Debug TRT Mismatch

Goal

Find the first boundary where TensorRT and the declared reference disagree. Preserve the failing workload, sampling settings, model revision, bundle, and runtime strategy while narrowing the problem. Do not hide a mismatch by loosening validation thresholds.

Preflight

Record the exact revision and environment:

git rev-parse HEAD
nvidia-smi --query-gpu=name,driver_version --format=csv,noheader
python3 -c "import tensorrt as trt; print(trt.__version__)"
PYTHONPATH=python:. python3 -c "import tensorrt_model_connect; print('package: OK')"
test -x ./build/trtmc

If the local checkout does not have the required GPU or TensorRT environment, use the existing team container rather than changing the investigation:

docker ps -a --filter "name=trtmc-dev-gb300" --format "{{.Names}} {{.Status}}"
./scripts/bootstrap_workspace.sh --id <team-id> --branch "$(git branch --show-current)" --detach

Run later commands inside trtmc-dev-gb300-<team-id> when needed.

Reproduce Through The Owned Validation Path

List model-first workloads and dry-run the failing one:

PYTHONPATH=python:. python3 tools/trtmc_validate.py --list
PYTHONPATH=python:. python3 tools/trtmc_validate.py \
  <model> <workload> \
  --dry-run \
  --output /tmp/trtmc-validation

The model binding in tests/validation/model_workloads.yaml, workload contract in tests/validation/workloads.yaml, family manifests, and their sidecars are the source of truth for inputs, sampling, and comparison gates. Read tools/validation/README.md before changing the engine: the persisted task_eval artifact key remains intentionally stable even though executable code moved to tools/validation/. Keep reference generation and comparison separate in the report; an execution failure is not a numerical mismatch.

Route By Model Capability

Inspect the model's Python, C++, and E2E MODEL.toml entries before selecting a debugger:

scripts/validate_family.sh <model> already routes declared family capabilities. Do not force decoder-only tools onto audio, diffusion, or another family that does not declare that profile.

Decoder Escalation

Start with logits:

PYTHONPATH=python:. python3 tools/diff_logits.py \
  --model <model> \
  --prompt "The capital of France is" \
  --max-new-tokens 10 \
  --atol 1e-3 \
  --json /tmp/diff-logits.json \
  --verbose

Interpret the first divergent step:

If logits identify a graph mismatch but not its origin, compare layers:

PYTHONPATH=python:. python3 tools/diff_layers.py \
  --model <model> \
  --prompt "Hello" \
  --atol 0.05 \
  --verbose

Use the testcase's declared tolerance as the authority. 0.01, 0.05, and 0.1 are investigation starting points, not replacement pass criteria.

Modality-Specific Escalation

Vision-language:

PYTHONPATH=python:. python3 tools/diff_vl.py \
  --bundle /tmp/model.bundle \
  --image /path/to/test.jpg \
  --model <model> \
  --binary ./build/trtmc \
  --hf-python <python> \
  --debug-layers

Audio:

PYTHONPATH=python:. python3 tools/diff_audio.py \
  --bundle /tmp/model.bundle \
  --binary ./build/trtmc \
  --model <model> \
  --hf-python <python> \
  --stage 1

Escalate audio stages only after the earlier stage passes. Stage 4 checks greedy token parity and does not replace waveform or distribution checks.

Diffusion:

PYTHONPATH=python:. python3 tools/debug_diffusion_pipeline.py \
  --bundle /tmp/model.bundle \
  --model-id <model> \
  --num-steps <small-reproducer-steps>

Use deterministic inputs and seeds. Localize preprocessing, encoder, denoiser, scheduler, and decoder boundaries before changing a family implementation.

Runtime Boundary

When Python/reference parity passes but the C++ result differs:

PYTHONPATH=python:. python3 tools/test_runner_parity.py \
  --bundle /tmp/model.bundle \
  --binary ./build/trtmc \
  --hf-python <python> \
  --prompt "The capital of France is" \
  --max-new-tokens 20

Inspect tokenizer inputs, bundle metadata, runtime strategy selection, masks, positions, and cache state. Keep the Python debug path aligned with the C++ runtime contract.

Operation Isolation

After identifying a layer or stage, run the family-owned tests nearest to the implementation. Graph helpers belong under the owning family; root graph helper modules are intentionally absent and are enforced by tests/tools/test_model_plugin_encapsulation_static.py.

For a missing focused test, create the smallest deterministic TensorRT graph that reproduces the operation and compare it to the reference. Do not broaden a model-local fix into shared runtime code without proof that multiple families share the same contract.

Report

Include:

• exact repository SHA, model revision, model/workload, runtime strategy, and bundle path or hash;
• exact reproduction and focused-debug commands;
• execution, reference, comparison, and validation status as separate facts;
• first divergent sample, token, layer, stage, or operation and its metrics;
• what passed and was ruled out;
• smallest ownership-aligned fix and the validation that must pass afterward.

When the mismatch came from Internal CI, keep private logs, artifacts, runner details, and package coordinates private. Source PRs may cite only the sanitized exact-head TRTMC Internal CI / Automated premerge gate status and public reproduction evidence.