Nvidia tensorrt llm ad graph dump

AutoDeploy: Graph dumps (AD_DUMP_GRAPHS_DIR)

Where this skill applies

This file is part of trtllm-agent-toolkit. Commands and paths such as examples/auto_deploy/ and tensorrt_llm/ are relative to a TensorRT-LLM source checkout, not the plugin repository.

When to use this skill

• You need to see how the FX graph changes after each registered transform runs.
• You are verifying that a subgraph exists, that a fusion matched, or that metadata / wrappers (getitem, view, reshape) appeared or disappeared between dumps.
• You are pairing log output with on-disk graph files while debugging AutoDeploy.

Related skills in this plugin

Skill	Use it for
ad-layer-visualizer	Extracting and visualizing a single decoder layer from a dump as a DOT/PNG diagram.
ad-add-fusion-transformation	Implementing or reviewing fusion passes once you know what the graphs show.
trtllm-codebase-exploration	Searching the TRT-LLM tree for transforms, custom ops, and patterns.
trtllm-code-contribution	Tests and contribution hygiene after you change TRT-LLM.

Environment variable

Set:

export AD_DUMP_GRAPHS_DIR=/path/to/output/dir

Implementation: GraphWriter.DUMP_GRAPHS_ENV == "AD_DUMP_GRAPHS_DIR" in tensorrt_llm/_torch/auto_deploy/utils/graph_writer.py.

If unset, no graph files are written.

When dumps are produced

After each transform application, BaseTransform calls graph_writer.dump_graph(mod, t_name, self.config.stage.value) from tensorrt_llm/_torch/auto_deploy/transform/interface.py (immediately after _visualize_graph). So the dump reflects the module after that transform has run.

Rank / process behavior

From GraphWriter.dump_graph:

• Dumps run only when AD_DUMP_GRAPHS_DIR is set.
• If ADLogger.rank is set and is not 0, dumping is skipped (non–rank-0 processes do not write files).

Directory lifecycle

On the first dump on rank 0, GraphWriter removes the target directory if it already exists, then recreates it. Do not point AD_DUMP_GRAPHS_DIR at a directory that must be preserved without copying it first.

File naming and ordering

Files are named:

{NNN}_{<stage.value>}_{<transform_key>}.txt

• NNN is a monotonically increasing three-digit counter (001, 002, …) in run order across all dumps in that process.
• The middle segment is each transform’s config.stage value (same enum/string used in default.yaml under each transform’s stage: field).
• The last segment is the transform’s registry key (transform_name passed into dump_graph).

So lexicographic sort by filename matches pipeline order for that run.

File contents

Each file is text and starts with headers similar to:

# Transform: <transform_key>
# Stage: <stage.value>
# GraphModules found: <count>

Then, for every torch.fx.GraphModule found under mod.named_modules() (including the root), the writer emits a section title and an SSA-style listing with shape/dtype metadata via dump_ssa_with_meta() in the same module.

Use this to compare operator chains, consumers, and node.meta shape/dtype hints across consecutive files.

Example: capture dumps from a registry build

From the root of the TensorRT-LLM clone (adjust the script and flags to your workflow):

AD_DUMP_GRAPHS_DIR=/tmp/ad-graphs \
  python examples/auto_deploy/build_and_run_ad.py --model <hf-model-id> --use-registry

Pick any AutoDeploy entrypoint you already use; the requirement is only that the code path runs the transform pipeline with AD_DUMP_GRAPHS_DIR set in the environment.

Logs vs dump files

While a transform runs, logging is patched so messages can be prefixed with [stage=<stage.value>, transform=<transform_key>] (see with_transform_logging in transform/interface.py). Transform summaries log [SUMMARY] with matches=<n> or skipped / disabled (_log_transform_summary). Use those lines together with the numbered dump files to tie match counts to graph shape before and after a specific transform.

Pitfalls

• Stale directory: Because the dump dir is deleted on first use, a second run in the same shell without changing AD_DUMP_GRAPHS_DIR overwrites prior output.
• No GraphModules: If the module has no GraphModule children, dump_graph returns without creating a new file for that step (see early return in graph_writer.py).
• Distributed: Only rank 0 writes; other ranks skip silently.

Source references

• tensorrt_llm/_torch/auto_deploy/utils/graph_writer.py — env var, filenames, SSA dump.
• tensorrt_llm/_torch/auto_deploy/transform/interface.py — call site after each transform; log prefix decorator.