Analyze TPU pretraining profiles to determine HBM peak occupancy and buffer allocation.
npx skills add https://github.com/primatrix/skills --skill memory-profile
Please help me install this Agent Skill. Skill: memory-profile Source: https://github.com/primatrix/skills/tree/main/plugins/tpu-perf/skills/memory-profile Command: npx skills add https://github.com/primatrix/skills --skill memory-profile
This Skill provides deep insights into the HBM (High Bandwidth Memory) occupancy and memory allocation during TPU pretraining, helping to optimize resource usage and identify bottlenecks.
Analyze the HBM occupancy for the profile directory 'tpu_profile'.
High-intent search queries and answers about installing and using this skill.
TPU pretraining HBM occupancy is analyzed by parsing HLO buffer assignments from '.hlo_proto.pb' files and runtime allocator events from '.xplane.pb' files to determine peak memory usage and identify contributing buffers.
High HBM usage during TPU pretraining is caused by specific HLO instructions and buffer allocations that contribute to the static memory peak. Tracking runtime buffer lifetimes and sizes helps pinpoint these memory bottlenecks.
Track buffer allocation in TPU runtime memory by parsing runtime allocator events captured within '.xplane.pb' profile files to monitor the lifetime and size of buffers allocated and deallocated during the trace.
Yes, HLO proto files are required to analyze TPU memory allocation. The analysis specifically requires HLO buffer assignment data from '*.hlo_proto.pb' files to identify the HLO instructions contributing to peak HBM occupancy.
The best way to find HLO instructions contributing to HBM peak usage is to cross-reference static HLO buffer assignment data from '.hlo_proto.pb' files with runtime allocator events from '.xplane.pb' profiles to isolate the specific buffers at the peak.
No, optimizing TPU resource usage fully requires runtime allocator events from '*.xplane.pb' files. Without these events, the analysis cannot track the dynamic lifetime and size of buffers allocated and deallocated during the runtime trace.
primatrixAnalyze communication primitives in TPU pretraining profiles for bandwidth utilization and overlap.
ByteDance-SeedProfile deep learning training runs by parsing Chrome traces and memory snapshots.
novoleiAutomate PDF text and table extraction with Python libraries.
yeismeProfile CPU and memory metrics to identify software bottlenecks.
