MarkBase Admin
ac75faa0cc
CI / build-and-test (push) Has been cancelled
- E4B-MarkBase model (42 layers, 4.4GB) loaded successfully - All Phase 1-6 tests passed (model loading, forward pass, vision/audio towers, token generation, performance) - All stress tests passed (5/5 in 127.6s) - Concurrent inference - Memory stress (67.5 tok/s, 0 NaN) - Continuous generation - Batch processing - Long-running stability - Swift Metal inference engine with multimodal support
39 lines
1.4 KiB
Swift
39 lines
1.4 KiB
Swift
/// Metadata for a single tensor stored in a SafeTensors file.
|
|
public struct TensorDescriptor: Sendable, Codable {
|
|
public let name: String
|
|
public let dtype: TensorDType
|
|
public let shape: [Int]
|
|
/// Byte offset from the start of the safetensors data section.
|
|
public let dataOffset: Int
|
|
/// Byte size of the tensor data.
|
|
public let dataSize: Int
|
|
|
|
/// Total number of elements.
|
|
public var elementCount: Int { shape.reduce(1, *) }
|
|
|
|
/// Check if shape is compatible with a given dim count.
|
|
public func hasRank(_ rank: Int) -> Bool { shape.count == rank }
|
|
|
|
/// For quantized tensors: returns the grouping factor (elements per group).
|
|
/// MLX default: 64 elements per quantization group (for Gemma 4 E4B 4-bit).
|
|
public var quantizationGroupSize: Int { 64 }
|
|
}
|
|
|
|
/// Group of tensors that together represent a quantized linear layer.
|
|
/// weight: U32 packed (shape: [outDim, inDim / 32 * 4])
|
|
/// scales: BF16 (shape: [outDim, inDim / 32])
|
|
/// biases: BF16 (shape: [outDim, inDim / 32])
|
|
public struct QuantizedTensorGroup: Sendable {
|
|
public let name: String
|
|
public let weight: TensorDescriptor
|
|
public let scales: TensorDescriptor
|
|
public let biases: TensorDescriptor
|
|
|
|
/// Output dimension.
|
|
public var outDim: Int { weight.shape[0] }
|
|
/// Input dimension (pre-quantization).
|
|
public var inDim: Int { scales.shape[1] * 32 }
|
|
/// Block size (elements per group).
|
|
public let groupSize: Int = 64
|
|
}
|