Hunyuan Gamecraft single-gpu implementation on Fal's serverless cloud
Python
0
0 commits
updated Aug 17, 2025
This is my trying Fal Serverless to deploy a custom container image and host it at: https://fal.ai/models/anthonywu/gamecraft (not promotable for now until bugs fixed)
The model deployed is Hunyuan Gamecraft which as of Aug 16 does not have a inference provider in the USA, so I thought it was a neat challenge to push the boundaries of my CUDA understanding as well as pushing GPU inference timeout limits.
action_list and action_speed_list which is where the camera moves for each iteration, and how fast the camera is moving (think: game controllers generate this data)docker build a amd64-destined container was easier than I expected. I've used Docker for 10+ years but never had to use emulation during docker packaging. This was frictionless.torch_run command, assuming the user would have shell access on a Nvidia host to do the inference. I had to re-write the inference flow to be compatible with how fal.Apps are pickled and packaged.fal CLI must run in a venv with a Python version that matches the Python that will execute the codeimport statements must be inside of fal.App's instancemethods. For this, I have to mute the linters and checkers.Hunyuan Gamecraft single-gpu implementation on Fal's serverless cloud
Python
0
0 commits
updated Aug 17, 2025
This is my trying Fal Serverless to deploy a custom container image and host it at: https://fal.ai/models/anthonywu/gamecraft (not promotable for now until bugs fixed)
The model deployed is Hunyuan Gamecraft which as of Aug 16 does not have a inference provider in the USA, so I thought it was a neat challenge to push the boundaries of my CUDA understanding as well as pushing GPU inference timeout limits.
action_list and action_speed_list which is where the camera moves for each iteration, and how fast the camera is moving (think: game controllers generate this data)docker build a amd64-destined container was easier than I expected. I've used Docker for 10+ years but never had to use emulation during docker packaging. This was frictionless.torch_run command, assuming the user would have shell access on a Nvidia host to do the inference. I had to re-write the inference flow to be compatible with how fal.Apps are pickled and packaged.fal CLI must run in a venv with a Python version that matches the Python that will execute the codeimport statements must be inside of fal.App's instancemethods. For this, I have to mute the linters and checkers.