Support custom layer splits from GPUShardingSpec.LayerSplit

[Defilan]

Context

internal/controller/inferenceservice_controller.go has a calculateTensorSplit() function that ignores the sharding.LayerSplit field from the CRD:

// TODO: Support custom layer splits from sharding.LayerSplit
func calculateTensorSplit(gpuCount int32, _ *GPUShardingSpec) string {

Currently all GPUs get equal split ratios ("1,1" for 2 GPUs, "1,1,1" for 3, etc.) regardless of what the user specifies in spec.hardware.gpu.sharding.layerSplit.

Problem

Users with asymmetric GPU configurations (e.g., mixed VRAM sizes) cannot control how layers are distributed. The layerSplit field exists in the CRD but is silently ignored.

Proposed Solution

• Parse sharding.LayerSplit (e.g., ["0-15", "16-31"]) into llama.cpp --tensor-split ratios
• Fall back to equal split when LayerSplit is empty (current behavior)
• Validate layer ranges don't overlap and cover the full model

Location

internal/controller/inferenceservice_controller.go:660