransferQueue/TransferQueue#160 [BREAKING][Feat] Add fine-grained clear method & optimize log TransferQueue/TransferQueue#159 [Perf] Add performance monitor tool & optimize log output TransferQueue/TransferQueue#132
- Better metadata print [Feat] Add select_samples & improve print readability for BatchMeta TransferQueue/TransferQueue#133
Minor Issues & Code Clean
- Provide a general dynamic socket function for both Client & Storage
- Make dtypes & shapes compulsory for
update_production_status - Optimize
_merge_tensors_to_tensordictto serve as a general function since similar codes may scatter everywhere inside TQ - Split the zero copy optimization from zmq_utils.py to serial_utils.py providing a general serialization util for tensordict (both device-side and cpu-side) [Refactor] Provide a general zero-copy serial util TransferQueue/TransferQueue#140
Documentary
- Build doc web site
- Provide tutorials [Tutorial] Provide tutorials for TransferQueue usage TransferQueue/TransferQueue#141
Metadata
Metadata
Assignees
Labels
No labels