launch

type Config struct {
	// BinPath is the gated runtime server executable to run (an installed build).
	BinPath string
	// WeightsPath is the verified GGUF weights file on disk to load.
	WeightsPath string
	// Model is the catalog entry being served; its ChatTemplate is the trusted contract.
	Model catalog.ModelSpec
	// Format is the weight encoding being served, which selects the runtime engine. Empty
	// is treated as GGUF/llama.cpp so existing callers keep their behavior.
	Format catalog.Format
	// Port is the loopback TCP port the server listens on. The caller picks a free one.
	Port int
	// CtxSize is the context window to run with; 0 lets the runtime use the model default.
	CtxSize int
	// CPUOnly forces the model to run entirely on the CPU with no GPU offload, the last-resort
	// footprint for a model that will not fit in device memory. The default offloads to the GPU
	// as the runtime sees fit. It overrides GPULayers and MoECPULayers.
	CPUOnly bool
	// GPULayers, when positive, is the number of model layers offloaded to the GPU
	// (llama.cpp -ngl). Zero leaves the offload to the runtime's default. Ignored when
	// CPUOnly is set.
	GPULayers int
	// MoECPULayers, when positive, keeps that many mixture-of-experts layers on the CPU
	// while the rest of the model stays on the GPU (llama.cpp -ncmoe). This is what lets a
	// large-total, small-active MoE run on a modest GPU: hot weights in VRAM, the cold
	// expert pages in system RAM. Ignored when CPUOnly is set.
	MoECPULayers int
	// KVCacheType, when set, quantizes the KV cache to this element type (llama.cpp
	// --cache-type-k/--cache-type-v), trading a little quality for room to hold more
	// context in the same VRAM. It must be a recognized type or the plan is refused.
	KVCacheType string
	// DraftWeightsPath, when set, names a small draft model the runtime uses for
	// speculative decoding: the draft proposes tokens the primary verifies in a batch,
	// raising throughput at no quality cost. Empty disables speculation.
	DraftWeightsPath string
	// DraftMax and DraftMin bound how many tokens the draft proposes per step (llama.cpp
	// --draft-max/--draft-min). They apply only when DraftWeightsPath is set; zero leaves
	// the runtime default.
	DraftMax int
	DraftMin int
	// Quantization, when set, names the on-disk quantization scheme the vLLM engine loads
	// the weights with (vLLM --quantization), for example "awq", "fp8", or "modelopt_fp4"
	// for an NVFP4 checkpoint. It must be a recognized scheme or the plan is refused. It
	// applies only to the vLLM engine; the llama.cpp engine reads the quantization from
	// the GGUF file itself.
	Quantization string
	// GPUMemoryUtilization, when positive, is the fraction of GPU memory vLLM may reserve
	// for weights and the KV cache (vLLM --gpu-memory-utilization), a value in (0,1]. Zero
	// leaves vLLM's default. vLLM engine only.
	GPUMemoryUtilization float64
	// KVCacheDtype, when set, quantizes the vLLM KV cache to this element type (vLLM
	// --kv-cache-dtype), for example "fp8". It must be a recognized vLLM type or the plan
	// is refused. This is the vLLM counterpart to KVCacheType, which is llama.cpp's; a plan
	// uses whichever matches its engine. vLLM engine only.
	KVCacheDtype string
	// Containerized records that the vLLM server runs inside a container, so it must bind
	// all interfaces inside the container's isolated network namespace for the engine's
	// loopback-only published port to reach it. The off-host exposure stays closed: the
	// container tier publishes the port bound to the host's loopback only, so binding
	// 0.0.0.0 inside the container is reachable from the host loopback and nowhere else, and
	// the client still targets the loopback BaseURL. A non-containerized server binds
	// loopback directly. vLLM engine only.
	Containerized bool
	// APIKey, when set, is required by the server on every request, so even a local
	// process cannot reach the model without the token the caller holds. Optional.
	APIKey string
	// ModelEmbedsTemplate records that the weights carry their own chat template, a
	// signal from inspecting the file. The plan overrides it regardless; the field only
	// surfaces that an override was necessary. Optional.
	ModelEmbedsTemplate bool
}

type Engine string

type Plan struct {
	// Argv is the full command: the binary followed by its flags, ready to execute.
	Argv []string
	// Host and Port are the loopback address the server binds.
	Host string
	Port int
	// BaseURL is the OpenAI-compatible endpoint a client targets once the server is up.
	BaseURL string
	// TemplateOverridden is true when the model shipped its own chat template that this
	// plan replaced with the trusted one.
	TemplateOverridden bool
}