Method estimateModelVRAM

core/services/nodes/router.go:862–900

estimateModelVRAM estimates the VRAM required for a model using the unified estimator.

(ctx context.Context, opts *pb.ModelOptions)

// estimateModelVRAM estimates the VRAM required for a model using the unified estimator.
func (r *SmartRouter) estimateModelVRAM(ctx context.Context, opts *pb.ModelOptions) uint64 {
	estCtx, cancel := context.WithTimeout(ctx, 10*time.Second)
	defer cancel()

	ctxSize := uint32(opts.ContextSize)
	if ctxSize == 0 {
		ctxSize = 8192
	}

	input := vram.ModelEstimateInput{
		Options: vram.EstimateOptions{
			GPULayers: int(opts.NGPULayers),
		},
	}

	// Try model file as a local file for GGUF metadata estimation
	if opts.ModelFile != "" {
		if _, err := os.Stat(opts.ModelFile); err == nil {
			input.Files = append(input.Files, vram.FileInput{URI: opts.ModelFile, Size: 0})
		}
	}

	// Try HF repo from model name (e.g. "org/model")
	if opts.Model != "" {
		if repoID, ok := vram.ExtractHFRepoID(opts.Model); ok {
			input.HFRepo = repoID
		}
	}

	if len(input.Files) == 0 && input.HFRepo == "" && input.Size == "" {
		return 0
	}

	result, err := vram.EstimateModelMultiContext(estCtx, input, []uint32{ctxSize})
	if err != nil {
		return 0
	}
	return result.VRAMForContext(ctxSize)
}

// installBackendOnNode sends a NATS backend.install request-reply to the node
// and returns the gRPC address. Concurrent identical calls (same nodeID +