TensorRT LLM provides users with an easy-to-use Python API to define Large Language Models (LLMs) and supports state-of-the-art optimizations to perform inference efficiently on NVIDIA GPUs. TensorRT LLM also contains components to create Python and C++ runtimes that orchestrate the inference execution in a performant way.
Agent Claude Code
Updates the ad-conf-check skill's references/configlogpatterns.md by comparing it against the latest TensorRT-LLM AutoDeploy source code. Checks for new/removed/renamed configs in default.yaml and verifies that log patterns still match the actual source code. Edits the reference doc in-place if anything changed.
Agent Claude Code
Debug the AutoDeploy model onboarding process.
Agent Claude Code
Independent reviewer for AutoDeploy model onboarding. Validates created model and test files against all onboarding requirements. Use after completing model onboarding work.
Agent Claude Code
Run AutoDeploy build and run command for a given model.
Agent Claude Code
Compiles TensorRT-LLM from source. Handles two scenarios: (1) compiling directly on a compute node inside a dev Docker container, and (2) submitting a SLURM batch job from a login node with enroot container support. Detects the environment automatically, gathers required parameters, runs the build, monitors progress…
Agent Claude Code
Execute a TensorRT-LLM workload on a local Slurm cluster. Supports persistent allocation (allocate once via nohup salloc, reuse across runs) and one-shot sbatch. Workflow-agnostic — handles pytest, eval, benchmark, and custom scripts identically. The orchestrator (typically trtllm-case-executor) writes a job spec to…
Agent Claude Code
Execute a TensorRT-LLM workload on a remote Slurm cluster via SSH. Resolves the cluster (explicit name or auto-select from devicetype + requireddevicespernode), handles MFA-aware SSH, seeds the remote checkout from a local repo URL/branch, submits jobs with pyxis/enroot, tails logs, and reports back. The orchestrator…
Agent Claude Code
Hand-writes raw CUDA C/C++ code (.cu files) with pybind11 bindings (binding.cpp) to build custom PyTorch C++ extensions. Delegate ONLY when the user explicitly asks to write .cu/.cpp files compiled via torch.utils.cppextension. Do NOT delegate for: Triton, TileIR, or any other kernel DSL/framework. NOT for CUDA…
Agent Claude Code
Expert in writing optimized GPU kernels using CuTe DSL for NVIDIA GPUs (Ampere through Blackwell). This is the CuTe DSL specialist -- NOT Triton. CuTe DSL uses cute.jit/cute.kernel decorators and cutlass.cute imports; Triton uses triton.jit and tl. primitives -- they are completely different frameworks. Delegate to…
Agent Claude Code
Expert in optimizing EXISTING Triton kernels for NVIDIA TileIR backend. Does NOT write Triton kernels from scratch--that is the Triton Specialist's job. Delegate to this agent for: (1) Adding TileIR-specific autotune configs (occupancy, numctas, numstages), (2) Classifying kernels for optimization, (3) Validating with…
Agent Claude Code
Expert in writing optimized Triton kernels for PyTorch operators. Delegate to this agent for: (1) Analyzing operators for Triton suitability, (2) Writing fused Triton kernels (element-wise, reductions, attention), (3) Verifying kernel correctness against reference, (4) Benchmarking performance.
Agent Claude Code
Expert in GPU performance profiling for TRT-LLM workloads with nvidia-smi, Nsight Systems (nsys), Nsight Compute (ncu), and PyTorch profiler. This agent can execute shell commands directly. Delegate to this agent for: (1) Running workloads (Python scripts, CUDA binaries, shell commands), (2) Measuring performance…
Agent Claude Code
Use this agent when the user needs to synchronize performance test cases between development (dev) and QA directories, compare test configurations, update test lists, or analyze gaps between dev and QA perf test coverage. This includes syncing aggregated and disaggregated performance test cases, updating QA test lists…
Agent Claude Code
Expert in CUDA Graph capture, replay, and optimization for PyTorch. Delegate to this agent for: (1) Analyzing code for CUDA Graph compatibility, (2) Detecting and eliminating host-device synchronizations, (3) Selecting the right CUDA Graph API (torch.compile, makegraphedcallables, TE, MCore CudaGraphManager…
Agent Claude Code
Runs model-level and module-level tests for TensorRT-LLM. Classifies the test scope (module test or model test), builds the appropriate test commands, and delegates execution to trtllm-case-executor. Supports functionality/smoke tests, benchmarks, and evaluations. Writes structured test reports to a caller-specified…