Interface: GpuBudget

Defined in: src/batteries/llm/chat_common/gpu_budget.ts:37

A snapshot of the WebGPU device's memory budget, as reported by the adapter/device limits.

Remarks

All sizes are bytes. maxBufferSize and maxStorageBufferBindingSize are the hard per-allocation ceilings a single ONNX tensor buffer cannot exceed — the practical wall an over-large context window hits first (e.g. ~4 GiB on Apple Metal). These are LIMITS, not live usage; pair with GpuBufferInstrument for live high-water-mark tracking.

Properties

Property	Type	Description	Defined in
adapterInfo?	{ architecture?: string; description?: string; device?: string; vendor?: string; }	Best-effort adapter description (vendor/architecture/device), when the runtime exposes it.	src/batteries/llm/chat_common/gpu_budget.ts:43
adapterInfo.architecture?	string	-	src/batteries/llm/chat_common/gpu_budget.ts:43
adapterInfo.description?	string	-	src/batteries/llm/chat_common/gpu_budget.ts:43
adapterInfo.device?	string	-	src/batteries/llm/chat_common/gpu_budget.ts:43
adapterInfo.vendor?	string	-	src/batteries/llm/chat_common/gpu_budget.ts:43
available	boolean	true when these numbers came from a real navigator.gpu device; false/absent ⇒ unavailable.	src/batteries/llm/chat_common/gpu_budget.ts:45
maxBufferBytes	number	Largest single GPUBuffer the device will allocate, in bytes (GPUSupportedLimits.maxBufferSize).	src/batteries/llm/chat_common/gpu_budget.ts:39
maxStorageBufferBindingBytes	number	Largest storage-buffer binding, in bytes (GPUSupportedLimits.maxStorageBufferBindingSize).	src/batteries/llm/chat_common/gpu_budget.ts:41