kitchen and panel_kitchen

comfy-kitchen is ComfyUI’s kernel library — fp8 / NVFP4 / MXFP8 / INT8 quantized tensors and GEMMs, fused RoPE and AdaLN, and its own INT8 attention, with hip / cuda / triton / eager backends. It is not a workflow library. Two tools share the same actions:

kitchen { action: "status" | "assess" | "apply" } — MCP clients. assess takes a workflow JSON.
panel_kitchen — the same actions against the live canvas, fenced like every other panel mutation.

A failed probe is unknown, never a no. A recommendation fires only when every fact it needs is known.

status — what this GPU can run

status reports:

Field	Meaning of `unknown`
`present`	No version line, no `comfy_package_versions[comfy-kitchen]`, and no import probe. Not a report that kitchen is missing.
`version`	Kitchen is present but the version token was not in the log, packages, or probe.
`backends.hip/cuda/triton/eager`	No `Found comfy_kitchen backend …` line and no `ck.list_backends()` probe.
`attention.ck`	Need both `--use-ck-attention` in argv and `Using Comfy Kitchen attention` in the log. A missing log line is not a no.
`gpu.fp8`	SM ≥ 8.9 (Ada). Unknown when the GPU name is not in the known table and there is no compute-capability probe.
`gpu.nvfp4` / `gpu.mxfp8`	SM ≥ 10.0 (Blackwell). Same unknown rule.
`gpu.int8Attention`	`ck.int8_attention_is_available()` was not probed (remote, no python, or timeout).
`flagSupport.ckAttention`	`comfy/cli_args.py` was not readable and argv does not already carry `--use-ck-attention`. A June ComfyUI checkout pins kitchen 0.2.10 and lacks the flag.
`model.quant`	Remote session, or `COMFYUI_PATH` unset, or the safetensors header was unreadable. Filename is not used as a substitute.

Local vs remote. --comfyui-url / Comfy Cloud gets log + /system_stats only. The import probe and model.quant stay unknown there; the tool says so rather than guessing from the filename.

assess — four recommendations

assess walks UNETLoaders. Each rec is a conjunction:

id prefix	Facts that must all be known	Change
`fp8_unet_fast:<node>`	kitchen present, GPU fp8, `weight_dtype` is `default`, file quant is unquantized (`none`)	set `weight_dtype` to `fp8_e4m3fn_fast`. Widget, no restart.
`ck_attention`	`--use-sage-attention` absent, sageattention not installed (a log import-failure or a trusted probe, not silence), kitchen present, INT8 available, flag not already on	add `--use-ck-attention`. Restart, `confirm: true`.
`nvfp4_swap:<node>`	GPU nvfp4, loader is bf16/fp8, an NVFP4 sibling of that filename is already listed locally	swap `unet_name`. No download guess.
`triton_backend`	ROCm host, kitchen present, triton ≥ 3.7, `--enable-triton-backend` not on argv	add `--enable-triton-backend`. Restart, `confirm: true`.

MXFP8, SVDQuant, and AWQ appear in status but are not recommended until a loader path in ComfyUI exposes them.

apply — change, then prove or revert

Widget recs (fp8_unet_fast, nvfp4_swap) do not need confirm. kitchen mutates the supplied workflow JSON. panel_kitchen writes through graph_set_widget (Ctrl+Z).
Flag recs need confirm: true. restart_comfyui / panel_restart_comfyui replay the previous argv and do not inject a new flag. Apply names the flag and tells you to add it to the launch line, then restart. That is not a new restart path.
Proof. panel_kitchen apply queues a panel_run on the current seed after a widget change. The report shape is s/it A → B, peak VRAM C → D GB, output non-black, N widget changed, M flags changed. A black or NaN after-run, or a slower after-run, reverts the widget. kitchen apply without a runner reports proof.status: "not_run" rather than claiming a bench it did not run.
Pass skip_proof: true on panel_kitchen to apply without queueing a run.

Proactive hint

Once per process, when kitchen is present and rec 1 or rec 2 is certain, status / assess includes one sentence naming the gap and the tool. Nothing runs unasked.

Flags this tool can name

--use-ck-attention, --enable-triton-backend, --disable-triton-backend, --fast fp8_matrix_mult, --fast cublas_ops. On master these are ComfyUI flags. --enable-triton-backend is not SwarmUI-only.

Getting Started

Configuration

Tools

kitchen and panel_kitchen

kitchen and panel_kitchen

status — what this GPU can run

assess — four recommendations

apply — change, then prove or revert

Proactive hint

Flags this tool can name

​kitchen and panel_kitchen

​status — what this GPU can run

​assess — four recommendations

​apply — change, then prove or revert

​Proactive hint

​Flags this tool can name

kitchen and panel_kitchen

status — what this GPU can run

assess — four recommendations

apply — change, then prove or revert

Proactive hint

Flags this tool can name