RunPod Serverless engineer for production GPU image generation

Looking for proposals I’m running production GPU image generation on RunPod Serverless. The product needs pay-per-use scale-to-zero (workersMin = 0) and a cold path that is actually fast. Always-on workers are not not a solution, as the pipeline needs to be able to scale up and down as need be and sometimes that means everything parked (with ability for a quick restart) Stack (what you’ll touch) - Custom Docker GPU worker (our image, our handler/startup) - Multi-GB model weights on a network volume - Full product image-generation workflow (real customer path) - FlashBoot available So this is a real model + volume + serverless load problem, not an empty container demo. Stack weight (so expectations are right) This is a production image stack, not a light demo: multi-GB model weights on a network volume, custom GPU worker, and a full product generation path. Several pieces have to work together (assignment, volume load, startup, real gen). Treat it as a heavy cold-start / serverless problem, not a small “spin up an empty container and optimize a light model” job. Where things stand The stack exists and can generate (does fine when always on but moving to serverless). When a worker is already warm and the system is behaving, generation can be fine. We have not hit the finish line: true cold after scale-to-zero is not acceptably fast, and under min=0 we also see unreliable job assignment (jobs sitting while workers look available). We’ve tried some cold-start / startup changes on our side; however still running into issues. What needs to be resolved 1. Reliable job assignment and completion with min=0 2. Fast true-cold product generation after workers go to zero 3. Leave capacity at zero when you’re done What “done” means min=0 works, real product jobs complete, cold starts are fast enough for production use, and you can show how you verified it. Private details (IDs, model names, config) after hire. Who this is for People who have shipped RunPod Serverless cold starts with min=0 and large models / network volumes. Docker GPU workers, handler/startup work, debugging via API/console.