Forge - GPU Kernel Optimization
@RightNow-AI
Turn slow PyTorch into fast CUDA/Triton kernels. 32 parallel swarm agents optimize your code on real datacenter GPUs (B200, H200, H100, A100) with up to 14x speedup over torch.compile.
Overview
What is Forge - GPU Kernel Optimization?
Forge is an MCP server that optimizes PyTorch code into fast CUDA/Triton kernels using 32 parallel swarm agents. It is designed for AI developers and agents who want to accelerate GPU operations without manual kernel engineering. The server provides tools to optimize existing kernels, generate new ones from descriptions, and benchmark results on datacenter GPUs (B200, H200, H100, A100, L40S, T4).
How to use Forge - GPU Kernel Optimization?
Add the server to an MCP client (e.g., Claude) with the command claude mcp add forge-mcp -- npx -y @rightnow/forge-mcp-server. Authenticate via browser-based OAuth using the forge_auth tool, then call forge_optimize (to optimize PyTorch code) or forge_generate (to generate kernels from a description). Additional tools include forge_credits to check balance, forge_status to monitor jobs, forge_cancel to cancel jobs, and forge_sessions to list past sessions.
Key features of Forge - GPU Kernel Optimization
- Optimize existing PyTorch code into faster Triton/CUDA kernels
- Generate production-ready kernels from a natural language description
- 32 parallel swarm agents (Coder+Judge pairs) compete for the fastest kernel
- Real GPU benchmarking on datacenter hardware (B200, H200, H100, A100, L40S, T4)
- Up to 14x faster than torch.compile(mode='max-autotune')
- One-click browser-based OAuth, no API keys needed
Use cases of Forge - GPU Kernel Optimization
- Accelerate custom PyTorch operations in AI models
- Replace slow Python loops with optimized GPU kernels without manual CUDA coding
- Generate kernels for novel operations described in natural language
- Benchmark and compare kernel performance across multiple GPU architectures
FAQ from Forge - GPU Kernel Optimization
What authentication method does Forge use?
One-click browser-based OAuth; no API keys are needed.
What GPUs are used for benchmarking?
Datacenter GPUs: B200, H200, H100, A100, L40S, and T4.
How much does Forge cost?
Pay-as-you-go at $15/credit, with 25% off for 10+ credits. A free trial is included – optimize one kernel with no credit card required.
How many parallel agents work on optimization?
32 parallel swarm agents, each a Coder+Judge pair that competes to find the fastest kernel.
What tools does Forge provide?
Seven tools: forge_auth, forge_optimize, forge_generate, forge_credits, forge_status, forge_cancel, and forge_sessions.
