IntelliKit

What’s in the box

IntelliKit is a set of Python tools for AMD-focused performance and validation. Most of the stack targets GPUs through ROCm, turning hardware counters, traces, and dispatch data into clear APIs you can use from Python. uprof_mcp adds AMD uProf for host-side CPU hotspot analysis. For LLM-style workflows you also get MCP servers and agent skills — installable SKILL.md playbooks for Cursor, Claude, Codex, and GitHub Copilot.

Tool	Role	Description
Kerncap	Isolate	Capture kernel dispatches, build standalone reproducers for HIP and Triton
Metrix	Profile	Human-readable metrics from hardware counters: bandwidth, cache, compute
Linex	Profile	Source-line timing and stall analysis — map GPU performance to your code
Nexus	Inspect	Intercept HSA packets to see what ran on the GPU: assembly and HIP source
Accordo	Validate	Prove an optimized kernel still matches a reference implementation
ROCm MCP	MCP	HIP compiler, HIP docs, and rocminfo servers for LLM agents
uProf MCP	CPU	MCP bridge to AMD uProf for host-side CPU hotspot analysis

Install

# Install all tools
curl -sSL https://raw.githubusercontent.com/AMDResearch/intellikit/main/install/tools/install.sh | bash

# Install agent skills (Cursor, Claude, Codex, GitHub Copilot)
curl -sSL https://raw.githubusercontent.com/AMDResearch/intellikit/main/install/skills/install.sh | bash

The workflow

Isolate a kernel with Kerncap, profile it with Metrix and Linex, inspect execution with Nexus, wire agents via MCP servers and skills, add uProf for host-side analysis, then lock in correctness with Accordo.

from metrix import Metrix
from nexus import Nexus
from accordo import Accordo

# 1) Profile — human-readable GPU metrics
profiler = Metrix()
baseline = profiler.profile("./app", metrics=["memory.hbm_bandwidth_utilization"])

# 2) Inspect — see what ran on the GPU
trace = Nexus().run(["./app"])
for kernel in trace:
    print(kernel.name, len(kernel.assembly), "instructions")

# 3) Validate — check correctness after optimization
validator = Accordo(binary="./app", kernel_name="my_kernel")
ref = validator.capture_snapshot(binary="./app_ref")
opt = validator.capture_snapshot(binary="./app_opt")
result = validator.compare_snapshots(ref, opt, tolerance=1e-6)
print(f"{'PASS' if result.is_valid else 'FAIL'}")