facebookresearch / HolisticTraceAnalysis

A library to analyze PyTorch traces.

http://hta.readthedocs.io

facebookresearch/HolisticTraceAnalysis Issues

Critical path analysis - matching the kernel related to cudaEventRecord with stream
Updated 2 months ago1
Simulating critical paths with optimization heuristics
Updated 2 months ago7
Updated timeline analysis
Closed 2 months ago
New Trace Callstack processing out of order
Updated 3 months ago
min(arg) is an empty sequence -> issues creating analyzer
Closed 3 months ago10
HTA expects rank to be specified in a trace file
Closed 3 months ago3
single GPU support
Closed 4 months ago1
How to plot timeline with the label added by torch.profiler.record_function?
Updated 4 months ago
Trace Event Filter
Closed 5 months ago
Pandas 2; Value error (Pandas 1.X works fine)
Closed 5 months ago2
Replace Call Stack Graph Implementation With Its Enhanced Version
Closed 6 months ago2
Update analyzers to use the new trace call stack graph implementation
Closed 6 months ago1
Introduce Memory Analyzer feature
Closed 7 months ago7
Bug - Incorrect variable name in Trace.load_traces() method
Closed 7 months ago
When parsing the trace files the last profiling step is not loaded
Closed 7 months ago2
The regenerated json file is garbled
Closed 7 months ago2
Expose more event features in trace representation
Closed 7 months ago1
CUPTI Counters analysis does not work when counters are float
Closed 9 months ago
Support traces collected from H100
Closed 9 months ago
Execution Trace Correlation Support
Updated 10 months ago
comm_comp_overlap_df = analyzer.get_comm_comp_overlap() error
Closed a year ago2
Add copyright header to cupti_profiler_demo.py
Closed a year ago
Unit of time specified in gpu_kernel_breakdown is incorrect
Closed a year ago
CUPTI Profiler Counter analysis in HTA
Closed a year ago3
Optimizing Trace load and parsing functionality
Updated a year ago3
Put back circleci config
Closed a year ago
Optimize the multiprocessing strategy when parsing multiple traces
Closed a year ago
Change log level for get_memory_usage function.
Closed a year ago
Parallelize `create_rank_to_trace_dict`
Closed a year ago
Update ShipIt sync
Closed a year ago
Memory bandwidth shows negative values
Closed a year ago
pd.concat fails when dataframe is empty
Closed a year ago
Maintenance
Closed a year ago
Implement graceful exit when rank info is not available in the trace
Closed a year ago
Dictionary lookups fail when runtime kernels are absent in the trace
Closed a year ago2
Release requirements
Closed a year ago
Generic kernel breakdown and overlap analysis
Updated a year ago1
Filter out gpu_user_annotation events at parsing time
Closed a year ago
Add code coverage
Closed a year ago
Bug fixes
Closed a year ago