NVIDIA Optimize Model Precision
Name the objective first: bundle size, device memory, setup time, Task API latency, throughput, or another measured family outcome. The best candidate is the lowest-cost option that still passes the existing family contract.
Establish a baseline
Use one family manifest and testcase. Record the repository and checkpoint revisions, hardware, build request, seed, input, bundle path, correctness result, warmups, iterations, and measured boundary.
Build the current declared configuration and run its E2E before exploring
alternatives. The public build interface exposes direct values for
--precision, --quantization, and repeated --fp32-layer selections.
Attempt only values the owning family implements.
python -m tensorrt_model_connect build <model-id-or-path> \
--revision <revision> --precision <fp32-or-fp16-or-bf16> \
--output <candidate.bundle>
trtmc inspect <candidate.bundle>
Change one effective value per attempt. If a candidate fails, localize the first mismatch in the owner family before changing another value. Do not add cross-family precision policy or weaken a threshold, dataset, seed, or oracle.
Qualify a candidate
Run the exact family E2E testcase with the candidate configuration. Only after it passes, compare Task API performance with matched inputs and protocol:
trtmc-bench run --model <model> --case <case> \
--bundle <candidate.bundle> --no-build \
--warmup 3 --iterations 10 --runtime-root <runtime-root> \
--output <result-dir>
Use tools/perf_matrix.py when the claim targets a checked-in release entry.
Report every attempt in a compact table, including effective request values,
correctness, measured objective, confounders, and unrun targets. It is valid
for the original configuration to remain the best qualified result.