denv.yaml

One file, checked into your repository. denv recipe validate reads recipes; denv release reads release; denv env apply reads environment. Providers, fleets and customers belong to your organization and are not in this file.

Full example

recipes:
  - name: ds-flash-h200
    version: 1
    extends: deepseek-v4.1-flash/sglang@2
    hardware: [h200]
    engine: { name: sglang, version: 0.6.1 }
    args: { tp: 8, ep: 8, enable-dp-attention: true }

release:
  model: ds-flash
  weights: hf://deepseek-ai/DeepSeek-V4.1-Flash-1017@c41a7e2
  recipe: deepseek-v4.1-flash/deepinfer@3
  topology:
    mode: disaggregated
    prefill: { gpus: 4, replicas: 2 }
    decode: { gpus: 64, replicas: 1 }
    placement: same-domain
  kv-cache:
    tiers: [gpu, host-memory, nvme]
    share: stage
  api: [chat-completions, responses, messages]
  health:
    startup-timeout: 30m
    warmup: 64
    golden: tests/golden.jsonl
  parity:
    reference: { url: https://api.deepseek.com, secret: deepseek-api-key, model: deepseek-v4.1-flash }
    suite: parity/flash.jsonl
    pass: { tool-calls: 99%, exact: 95% }

environment:
  name: ds-flash
  stages:
    - name: canary
      fleets: [gb200-rack]
      replicas: 1
      promote: { bake: 1h }

    - name: prod-us
      fleets: [gb200-rack, h200-use1]
      recipes: { h200-use1: ds-flash-h200@1 }
      replicas: { min: 2, max: 6 }
      routing: { policy: priority, order: [gb200-rack, h200-use1] }
      rollout: { batch: 1, min-healthy: 75%, bake: 2h }
      alarms: { ttft-p99: 1500ms, error-rate: 1%, throughput-drop: 15%, parity-fail: 1 }
      promote: { approval: required }

recipes[]

Your own recipes. Catalog recipes (denv recipe list) need no entry here.

Field	Required	Meaning
name, version	yes	A recipe is referenced as name@version. Versions are immutable once used by a release.
extends	no	Start from another recipe and override only what differs.
models	no	Model families this recipe accepts, such as deepseek-v4.1. Inherited when extends is set.
engine	yes	{ name, version }. name is vllm, sglang, tensorrt-llm or deepinfer.
hardware	yes	Hardware ids this recipe is tested on: h100, h200, b200, b300, gb200-nvl72, gb300-nvl72, mi300x, mi355x, ascend-950, l40s. See Hardware.
topology	no	Default topology; a release may override it.
args	no	Engine arguments, validated against the engine version.
adapter	no	API adapter that turns requests into prompts and output into responses: engine (the engine's own OpenAI server) or a model-specific adapter such as deepseek-recipe. Default from extends or engine.
parity	no	Default parity suite for the model family.

release

Field	Required	Meaning
model	yes	Name shared by every release of this model; also the model value callers send.
weights	yes	Pinned: hf://org/repo@revision, s3://bucket/path@<versionId>, gs://bucket/path#<generation>, or denv://<model>/<version>. A branch is resolved to a commit when the release is created.
weights-auth	no	Secret used to read private weights.
recipe	yes	name@version, from the catalog or recipes.
overrides	no	Any recipe field (engine, args, adapter) changed for this release only.
topology.mode	no	aggregated or disaggregated. Default from the recipe.
topology.gpus, topology.replicas	aggregated	GPUs per replica and replica count per stage, unless the stage sets replicas.
topology.prefill, topology.decode	disaggregated	{ gpus, replicas } for each pool.
topology.placement	no	same-domain keeps prefill and decode of one serving unit inside one scale-up domain (default for disaggregated); any allows crossing domains.
kv-cache.tiers	no	Ordered tiers: gpu, host-memory, nvme. Default [gpu, host-memory].
kv-cache.share	no	replica (no sharing) or stage (prefixes shared across the stage's replicas). Default stage.
adapters	no	LoRA adapters, each pinned like weights. Callers select one with model: <model>:<adapter>.
api	no	Surfaces to expose: chat-completions, responses, messages, completions, embeddings. Default [chat-completions, responses].
health	no	startup-timeout, warmup request count, golden prompts file. See Rollout policy.
parity	no	reference (url, secret, model), suite file (JSONL; see Run parity checks), pass thresholds as per-kind pass shares (exact, tool-calls, logprobs). See Parity.

The release id (v8) and digest are assigned when it is created. The digest covers the resolved weights revision, recipe version, resolved engine image digest, overrides, topology, KV cache settings, API list, health and parity configuration.

environment

Field	Required	Meaning
name	yes	Lowercase, digits, hyphens.
stages	yes	Ordered; a release is promoted from each stage to the next.

stages[]

Field	Required	Meaning
name	yes	Part of the endpoint URL.
fleets	yes	Fleets this stage serves from. Each must offer hardware the release's recipe supports.
recipes	no	Per-fleet recipe overrides, for serving the same weights on different hardware: { <fleet>: <recipe>@<v> }.
replicas	yes	A number or { min, max }. For disaggregated releases, a replica is one serving unit (its prefill and decode pools).
routing	no	{ policy: priority, order }, { policy: weighted, weights }, { policy: latency }, { policy: cost }. Default latency. KV cache locality is applied within each policy.
rollout	no	batch, min-healthy, bake, on-alarm: rollback | pause. See Rollout policy.
alarms	no	ttft-p99, itl-p99 (inter-token latency), error-rate, throughput-drop, parity-fail, and custom: [{ metric, above | below, for }].
promote	no	bake, approval: required, windows, auto: true.
residency	no	eu, us, or a list of regions. Fleets outside it are refused; the router never fails over outside it.

Validation

--dry-run on denv recipe validate, denv release and denv env apply checks without changing anything: weights resolve and are pinned, the engine version supports the hardware and arguments, every stage's fleets can fit the topology (including domain size for same-domain), the parity reference is reachable, residency holds.