M8 · 大模型推理

An LLM in the graph is a tool, not an encoder. It's not the T5 or Qwen3 inside your checkpoint doing the forward pass - it's a node you bolt on that rewrites a rough idea into a structured prompt, or turns a caption into tags, before the text ever reaches the sampler. M8 · LLM Inference is the API half of that idea: any OpenAI-compatible endpoint, your key, your call.

Why you'd wire an LLM into a workflow at all

Because the 2026 architecture rewards it. When the checkpoint's text encoder is itself a language model reading an instruction, having a second language model write that instruction is a translation between two things that speak the same language. The job is short structured rewriting, which is why you want something small and obedient - not a reasoning model that deliberates and leaks its scratch-work into your prompt. The honest split in this pack: this node for a frontier endpoint you'll pay for, M8 · Local LLM Inference for a local GGUF that's free per call and won't refuse anything.

How it works

Plain OpenAI chat-completions. You pick a provider - deepseek, openai, openrouter, siliconflow, ollama, custom - and it fills that provider's default base_url and its thinking options. Default is DeepSeek. The model dropdown is empty until you hit Refresh models, which asks the endpoint for its list; you can also type a model name by hand, which is what you do for Ollama or a model released after the list was fetched.

Two prompt boxes:

• system_prompt - persona, tone, output format
• user_prompt - this turn's actual question, and it can be converted to an input socket (right-click) so an upstream node's text feeds it

thinking runs off/low/medium/high. It's a provider-specific parameter, and if the endpoint rejects it the node drops the field and retries once rather than dying - the log tells you it happened. show_thinking renders the model's reasoning in the panel only; it never leaves the node. That's the right call: wire reasoning text into a prompt encoder and you've fed it a monologue.

The output is exactly one thing, text (STRING). temperature defaults to 1.0 - drop it to 0.3–0.6 for prompt writing, which is steady work, not creative work. max_tokens is a cap, not a target. timeout defaults to 120 seconds; if you're using a thinking model, give it 300 or more or you'll be timing out on answers that were about to arrive.

Skills, images, audio, and the escape hatch

Wire M8 · Skill Loader into skill and its text is appended after the system prompt, wrapped so the model can tell instructions from reference material. You can also type / in the prompt box to reference an uploaded skill by name - the backend injects it only when the reference resolves to exactly one skill. Ambiguous or unmatched means nothing gets injected, which is deliberate: guessing wrong dumps the wrong reference into your context, which is worse than omitting it. URLs, drive paths and things like 3/4 aren't treated as references.

skill_auto (off by default) injects every uploaded skill and lets the model pick. Leave it off unless your library is tiny - every turn then carries a few irrelevant packs and burns tokens on them.

image and audio are optional multimodal inputs; audio gets converted to WAV. Feed them to a model that supports them or the endpoint returns 400 (M8-LLM-010). extra_params takes a JSON object merged straight into the request body - {"top_p": 0.9}, or any vendor-specific field this pack doesn't know about.

The key handling is the good part

Two ways to give it a key, and they are not equivalent. Typing it on the node works, and gets written into the workflow file - hand that workflow to someone and you've handed them your key. The other button saves it server-side into m8/data/credentials.json, clears the field, and leaves the workflow clean.

There's a guardrail worth knowing about: the stored key is only sent when base_url matches that provider's default or saved address. A shared workflow can set base_url to someone else's server with the key field blank, and the node will refuse with M8-LLM-021 rather than post your credential to a stranger's endpoint.

Install and the errors you'll meet

Same as the rest of the pack - ComfyUI Manager, or git clone https://github.com/chenr5934-tech/ComfyUI-M8nodes.git into custom_nodes and restart. No pip step.

When something's wrong you get a code, not a shrug: M8-LLM-004 for a non-200 response (401 means key, 402 balance, 404 wrong address, 429 rate limit), M8-LLM-001 for no key configured, M8-LLM-013 when the request body is too big - usually too many images or a fat skill pack, and the node warns in the log when a pack passes 20,000 characters. M8-LLM-005 means the endpoint answered but the response carried no text, which almost always means the model name is wrong or the endpoint isn't really OpenAI-compatible.

One last thing, and it isn't paranoia. An API node holds a credential and calls the network by design, which is why the worst security incident in this ecosystem's history was an LLM-vision node that shipped credential-stealing malware. Prefer keys saved server-side over keys baked into the workflow.