tlhub 0.1.0

Local TORCH_TRACE run hub with artifact browsing and diffing.

tlhub

tlhub is a local hub for TORCH_TRACE runs.

Instead of manually setting TORCH_TRACE, running a program, and then feeding the logs into tlparse, tlhub folds that loop into one tool:

1. Wraps a command and sets TORCH_TRACE for it.
2. Indexes the resulting trace logs into browsable artifacts.
3. Keeps a local daemon-backed UI where runs can be viewed, compared, diffed, and deleted.

tlhub does not reuse tlparse's HTML renderer. It parses the structured trace logs directly and renders its own viewer, but it now targets the same major report surfaces as tlparse: compile summaries, stack trie, failures and restarts, raw JSONL, export diagnostics, provenance tracking, vLLM views, and multi-rank analysis.

Install

python3 -m pip install tlhub

From source:

python3 -m pip install -e .

Quick start

Prefix any command with tlhub:

tlhub python train.py

That will:

• create a new run directory under TLHUB_HOME or ~/.local/share/tlhub
• set TORCH_TRACE for the wrapped command
• check for the background daemon and start it if needed
• ingest the resulting trace logs
• print a run URL you can open in the browser

Open the dashboard without running a command:

tlhub

Stop the daemon:

tlhub --stop

There are no user-facing subcommands. Run deletion and cross-run comparison happen in the web UI.

Viewer parity targets

The viewer now exposes the main tlparse surfaces in tlhub form:

• run dashboard
• per-run compile directory
• compile-detail pages with stack, output files, compile metrics, custom op info, symbolic shape specialization, created symbols, unbacked symbols, and guard-added-fast data
• stack trie over compile stacks
• failures and restarts summary
• raw.jsonl shortraw-style output with the string table prepended
• export diagnostics plus guard detail pages
• provenance-tracking pages that align pre-grad graphs, post-grad graphs, and generated code via node mappings
• vLLM-specific summary pages with piecewise split graphs, compile config, and per-subgraph artifact listings
• multi-rank diagnostics:
  • compile id divergence
  • cache-pattern divergence
  • collective divergence
  • tensor-meta divergence
  • runtime delta summaries
  • execution-order summaries
• combined and derived report artifacts such as:
  • chromium_events.json
  • runtime_estimations.json
  • chromium_trace_with_runtime.json
  • collective_schedules.json
  • collectives_parity.json
  • execution_order_report.json
  • compile_directory.json

The diffing workflow is built into the UI. You can either:

• compare two runs and let tlhub line up artifacts by family plus occurrence index
• pick any two artifacts manually and diff them directly

This is especially useful for FX graphs, Inductor output code, and report JSON. FX-graph diffs also get a graph-aware semantic view so added, removed, and retargeted nodes are visible before you drop to raw line diffs.

What gets indexed

The ingester understands the same raw TORCH_TRACE log shape that tlparse reads:

• glog-formatted structured log lines
• JSON envelopes
• tab-indented payload bodies following has_payload
• string-table ("str") entries used by stack frames

It extracts and stores artifacts such as:

• graph payloads like dynamo_output_graph, aot_*_graph, and inductor_*_graph
• graph_dump
• inductor_output_code
• dynamo_guards
• generic artifact payloads with string or json encoding
• memoizer_artifacts
• dump_file
• exported_program
• synthetic report outputs derived from the run

Artifacts are grouped by a stable family key plus occurrence index so that matching outputs from two runs can be lined up and diffed.

Web UI

The local UI provides:

• a run dashboard
• a report section for raw.jsonl and derived diagnostics
• compile-detail pages
• export-guard detail pages
• provenance detail pages
• per-run artifact browsing
• artifact-family matching across two runs
• arbitrary artifact-to-artifact diffing
• run deletion

For graph-like text artifacts, the viewer records lightweight summaries such as node counts and op buckets to make diffs easier to scan. Synthetic reports are also first-class artifacts, so you can diff analysis outputs across runs, not just raw payloads.

Daemon behavior

The daemon is local-only and binds to 127.0.0.1.

Every normal tlhub invocation checks whether the daemon is already up and starts it automatically when needed. tlhub --stop is the only normal daemon-management command.

Tests

python3 -m unittest discover -s tests -v

Publish

Build and validate the distribution locally:

uv build
uv run --with twine python -m twine check dist/*

Publish directly from your machine:

uv publish

The repo also includes a GitHub Actions workflow at .github/workflows/publish.yml that publishes on tags like v0.1.0.

Before using the workflow, configure PyPI Trusted Publishing for:

• owner or user: bobrenjc93
• repository: tlhub
• workflow: publish.yml
• environment: pypi

Release flow:

1. Update src/tlhub/__init__.py with the new version.
2. Commit and push main.
3. Create and push a matching tag such as v0.1.0.