Strata 连接 / Connection

Every other node in this pack has a connection input, and this is the node that feeds it. It has exactly one widget and one output, which makes it look like a placeholder. It isn't: it's the boundary between your workflow and a separate inference process, and almost every "why is Strata not answering" error comes back to what's in here.

What it actually is

Strata is not a model loaded into ComfyUI. It's a local language-and-vision engine that runs as its own process - the same "two processes or one" fork the KB describes for local LLM nodes (the KB's llm-in-comfyui notes § 3): some nodes talk to an Ollama server they don't own, some load the weights in-graph. This pack takes the first path, then adds a lifecycle layer on top so it can shut the engine down again before your sampler wants the card.

StrataT8Connection reads a saved local profile by name and hands out a STRATA_CONNECTION object. That's it. No network call, no GPU work, nothing heavy happens at import. The workflow stores only the profile name - the JSON with paths, port, thresholds and the API key lives outside custom_nodes, in %LOCALAPPDATA%/Strata-T8-ComfyUI (Linux: ~/.config/Strata-T8-ComfyUI, or wherever STRATA_COMFY_HOME points). That's deliberate: you can hand someone a workflow without handing them your key.

Inputs and outputs

• profile (COMBO) - populated from the profiles/ folder of that config directory. Names are restricted to 1–64 characters of letters, digits, _ and -, and Windows reserved device names (CON, NUL, COM1…) are rejected.
• connection (STRATA_CONNECTION) - wire it to every Strata node that needs it. One Connection node can fan out to all of them.

There is one nice trick in here: the node fingerprints the profile file's SHA-256 as its cache key. Change the port or raise min_free_vram_mib in the file and the downstream nodes re-run instead of serving you a cached answer from the old settings.

The profile, because it isn't in the node

Open the Strata-T8 sidebar, name a profile, paste the JSON, save, refresh, then pick the name here. Two modes:

• managed - the node pack starts and stops the portable runtime itself. You must give it runtime (D:/Strata-T8) and data_dir (D:/Strata-data); the URL is forced to http://127.0.0.1:<port>, lifecycle control is forced on, and same_gpu defaults to true.
• external - you already have a service running; give it a url and the pack only sends requests. allow_lifecycle and same_gpu both default to false, and same-GPU mode refuses to combine with allow_lifecycle:false or with a non-local host.

Other knobs worth knowing: vision (gpu / cpu / no), context (1024–131072, default 32768), timeout_s 1800, cleanup_timeout_s 90, min_free_vram_mib 12288, min_free_ram_gib 60. Those last two are the bouncer. They're what stops Strata from loading on top of your live diffusion model.

Install

ComfyUI-Manager → search Strata-T8 (node ID strata-t8, publisher t8star), or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-Strata-T8.git
python -m pip install -r Comfyui-Strata-T8/requirements.txt

That requirements.txt is one line - jsonschema. Image and video handling reuses ComfyUI's existing PyTorch, NumPy, Pillow and PyAV; the engine's own Python and CUDA/HIP stack stay in the portable runtime. Nothing downloads on install: the pack ships no weights, and the main IQ3_S data is ~83.6GB plus ~5.2GB of MTP tensors and a ~907MB BF16 vision encoder on the runtime side. Keep exactly one install directory - old comfyui-strata-t8 copies registering the same 12 node IDs cause duplicate-node confusion.

Where people get burned

The combo box lies when you have no profiles. With an empty profile folder it still lists default, and selecting it gives you "Create this local profile in the Strata panel first; workflows contain only its name." That's not a broken install, it's a missing profile.

The profile panel is loopback-only - it rejects anything that isn't a 127.0.0.1 connection with a matching Origin, so if you run ComfyUI on another machine and browse to it remotely, you cannot create the profile there. Set it up locally or write the JSON by hand.

Then the health checks, all of which report clearly: a managed runtime older than 0.1.39-t8.14 is refused ("update the portable runtime first"); a busy port is refused rather than silently reused; and if same_gpu is on, the service must be serving parallel=1 or you get "Same-GPU mode requires parallel=1". One more: same-GPU handoff currently requires the verified NVIDIA path - on Windows AMD, set vision:no and keep same_gpu:false.

Worth saying out loud that this author's whole bias is memory hygiene - T8mars also ships comfyui-purgevram, a node people recommend in "how do I unload models from VRAM" threads. Same instinct here: everything in this pack assumes the engine leaves before the sampler arrives.