Fix various issues in comm replay including hashing, srcOrDst assignment, and trace file naming#148

Closed

TaekyungHeo wants to merge 15 commits intofacebookresearch:mainfrom TaekyungHeo:develop-rebase

+212-318

Commits on Aug 12, 2024

Deprecate support for basic and Kineto trace parsing
sanshang-nv
committed
Refator _parseExecutionTrace function
sanshang-nv
committed
Fix ValueError message to correctly refer to "trace version" instead of "trace format"
sanshang-nv
committed
Modify _parse_proc_group_info to add node_id to returned value
sanshang-nv
committed
Convert recorded rank id from local to global in _parse_comms_op_node
sanshang-nv
committed
Refactor _parse_proc_group_info for readability and maintainability
sanshang-nv
committed
Parsing comm arguments from attributes field and being used in _parse_comms_op_node
sanshang-nv
committed
For undecimal process group name, replace raising exeption with warning message and continue.
sanshang-nv
committed
commsTraceParser.py: remove unused tensorDtypeMap and _getTensorInfoFromPyTorchETEntry()
sanshang-nv
committed
Run black

TaekyungHeo
authored and
sanshang-nv
committed
Fix bug of hash to inSplit and outSplit list

sanshang-nv
authored and
TaekyungHeo
committed
Fix srcOrDst assignment for "reduce", "broadcast", "gather", "scatter" in runComms

sanshang-nv
authored and
TaekyungHeo
committed
Fix typo in regenerateTensors description: "temsors" to "tensors"

sanshang-nv
authored and
TaekyungHeo
committed
Disable warm-up by default and correct help text for do_warm_up toggle

sanshang-nv
authored and
TaekyungHeo
committed
Update trace file naming convention to "rank{rank}.json" in directory mode

sanshang-nv
authored and
TaekyungHeo
committed