The traces provided in this directory (tests/traces/h100 and tests/traces/mi300) are for testing, validation, and educational purposes only.
They are derived from open-source models and are intended to help contributors and users verify TraceLens functionality, experiment with analysis workflows, and learn about performance behavior across hardware.
These traces do not represent:
- Official or benchmarked performance numbers.
- Any internal, production, or customer environment.
- Validated hardware configurations, tuning parameters, or training setups.
Performance metrics observed from these traces may differ substantially from real workloads. They should not be interpreted as indicative of actual device performance or official results.
| Directory | JAX version | Description |
|---|---|---|
jax_conv_minimal_legacy |
~0.6 | Older trace; .hlo_proto.pb sidecars are generated at runtime by xprof (ignored via .gitignore). Full perf model (FLOPS, bytes, shapes) works when sidecars are present. Used by test_jax_conv_analysis.py. |
jax_conv_minimal_08 |
0.8 | Comparable minimal 3D conv trace (same params). For E2E smoke tests; perf model metadata limited until #425. |
Both use the same convolution parameters (16→5120 channels, 1×2×2 kernel, stride 1×2×2, bf16). Generated via jax-minimal/traces/generate_conv_trace.py.
Python function events have been removed from these traces to reduce test data size while maintaining identical NCCL analysis results.