Skip to main content
by artokun · August 21, 2026 · kitchen · fp8 · NVFP4 · ComfyUI The agent could not even say whether comfy-kitchen was present. Users asked it to make their graph faster with kitchen; it guessed. kitchen and panel_kitchen are the see / assess / apply tools for that, and apply is not done until a before/after run on the same seed says so — or until the widget is reverted. Install comfyui-mcp and the Panel — then start the agent with npx -y comfyui-mcp@latest connect and click Connect in the panel (no API keys; sign in with claude once).

​
What a panel_kitchen apply report is

The deliverable is a line like this, filled from that run, not from a canned bench:
If the after-run is black, NaN, or slower, the widget is reverted and the same line is the failure report. Numbers in docs are the shape of that report. The values come off panel_kitchen apply on your graph. Call kitchen action:“status” first if you are not on the canvas; panel_kitchen if you are.

​
What kitchen actually is

comfy-kitchen 0.2.31 (PyPI and GitHub tag v0.2.31; ComfyUI master pins that version) is the kernel library: fp8 / NVFP4 / MXFP8 / INT8 tensors and GEMMs, fused RoPE and AdaLN, INT8 attention, backends hip / cuda / triton / eager. A June ComfyUI checkout still pins 0.2.10 and lacks --use-ck-attention. status reports ComfyUI-side flag support, not only the kitchen version. ComfyUI wires it in at four places the tool reads:
  • Startup log: comfy-kitchen version: X, Found comfy_kitchen backend {name}: {status}, Found triton X. Enabling comfy-kitchen triton backend., Using Comfy Kitchen attention.
  • Flags: --use-ck-attention (INT8 attention, no sageattention wheel), --enable-triton-backend / --disable-triton-backend, --fast fp8_matrix_mult, --fast cublas_ops. These are ComfyUI master flags. --enable-triton-backend is not SwarmUI-only.
  • Hardware: fp8 compute at SM ≥ 8.9 (Ada), NVFP4 / MXFP8 at SM ≥ 10.0 (Blackwell).
  • Quantized checkpoints: safetensors _quantization_metadata / .comfy_quant in fp8_e4m3fn / fp8_e5m2 / nvfp4 / mxfp8. UNETLoader.weight_dtype fp8_e4m3fn_fast turns on fp8_optimizations.

​
Three actions

​
status

Kitchen version, backends loaded, INT8 attention, GPU fp8 / NVFP4 / MXFP8, launch flags. Remote (--comfyui-url, Comfy Cloud) gets log + /system_stats only; the import probe and model.quant are unknown. A failed probe is unknown, never a no.

​
assess

Walks UNETLoaders. A rec fires only when every input is known:
  1. weight_dtype: default on a bf16 UNETLoader, GPU has fp8, kitchen present → fp8_e4m3fn_fast. Reversible widget, no restart.
  2. --use-sage-attention absent, sageattention not installed, kitchen INT8 available → --use-ck-attention. Restart, confirm.
  3. Blackwell GPU, model in bf16/fp8, an NVFP4 sibling already on disk → swap. No URL guessed from the filename.
  4. ROCm host, kitchen present, triton ≥ 3.7, triton backend off → --enable-triton-backend. Restart.
MXFP8 is reported in status, not recommended, until a loader path exposes it.

​
apply

One recommendation_id from assess. Widget edits do not need confirm. Flags and downloads need confirm: true. Flag apply names the launch flag; restart_comfyui / panel_restart_comfyui replay the previous argv and do not inject a new one — add the flag to the launch line, then restart. panel_kitchen apply then queues a panel_run on the current seed and reverts on black / NaN / slower.

​
INT8 attention without sageattention

The triton-sageattention skill used to send you through a wheel-matched install. If kitchen action:“status” says INT8 attention is available, use --use-ck-attention and skip that dance. See the kitchen usage page.