wafer-cli 0.2.32

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

Wafer CLI

Wafer CLI gives coding agents direct access to GPU docs, trace analysis, and remote kernel evaluation. It helps you develop and optimize GPU kernels even when you are not working on a machine with a GPU.

Key features

• Query GPU documentation with citations
• Analyze GPU traces and profiles
• Evaluate kernels on remote GPUs for correctness and performance
• Run commands on GPU targets (remote or local)
• Manage persistent workspaces

Quick start

uv tool install wafer-cli
wafer login
wafer remote-run -- nvidia-smi

Common commands

wafer workspaces list
wafer workspaces create my-workspace --wait
wafer agent -t ask-docs --corpus cuda "What causes shared memory bank conflicts?"
wafer agent -t trace-analyze --args trace=./profile.ncu-rep "Why is this kernel slow?"
wafer evaluate --impl kernel.py --reference ref.py --test-cases tests.json --benchmark
wafer nvidia ncu analyze profile.ncu-rep
wafer corpus list

Typical workflows

Query GPU documentation

Download a documentation corpus and ask questions with citations.

wafer corpus download cuda
wafer agent -t ask-docs --corpus cuda "What causes shared memory bank conflicts?"

Analyze performance traces

Use the trace analysis template or query trace data directly.

wafer agent -t trace-analyze --args trace=./profile.ncu-rep "Why is this kernel slow?"
wafer nvidia perfetto query trace.json \
  "SELECT name, dur/1e6 as ms FROM slice WHERE cat='kernel' ORDER BY dur DESC LIMIT 10"

Evaluate kernels on remote GPUs

Run correctness and performance checks on a remote target.

wafer evaluate \
  --impl ./kernel.py \
  --reference ./reference.py \
  --test-cases ./tests.json \
  --benchmark

Run commands on a remote GPU

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

Manage workspaces

wafer workspaces list
wafer workspaces create my-workspace --wait
wafer workspaces ssh <workspace-id>
wafer workspaces delete <workspace-id>

Install the CLI skill (optional)

wafer skill install
# or
wafer skill install -t <claude/codex>