One 32 GB card (RTX 5090) for the agent; fix the hf download commands; which setup for which GPU

by yycc - opened about 13 hours ago

←

Viggle AI org about 13 hours ago

•

Fix the download commands. The hf CLI in huggingface_hub 1.x takes one pattern per --include: with --include a b, b is read as a filename and every --include is ignored, so the Quick start fetched neither the LoRAs nor modular_model_index.json, and on Windows it fails on * in a filename (reported in #3, thanks). Each pattern now gets its own --include in README.md and COMFYUI.md, and both blocks start with pip install -U "huggingface_hub>=1.0", since 0.3x keeps only the last --include. Checked with hf download --dry-run: 144.1 GB base, 5.5 GB Remake, 7.6 GB ComfyUI.
Which setup for your machine: a table in the README Quick start, from no GPU (the Viggle app) through 12 to 24 GB (ComfyUI by hand), 32 GB (the agent with the ComfyUI worker), 80 GB (sample.py) and 96 GB or more (viggle-remake-worker).
mcp/README.md: what the agent needs (an MCP client, a painter, render workers), and ffmpeg 5.1 or newer.
mcp/comfy_worker/: a queue worker on ComfyUI 0.42 with the int8 transformers and dynamic VRAM, so the agent runs on one 32 GB card (1280x736 up to 345 frames, about 60 GB of RAM). The node pack gains RemakePrefix for hops; ffmpeg -vsync (removed in ffmpeg 9) becomes -fps_mode.

yycc changed pull request status to merged about 11 hours ago

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment