Load Phi

This is the node you reach for when you want an actual language model running in your graph, not just a text-encoder that makes embeddings. Load Phi spins up Microsoft's Phi-3.5-mini-instruct - a 3.8B-parameter chat model - and hands you a ready-to-run model plus its tokenizer. The obvious use is what most people actually do with Phi in ComfyUI: prompt generation and rewriting. Wire its output into the right place and you've got a node that takes your rough idea ("a cyberpunk bakery, moody lighting, close-up") and hands back a polished prompt you can feed a checkpoint. No API, no key, no cloud call - it's all running on your GPU.

The name is honest, which is rarer than it should be: this node only loads. Generation happens in its sibling, Run Phi. Think of the Load/Run pair as a ComfyUI-flavored version of "model + sampler" - load once, keep it in memory, run as many generations as you like without re-reading weights.

How it works

Peek at the source and it's a straight HuggingFace load. The node resolves your ComfyUI/models/microsoft/ folder, builds the path to the model directory, and calls AutoModelForCausalLM.from_pretrained with device_map="cuda", torch_dtype="auto", and trust_remote_code=True. The tokenizer comes from AutoTokenizer. Everything is local_files_only=True - a detail that matters more than it looks like it should, because it means there is no auto-download. If the model folder isn't there, the node fails, and the failure message will not hold your hand.

Inputs and outputs

Only one input you can set:

• model - a dropdown that lists exactly one choice, Phi-3.5-mini-instruct. It's an enum with a single entry, so don't go looking for a model-list scan like other loaders give you.

The node emits two outputs:

• phi_model - the loaded model object.
• phi_tokenizer - its tokenizer.

Both feed into Run Phi and nothing else; these are custom types (phi_model, phi_tokenizer) that only this pack understands.

Installing it

Install the pack either way you like:

# Option A: ComfyUI Manager → search "Phi" → install
# Option B: manual
cd ComfyUI/custom_nodes
git clone https://github.com/alexisrolland/ComfyUI-Phi.git
../../(your python) -m pip install -r ComfyUI-Phi/requirements.txt

Then the part that trips people: download the model yourself into a folder that ComfyUI registers as microsoft:

cd ComfyUI/models/microsoft          # create it if missing
git clone https://huggingface.co/microsoft/Phi-3.5-mini-instruct

The folder name must be exactly Phi-3.5-mini-instruct - the node builds the path from the enum, so a renamed folder breaks the load. The pack's requirements are light (accelerate, backoff, peft); notably transformers is not pinned anymore, so it borrows whatever ComfyUI already has - fine, but it's the classic no-isolation situation where another pack's upgrade can quietly break this one.

Where people get burned

• No auto-download. local_files_only=True means a missing folder = instant error. The README is explicit about manual download; ignore it and you'll stare at a red node.
• VRAM. 3.8B params in fp16 is roughly 7–8GB of weights sitting in memory while you also try to hold a diffusion checkpoint. On a 12GB card this is a squeeze; it's one reason many people run Phi via GGUF/llama.cpp instead for text-only jobs. This pack is not a GGUF loader - it wants the full HuggingFace safetensors repo.
• CUDA or nothing. device_map="cuda" means no CPU fallback; a machine without a usable GPU won't run this at all.
• trust_remote_code. The model loads custom modeling code from the repo. The pack's README recommends the forked Lexius/Phi-3.5-mini-instruct because the originals have known issues - worth knowing before you pull the plain microsoft repo and hit a weird generation bug downstream.