Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/nvidia/tensorrt-model-connect/optimize-model-precisionnpx skills add NVIDIA/TensorRT-Model-Connect --skill optimize-model-precisiongit clone --depth 1 https://github.com/NVIDIA/TensorRT-Model-ConnectWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00064 | $0.01488 |
| Opus 5 | $0.00032 | $0.00744 |
| Sonnet 5 | $0.00013 | $0.00298 |
| Haiku 4.5 | $0.00006 | $0.00149 |
Grade A, and why
optimize-model-precision scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 190 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Optimize Model Precision
Objective
Find the lowest-cost configuration that satisfies the model's existing correctness contract and improves a named resource or performance metric. “Best” must name the objective: bundle size, device memory, setup time, prefill, decode, throughput, or another model-owned metric.
Do not weaken an oracle, threshold, sample set, or pass criterion to make a configuration qualify. If a test appears wrong, stop and escalate it to a maintainer.
Establish The Owned Baseline
Resolve the model through all relevant descriptors:
- Python family
MODEL.tomland plugin; - C++ model
MODEL.tomland runtime strategy; - E2E family
MODEL.toml, manifest, testcases, thresholds, andperf_validation.jsonwhen present; - model-first binding in
tests/validation/model_workloads.yaml; - optimized implementation/profile/qualification descriptors when selected.
List and dry-run the reference-consistency workload:
PYTHONPATH=python:. python3 tools/trtmc_validate.py --list
PYTHONPATH=python:. python3 tools/trtmc_validate.py \
<model> <workload> \
--dry-run \
--output <baseline-plan-dir>
Build and validate the existing configuration before optimizing it. Record: repository SHA, exact model revision, target/hardware, runtime path, effective build options, bundle hash, workload and sample limit, seed/sampling, artifact paths, correctness metrics, and the performance protocol.
Build Matrix
Try only formats supported by the current CLI and owning family:
./build/trtmc build <model> -o <bundle>.bundle \
--precision fp16 \
--max-cache-length <N>
./build/trtmc build <model> -o <bundle>.bundle \
--precision fp16 \
--quantize <supported-format> \
--quant-calibration-samples <N> \
--max-cache-length <N>
The current quantization core resolves a QuantPlan; family hooks own
calibration data, adapters, exclusion patterns, and FP8 scales. Use the current
CLI help and website/docs/features/quantization.md for supported options.
Use --quant-scales for a reviewed generic scale artifact and the dedicated
FP8 scale flags only for their documented compatibility path. Do not bypass
the plan with ad hoc family Q/DQ code.
What ships with it
1 file beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 3d ago First seen · 190 lines · 64 tokens per session scan A 6a53f3dc414b
optimize-model-precision is a skill published in the GitHub repository NVIDIA/TensorRT-Model-Connect (188 stars, last pushed 3d ago), licensed Apache-2.0. It adds 64 tokens to every session and 1,488 once invoked, about $0.0003 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
graphsignal-profiler
Set up the Graphsignal Profiler for inference workloads — vLLM, SGLang, PyTorch, and dstack services. Use when the user wants GPU profiling, tracing, or monitoring for inference, asks about graphsignal-run or graphsignal.watch(), or asks about CUPTI / Prometheus / OTLP setup.
weights-and-biases
Track ML experiments with automatic logging, visualize training in real-time, optimize hyperparameters with sweeps, and manage model registry with W&B - collaborative MLOps platform.
huggingface-accelerate
Simplest distributed training API. 4 lines to add distributed support to any PyTorch script. Unified API for DeepSpeed/FSDP/Megatron/DDP. Automatic device placement, mixed precision (FP16/BF16/FP8). Interactive config, single launch command. HuggingFace ecosystem standard.
mlflow
Track ML experiments, manage model registry with versioning, deploy models to production, and reproduce experiments with MLflow - framework-agnostic ML lifecycle platform.
huggingface-hub
Hugging Face Hub — model discovery, download, inference, and upload.
ml-training
Machine learning model training authority — classifier and regressor training with scikit-learn and PyTorch, feature engineering, cross-validation, hyperparameter tuning, fine-tuning transformer models with HuggingFace, dataset splits, loss functions, learning rate schedules, and reproducible training runs.