Reference

Profiles

A profile is YAML laid over warmctl's defaults. Only schema, id, model, parallel and hardware are needed; every other key has a default.

qwen3-8b-h100-tp1.yaml

schema: warmctl.profile/v1
id: qwen3-8b-h100-tp1
model: {id: Qwen/Qwen3-8B, revision: b968826d…, served_names: [qwen3-8b]}
parallel: {tp: 1}
hardware: [h100-80gb-sxm]
vllm:
  max-model-len: 32768
  max-num-seqs: 64
  kv-cache-memory-bytes: 47244640256
  enable-chunked-prefill: true
graphs:
  mode: FULL_AND_PIECEWISE
  decode_batches: [1, 2, 4, 8, 16, 24, 32, 48, 64]
warmup:
  decode_prompt_tokens: 32
  prefill_tokens: [512, 2048, 8192]
resources: {memory: 96GiB}

The catalog's profiles ship in the binary: warmctl profile list. A key set to null removes the default. build --set vllm.<flag>=…, env.<NAME>=… or ipc=… overrides a key for one build.

Keys

Key	Default
`schema`		`warmctl.profile/v1`
`id`		the profile's name
`image`	pinned vLLM image	the engine's container image, by digest
`model.id`, `.revision`		Hugging Face id and the revision the weights are
`model.served_names`		names the API answers to
`model.container_path`		where the weights are mounted in the container
`parallel`		`tp`, and `dp`, `ep` for data and expert parallel, all within one host
`hardware`		the hardware ids the profile is meant for
`vllm`	see below	any `vllm serve` flag, without the dashes in front
`env`	see below	environment of the engine; `null` unsets a default
`graphs.mode`	FULL_DECODE_ONLY	CUDA graph mode: `FULL_DECODE_ONLY`, `FULL_AND_PIECEWISE` or `PIECEWISE`
`graphs.decode_batches`	[1, 2, 4, 8]	batch sizes captured as graphs
`graphs.compile_mode`	3	vLLM compilation level
`warmup.prefill_tokens`		prompt lengths warmed before the snapshot
`warmup.decode_prompt_tokens`		prompt length for the decode batches
`warmup.generated_tokens`	32	tokens generated per warmup request
`canary.mode`	warn	`bitwise`, `tolerance` or `warn`: how strictly a restore must match the donor
`canary.tolerance`	0.001	for `tolerance`
`canary.prefix_tokens`	512	length of the canary prompt
`resources`	64GiB · 16GiB · 16	`memory`, `shm`, `cpus`, `pids` of the container
`restore.warmup`	canary	what runs after a restore: `none`, `canary` or `full`
`restore.cuda_restore`	auto	`auto`, `serial` or `parallel`
`ipc`	auto	cross-process GPU memory: `auto`, `interpose` (NVIDIA cuInterpose) or `native`

Defaults that matter

The engine always runs with what makes a snapshot possible: enable-sleep-mode, enable-nccl-comm-suspend and warmctl's worker extension. It also gets max-model-len: 2048 and 1 GiB of KV cache, so a serving profile sets both. Its compile and JIT caches live under /workspace, which the snapshot carries, and Hugging Face runs offline. warmctl profile show ID prints everything a profile resolves to.

Hardware ids

Id	GPU
`h100-80gb-sxm`	NVIDIA H100 80GB HBM3, NVLink
`h200-141gb`	NVIDIA H200
`b200-180gb`	NVIDIA B200
`l40s-48gb`	NVIDIA L40S, PCIe

A snapshot's platform tag is the GPU count, the hardware id and the driver it was built with: 8xh100-80gb-sxm-580.173.02. Platforms and drivers explains why.