fix(profiling): make TraceLens and dynolog survive real GPU traces

[Cemberk]

Summary

#170 proved the TraceLens and dynolog tools work without a GPU. Running them on
an actual AMD GPU node against real rocprofv3 and dynolog captures proved they
did not work with one. This is the fallout.

• rocprofv3 JSON is not valid UTF-8. rocprofv3 copies HIP API
  const char * arguments into its JSON verbatim, so an argument that does not
  point at a string leaves raw bytes behind. TraceLens loads traces with orjson,
  which rejects the whole document, so twelve stray bytes cost the entire report
  on a 295 MB trace. The analyzer now detects undecodable bytes incrementally
  (no 295 MB read into memory) and analyzes a sanitized copy, leaving the
  original trace untouched.
• The dynolog trigger was arguing with a flag that does not exist.
  dyno gputrace has no --fail-on-no-process in the pinned v0.5.0, and exits 0
  whether or not it matched a process. Every attempt therefore looked like "no
  PyTorch process registered yet", and the trigger retried a request dyno had
  refused, twelve times, before giving up with a misleading diagnosis. It now
  reads processesMatched out of the response, and stops immediately when dyno
  rejected the request rather than when the workload is merely slow to register.
• The multi-rank collective report matched traces it could not use. It
  globbed the whole run directory and assumed every PyTorch trace carried its
  rank in its filename. Traces captured on demand are named after the process
  id, so TraceLens matched three files and rejected all of them. The report is
  now built only from rank-labelled traces, from a glob scoped to their own
  directory, and is skipped with an explanation when no trace carries a rank.
• --python was not actually isolating anything. _build_command fell back
  to searching PATH for TraceLens console scripts, which defeats the isolation
  that flag exists to provide: TraceLens pins protobuf and xprof and lives in a
  venv of its own.
• pip install -e . failed outright on this branch. The tracelens extra
  installs from git, and hatchling rejects direct references unless they are
  allowed explicitly. Nobody could install the branch to try any of the above.

Test plan

Every item above has a regression test that fails without its fix and needs no
GPU, so CI covers them. On the GPU node (banff-cyxtera-s83-5):

• tests/e2e/test_tracelens_dummy_pipeline.py + test_dynolog_dummy_pipeline.py — 35 passed
• tests/unit/test_tracelens_analyze.py + test_tracelens_report.py + tests/integration/test_tracelens_tools_config.py — 66 passed
• tests/e2e/test_tracelens_workflows.py (GPU-gated) — 9 passed, 2 skipped
• Real rocprofv3_lightweight + tracelens run end to end: the analyzer
  reported the trace as not valid UTF-8, sanitized it, and produced a
  SUCCESS row from the real TraceLens_generate_perf_report_rocprof
• Real torch_profiler_dynolog capture, then TraceLens on the resulting
  Kineto trace
• The whole tracelens or dynolog or profil surface: 127 passed. The three
  test_profiling_workflows.py failures (rpd, gpu_info_power,
  gpu_info_vram) are pre-existing and unrelated — they fail identically at
  ec4de0b, the commit before any TraceLens work. rpd dies inside the
  container on a missing librocm_sysdeps_sqlite3.so, and the gpu_info
  tools fail in gpu_info_pre.sh.

Note

Stacked on #170, so it targets cem/tracelens-dummy-tests.

[Copilot]

Pull request overview

This PR hardens the profiling toolchain (TraceLens analyzer + dynolog integration) to behave correctly on real AMD GPU nodes, addressing invalid UTF-8 rocprofv3 JSON traces, dynolog request/response handling, and multi-rank collective report trace selection. It also updates packaging metadata so the tracelens extra can be installed from git under hatchling.

Changes:

• Add incremental UTF-8 validation and “sanitized copy” analysis for rocprofv3 / PyTorch JSON traces, and keep --python isolation by avoiding PATH fallback.
• Fix dynolog trigger semantics by interpreting processesMatched from output (exit code is not meaningful) and failing fast on rejected requests; improve stop-script diagnostics.
• Restrict multi-rank collective report inputs to rank-labelled traces and add regression tests; enable hatch direct references for pip install -e ..

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tests/unit/test_tracelens_analyze.py	Updates collective-args expectations and adds coverage for skipping collective reports when ranks are not identifiable.
tests/fixtures/dummy_tracelens/TraceLens/Reporting/_dummy.py	Makes the dummy TraceLens fail on invalid UTF-8 JSON to emulate orjson/TraceLens behavior.
tests/e2e/test_tracelens_dummy_pipeline.py	Adds e2e regression tests for analyzing traces containing undecodable bytes via sanitized copies.
tests/e2e/test_dynolog_dummy_pipeline.py	Expands the dyno stub contract and adds tests for fail-fast behavior on rejected requests and accurate reporting.
src/madengine/scripts/common/tools/tracelens_analyze.py	Implements incremental UTF-8 detection + sanitization, improves collective report trace selection/globbing, and enforces --python environment isolation.
src/madengine/scripts/common/tools/dynolog_trigger.sh	Switches from non-existent flag semantics to parsing processesMatched, and fails fast when dyno rejects the request.
src/madengine/scripts/common/post_scripts/dynolog_stop.sh	Improves post-run messaging to distinguish “no process matched” vs “request rejected”.
pyproject.toml	Enables hatchling allow-direct-references to support git-based dependencies in extras.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.