wafer-cli 0.2.50

CLI for running GPU workloads, managing remote workspaces, and evaluating/optimizing kernels

Wafer CLI

Run GPU workloads, optimize kernels, and query GPU documentation.

Getting Started

# Install
cd apps/wafer-cli && uv sync

# Use staging (workspaces and other features require staging)
wafer config set api.environment staging

# Login
wafer login

# Run a command on a remote GPU
wafer remote-run -- nvidia-smi

Commands

wafer login / wafer logout / wafer whoami

Authenticate with GitHub OAuth.

wafer login          # Opens browser for GitHub OAuth
wafer whoami         # Show current user
wafer logout         # Remove credentials

wafer remote-run

Run any command on a remote GPU.

wafer remote-run -- nvidia-smi
wafer remote-run --upload-dir ./my_code -- python3 train.py

wafer workspaces

Create and manage persistent GPU environments.

Available GPUs:

• MI300X - AMD Instinct MI300X (192GB HBM3, ROCm)
• B200 - NVIDIA Blackwell B200 (180GB HBM3e, CUDA) - default

wafer workspaces list
wafer workspaces create my-workspace --gpu B200 --wait   # NVIDIA B200
wafer workspaces create amd-dev --gpu MI300X             # AMD MI300X
wafer workspaces ssh <workspace-id>
wafer workspaces delete <workspace-id>

wafer agent

AI assistant for GPU kernel development. Helps with CUDA/Triton optimization, documentation queries, and performance analysis.

wafer agent "What is TMEM in CuTeDSL?"
wafer agent -s "optimize this kernel" < kernel.py

wafer evaluate

Evaluate kernel correctness and performance against a reference implementation.

Functional format (default):

# Generate template files
wafer evaluate make-template ./my-kernel

# Run evaluation
wafer evaluate --impl kernel.py --reference ref.py --test-cases tests.json --benchmark

The implementation must define custom_kernel(inputs), the reference must define ref_kernel(inputs) and generate_input(**params).

KernelBench format (ModelNew class):

# Extract a KernelBench problem as template
wafer evaluate kernelbench make-template level1/1

# Run evaluation
wafer evaluate kernelbench --impl my_kernel.py --reference problem.py --benchmark

The implementation must define class ModelNew(nn.Module), the reference must define class Model, get_inputs(), and get_init_inputs().

wafer wevin -t ask-docs

Query GPU documentation using the docs template.

wafer wevin -t ask-docs --json -s "What causes bank conflicts in shared memory?"

wafer corpus

Download documentation to local filesystem for agents to search.

wafer corpus list
wafer corpus download cuda-programming-guide

---

Customization

wafer remote-run options

wafer remote-run --image pytorch/pytorch:2.5.1-cuda12.4-cudnn9-devel -- python3 script.py
wafer remote-run --require-hwc -- ncu --set full python3 bench.py   # Hardware counters for NCU

wafer evaluate options

wafer evaluate --impl k.py --reference r.py --test-cases t.json \
    --target vultr-b200 \    # Specific GPU target
    --benchmark \            # Measure performance
    --profile                # Enable torch.profiler + NCU

wafer push for multi-command workflows

WORKSPACE=$(wafer push ./project)
wafer remote-run --workspace-id $WORKSPACE -- python3 test1.py
wafer remote-run --workspace-id $WORKSPACE -- python3 test2.py

Profile analysis

wafer nvidia ncu analyze profile.ncu-rep
wafer nvidia nsys analyze profile.nsys-rep

---

Advanced

Local targets

Bypass the API and SSH directly to your own GPUs:

wafer targets list
wafer targets add ./my-gpu.toml
wafer targets default my-gpu

Defensive evaluation

Detect evaluation hacking (stream injection, lazy evaluation, etc.):

wafer evaluate --impl k.py --reference r.py --test-cases t.json --benchmark --defensive

Other tools

wafer perfetto <trace.json> --query "SELECT * FROM slice"   # Perfetto SQL queries
wafer capture ./script.py                                    # Capture execution snapshot
wafer compiler-analyze kernel.ptx                            # Analyze PTX/SASS

ROCm profiling (AMD GPUs)

wafer rocprof-sdk ...
wafer rocprof-systems ...
wafer rocprof-compute ...

---

Shell Completion

Enable tab completion for commands, options, and target names:

# Install completion (zsh/bash/fish)
wafer --install-completion

# Then restart your terminal, or source your shell config:
source ~/.zshrc  # or ~/.bashrc

Now you can tab-complete:

• Commands: wafer eva<TAB> → wafer evaluate
• Options: wafer evaluate --<TAB>
• Target names: wafer evaluate --target v<TAB> → wafer evaluate --target vultr-b200
• File paths: wafer evaluate --impl ./<TAB>

---

AI Assistant Skills

Install the Wafer CLI skill to make wafer commands discoverable by your AI coding assistant:

# Install for all supported tools (Claude Code, Codex CLI, Cursor)
wafer skill install

# Install for a specific tool
wafer skill install -t cursor    # Cursor
wafer skill install -t claude    # Claude Code
wafer skill install -t codex     # Codex CLI

# Check installation status
wafer skill status

# Uninstall
wafer skill uninstall

Installing from GitHub (Cursor)

You can also install the skill directly from GitHub in Cursor:

1. Open Cursor Settings (Cmd+Shift+J / Ctrl+Shift+J)
2. Navigate to Rules → Add Rule → Remote Rule (Github)
3. Enter: https://github.com/wafer-ai/skills
4. Cursor will automatically discover skills in .cursor/skills/

The skill provides comprehensive guidance for GPU kernel development, including documentation lookup, trace analysis, kernel evaluation, and optimization workflows.

---

Requirements

• Python 3.10+
• GitHub account (for authentication)