NVIDIA/TensorRT-Model-Connect is a collection of C++ reference implementations for deploying supported Hugging Face models with NVIDIA TensorRT, a system that runs trained models to produce inference results. It is for developers who want to build and run supported models or evaluate model integrations through TensorRT. The catalogue entries provide skills and instructions for using this model deployment workflow.
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/fp16-trt-networknpx skills add NVIDIA/TensorRT-Model-Connect --skill fp16-trt-networkgit clone --depth 1 https://github.com/NVIDIA/TensorRT-Model-ConnectWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/nvidia/tensorrt-model-connect/fp16-trt-network)<a href="https://agentmods.dev/skills/nvidia/tensorrt-model-connect/fp16-trt-network"><img src="https://agentmods.dev/badge/skills/nvidia/tensorrt-model-connect/fp16-trt-network.svg" alt="Measured on agentmods" height="20"></a>What 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 | $0.00059 | $0.01396 |
| Opus 5 | $0.00030 | $0.00698 |
| Sonnet 5 | $0.00012 | $0.00279 |
| Haiku 4.5 | $0.00006 | $0.00140 |
Grade A, and why
fp16-trt-network 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 5d 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 — 167 lines — stays where its author put it; the contents beside it link to each section on GitHub.
FP16/BF16 TensorRT Networks
Contract
TensorRT networks in this repository are strongly typed:
from tensorrt_model_connect import trt_compat
flags = trt_compat.network_creation_flags(strongly_typed=True)
network = builder.create_network(flags)
Follow an owning family's established direct flag when it is intentionally
backend-specific. For new backend-agnostic code, use trt_compat so TensorRT
versions without EXPLICIT_BATCH and the optional TensorRT-RTX backend share
one flag boundary. Backend selection, including --rtx, must happen before a
module imports TensorRT; never switch backends after TensorRT is loaded.
Precision follows tensor dtypes, typed constants, and explicit
network.add_cast(...) boundaries. Do not use BuilderFlag.FP16,
BuilderFlag.INT8, layer.setPrecision(), layer.setOutputType(), or direct
tensor.dtype mutation to override inference in a strongly typed network.
Keep changes in the owning family under
python/tensorrt_model_connect/families/<family>/. Root graph helper modules
are intentionally absent. Share a helper only within an ownership boundary
where shape, dtype, and layout semantics genuinely match.
Map Precision Before Editing
For every input, weight, constant, intermediate, state tensor, and output, record:
- storage dtype used to create the constant;
- TensorRT runtime dtype;
- shape and layout;
- the operation where a cast occurs;
- the required comparison dtype.
BF16 needs special care. Some family builders store constants in FP16-compatible
NumPy storage and explicitly cast them to trt.bfloat16. Do not assume a
NumPy dtype maps directly to the TensorRT dtype. Follow the owning family's
constant helper and checkpoint mapper.
FP32 Boundaries
Use the family implementation and reference numerics to decide boundaries. Common FP32 candidates include:
- normalization reductions and reciprocal/square-root arithmetic;
- softmax and probability normalization;
- batch/group statistics;
- final logits or comparison-sensitive outputs;
- unstable scale or calibration arithmetic.
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.
- 5d ago First seen · 167 lines · 59 tokens per session scan A 5840237bd66c
fp16-trt-network is a skill published in the GitHub repository NVIDIA/TensorRT-Model-Connect (219 stars, last pushed today), licensed Apache-2.0. It adds 59 tokens to every session and 1,396 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.
perf-host-analysis
Analyze host/CPU overhead in TensorRT-LLM inference from nsys traces. Detect whether host overhead is the bottleneck using GPU idle ratio, host prep exposed ratio, and per-phase evidence. For regressions, isolate forward steps via allreduce/NVTX patterns, compare host operation breakdowns across versions, and identify…
perf-host-optimization
Profiles and optimizes TensorRT-LLM host/CPU overhead using lineprofiler (with nsys support planned). Runs iterative profile-analyze-optimize-validate rounds. Use when GPU utilization is low or optimizing PyExecutor throughput.
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.
huggingface-hub
Hugging Face Hub — model discovery, download, inference, and upload.