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
91 lines
5.1 KiB
Swift
91 lines
5.1 KiB
Swift
import XCTest
|
|
@testable import MarkBase
|
|
|
|
class ModelScalesComparisonTest: XCTestCase {
|
|
|
|
func testCompareAllModelScales() throws {
|
|
print("\n═══════════════════════════════════════════════════════════════════")
|
|
print(" 26B-Standard vs 26B-A4B vs 31B Scales Comparison")
|
|
print("═══════════════════════════════════════════════════════════════════\n")
|
|
|
|
// 26B-Standard (single file, custom quant)
|
|
print("1. 26B-Standard:")
|
|
let standardPath = "/Users/accusys/MarkBaseEngine/models/gemma-4-26b-standard"
|
|
let standardReader = try SafeTensorsReader(path: "\(standardPath)/model.safetensors")
|
|
let standardTensors = standardReader.allTensors
|
|
|
|
let standardScales = standardTensors.first { $0.name.contains("embed_tokens.scales") }
|
|
if let s = standardScales {
|
|
let data = try standardReader.read(tensor: s)
|
|
let scales = data.withUnsafeBytes { ptr in
|
|
Array(ptr.assumingMemoryBound(to: Float.self).prefix(20))
|
|
}
|
|
print(" Scales shape: \(s.shape), dtype: \(s.dtype)")
|
|
print(" Sample: \(scales)")
|
|
let nanCount = scales.filter { $0.isNaN }.count
|
|
let negCount = scales.filter { $0 < 0 }.count
|
|
print(" NaN: \(nanCount), Negative: \(negCount)")
|
|
}
|
|
|
|
// 26B-A4B (sharded, MLX-vlm 0.4.3)
|
|
print("\n2. 26B-A4B:")
|
|
let a4bPath = "/Users/accusys/MarkBaseEngine/models/gemma-4-26b-a4b-it-4bit"
|
|
let a4bIndex = try SafeTensorsIndex(modelDir: a4bPath)
|
|
var a4bReaders: [String: SafeTensorsReader] = [:]
|
|
for shardFile in a4bIndex.weightMap.values {
|
|
if a4bReaders[shardFile] == nil {
|
|
a4bReaders[shardFile] = try SafeTensorsReader(path: "\(a4bPath)/\(shardFile)")
|
|
}
|
|
}
|
|
let a4bTensors = a4bReaders.values.flatMap { $0.allTensors }
|
|
|
|
let a4bScales = a4bTensors.first { $0.name.contains("embed_tokens.scales") }
|
|
if let s = a4bScales {
|
|
let reader = a4bReaders[a4bIndex.weightMap[s.name] ?? "model-00001-of-00003.safetensors"]!
|
|
let data = try reader.read(tensor: s)
|
|
let scales = data.withUnsafeBytes { ptr in
|
|
Array(ptr.assumingMemoryBound(to: Float.self).prefix(20))
|
|
}
|
|
print(" Scales shape: \(s.shape), dtype: \(s.dtype)")
|
|
print(" Sample: \(scales)")
|
|
let nanCount = scales.filter { $0.isNaN }.count
|
|
let negCount = scales.filter { $0 < 0 }.count
|
|
print(" NaN: \(nanCount), Negative: \(negCount)")
|
|
}
|
|
|
|
// 31B (sharded, MLX-vlm 0.4.3)
|
|
print("\n3. 31B:")
|
|
let model31BPath = "/Users/accusys/MarkBaseEngine/models/gemma-4-31b-it-4bit"
|
|
let model31BIndex = try SafeTensorsIndex(modelDir: model31BPath)
|
|
var model31BReaders: [String: SafeTensorsReader] = [:]
|
|
for shardFile in model31BIndex.weightMap.values {
|
|
if model31BReaders[shardFile] == nil {
|
|
model31BReaders[shardFile] = try SafeTensorsReader(path: "\(model31BPath)/\(shardFile)")
|
|
}
|
|
}
|
|
let model31BTensors = model31BReaders.values.flatMap { $0.allTensors }
|
|
|
|
let model31BScales = model31BTensors.first { $0.name.contains("embed_tokens.scales") }
|
|
if let s = model31BScales {
|
|
let reader = model31BReaders[model31BIndex.weightMap[s.name] ?? "model-00001-of-00004.safetensors"]!
|
|
let data = try reader.read(tensor: s)
|
|
let scales = data.withUnsafeBytes { ptr in
|
|
Array(ptr.assumingMemoryBound(to: Float.self).prefix(20))
|
|
}
|
|
print(" Scales shape: \(s.shape), dtype: \(s.dtype)")
|
|
print(" Sample: \(scales)")
|
|
let nanCount = scales.filter { $0.isNaN }.count
|
|
let negCount = scales.filter { $0 < 0 }.count
|
|
print(" NaN: \(nanCount), Negative: \(negCount)")
|
|
}
|
|
|
|
print("\n═══════════════════════════════════════════════════════════════════")
|
|
print(" Comparison Summary")
|
|
print("═══════════════════════════════════════════════════════════════════\n")
|
|
print("26B-Standard: custom quant, group_size=32")
|
|
print("26B-A4B: MLX-vlm 0.4.3, group_size=64, affine")
|
|
print("31B: MLX-vlm 0.4.3, group_size=64, affine")
|
|
print("\nIf 31B scales similar to 26B-A4B (±0.01), it may have same problem!")
|
|
print("═══════════════════════════════════════════════════════════════════\n")
|
|
}
|
|
} |