ETRecord profiling views

Visualize a matching ETRecord and ETDump to explore runtime timings and captured VGF profiling. Other model formats are covered in Usage and Model adapters.

Run the visualization

Supply the .etrecord (or .bin) and its matching .etdump (or .etdp). Add a matching .pte if you also want its VGF compiler graph. The following command installs the required ExecuTorch extras for this run and opens the visualization:

uv run --with "mlia-neural-technology[executorch]" \
  --with "mlia-model-explorer[executorch]" \
  mlia check model.etrecord \
  --target-profile neural-technology \
  --performance \
  --profiling-data run.etdump \
  --companion-model model.pte \
  --output-dir mlia-etrecord-result \
  --model-explorer open

This workflow currently supports only a .pte companion containing a VGF delegate, with matching VGF profiling in the ETDump. Other companion formats and non-VGF companion views are not supported. The Model Explorer ExecuTorch extra includes the required ETRecord, PTE and VGF adapters.

Omit --companion-model when no companion view is needed. Captures containing only ordinary runtime events remain supported and show available source timings.

For a multi-method record, add --neural-technology-profiling-data.etrecord-method decode, replacing decode with the captured method. Single-method selection is automatic.

Use --model-explorer files to generate files without opening the browser. The standalone page is written under the output directory at mlia-output/model-explorer/bundle/index.html. Analysis results are saved in mlia-output/mlia-output.json. Companion input is also supported through the Python API's companion_model argument. Supplying a companion does not enable visualization; --model-explorer must be requested explicitly.

Explore the results

View	Contents
Source Model	Source operators and available runtime timings.
segment_cascade_chain	The captured NX hierarchy and its profiling.
ETRecord Chain Mapping	Chains associated with source operators; the initial detailed view when available.
ETDump Backend Trace	Compiler operations and their measured chain groups.
PTE - VGF Backend Hierarchy	The optional companion compiler graph with associated chain groups.

Select a mapped operator or group to navigate between related views. Some views or links may be absent when the capture does not provide the required relationships. Unmatched compiler operations remain visible; a missing link does not establish that an operation was not executed.

Measurements and limitations

• Runtime timings are in milliseconds; VGF counters are in cycles. Cycle counters describe captured activity, not elapsed latency.
• PTE and ETDump views show VGF profiling on chain and cascade groups. A group's measurement is not an independent timing for every operator it contains. Source associations may be incomplete or approximate.
• Colours use the same scale for the same entity kind across views. The viewer selects total_cycles when available, otherwise op_cycles.
• Parent and child groups may describe overlapping work. Do not sum measurements across hierarchy levels. Shared runtime timings, P50, minimum and maximum values are not additive; unavailable group aggregates are omitted.
• Independent measurements are combined within each recorded execution before summarizing repeated executions. Combining group averages requires matching execution coverage.
• Totals describe recorded observations. Missing measurements are not estimated, and a successful visualization does not prove a complete inference was captured. Use the reported execution coverage and warnings when interpreting summaries. Model cycle summaries are unavailable when VGF instructions or segments cover different recorded executions.
• Companion matching checks structural compatibility and assumes you supplied the matching PTE. It does not prove that this PTE produced the capture. Before generating the companion view, its digest is compared with the analyzed file's recorded digest; a changed file requires analysis again.

Supported VGF captures

VGF decoding supports statistics modes 0 and 1 for Mali-G1 and Mali-G2:

Device	Mode 0 words per task	Mode 1 words per task
Mali-G1	16	32
Mali-G2	16	64

A Mali-G2 mode-0 layout with 64 words per task is not supported. Declared counter-layout metadata must use schema version 1; unsupported or conflicting settings are rejected.

Recorded mode and device settings are used when available. Legacy schema-v1 VGF captures do not record these settings, so analysis defaults to mode 1 and Mali-G2 when they are missing. Check the reported assumptions and warnings: successful decoding does not verify that those defaults match the capture.

For counter layouts, aggregation rules and capture-completeness details, see Neural Technology ETRecord profiling. See Artifacts for the generated files. If an adapter is missing or incompatible, install the extra or version named in the error. For companion matching errors, check that all artifacts belong to the same deployed program and selected method.