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
116 lines
6.3 KiB
Swift
116 lines
6.3 KiB
Swift
import XCTest
|
|
@testable import MarkBase
|
|
|
|
final class VisionTowerIndependentTest: XCTestCase {
|
|
|
|
func test12BVisionIndependent() throws {
|
|
print("\n═══════════════════════════════════════")
|
|
print(" 12B Vision INDEPENDENT Test (No TEXT)")
|
|
print("═══════════════════════════════════════\n")
|
|
|
|
let modelDir = "/Users/accusys/.cache/huggingface/hub/models--mlx-community--gemma-4-12B-it-4bit/snapshots/73bcf09092aa277861d5a191b989b666f7f32e8f"
|
|
|
|
print("Step 1: Create engine...")
|
|
let engine = try MarkBaseEngine(autoCompile: true)
|
|
print(" ✓ Engine created")
|
|
|
|
print("\nStep 2: Load VisionTower12B ONLY (no TEXT model)...")
|
|
let loadStart = Date()
|
|
let visionTower = try VisionTower12B.load(modelDir: modelDir, engine: engine)
|
|
let loadTime = Date().timeIntervalSince(loadStart) * 1000
|
|
print(" ✓ Vision tower loaded in \(loadTime) ms")
|
|
print(" Config: hiddenDim=\(visionTower.hiddenDim), patchDim=\(visionTower.patchDim)")
|
|
print(" Output: \(visionTower.outputDim)")
|
|
|
|
print("\nStep 3: Generate synthetic patch embeddings...")
|
|
let numPatches = 256
|
|
let patchDim = visionTower.patchDim
|
|
var patches: [Float] = []
|
|
for i in 0..<numPatches * patchDim {
|
|
patches.append(Float.random(in: -0.5...0.5))
|
|
}
|
|
print(" ✓ Patch shape: [\(numPatches), \(patchDim)]")
|
|
|
|
print("\nStep 4: Process vision...")
|
|
let patchBuffer = engine.device.makeBuffer(bytes: patches, length: patches.count * 4)!
|
|
let outputBuffer = engine.device.makeBuffer(length: numPatches * visionTower.outputDim * 4)!
|
|
|
|
let forwardStart = Date()
|
|
try visionTower.forward(patchEmbeddings: patchBuffer, numPatches: numPatches, outputBuffer: outputBuffer)
|
|
let forwardTime = Date().timeIntervalSince(forwardStart) * 1000
|
|
|
|
let ptr = outputBuffer.contents().assumingMemoryBound(to: Float.self)
|
|
let output = Array(UnsafeBufferPointer(start: ptr, count: numPatches * visionTower.outputDim))
|
|
|
|
print(" Output shape: [\(numPatches), \(visionTower.outputDim)]")
|
|
print(" Range: [\(output.min() ?? 0), \(output.max() ?? 0)]")
|
|
print(" NaN count: \(output.filter { $0.isNaN }.count)")
|
|
print(" Mean: \(output.reduce(0, +) / Float(output.count))")
|
|
print(" Time: \(forwardTime) ms")
|
|
|
|
XCTAssertFalse(output.contains { $0.isNaN }, "Vision output should not have NaN")
|
|
|
|
print("\n═══════════════════════════════════════")
|
|
print("✓ 12B Vision INDEPENDENT test passed")
|
|
print(" Load time: \(loadTime) ms (vs 3-4 min for TEXT model)")
|
|
print(" Forward time: \(forwardTime) ms")
|
|
print("═══════════════════════════════════════\n")
|
|
}
|
|
|
|
func testE4BVisionIndependent() throws {
|
|
print("\n═══════════════════════════════════════")
|
|
print(" E4B Vision INDEPENDENT Test (No TEXT)")
|
|
print("═══════════════════════════════════════\n")
|
|
|
|
let modelDir = "/Users/accusys/MarkBaseEngine/models/E4B-MarkBase"
|
|
|
|
print("Step 1: Create engine...")
|
|
let engine = try MarkBaseEngine(autoCompile: true)
|
|
print(" ✓ Engine created")
|
|
|
|
print("\nStep 2: Load VisionTower (E4B Full) ONLY (no TEXT model)...")
|
|
let loadStart = Date()
|
|
let vcfg = loadVisionConfig(modelDir: modelDir)
|
|
guard let reader = try? SafeTensorsReader(path: modelDir + "/model.safetensors") else {
|
|
throw NSError(domain: "Vision", code: -1, userInfo: [NSLocalizedDescriptionKey: "Failed to create safetensors reader"])
|
|
}
|
|
let visionTower = try loadVisionTower(reader: reader, config: vcfg, engine: engine)
|
|
let loadTime = Date().timeIntervalSince(loadStart) * 1000
|
|
print(" ✓ Vision tower loaded in \(loadTime) ms")
|
|
print(" Config: hiddenSize=\(visionTower.config.hiddenSize), layers=\(visionTower.config.numHiddenLayers)")
|
|
|
|
print("\nStep 3: Generate synthetic patch embeddings...")
|
|
let numPatches = 256
|
|
let patchDim = 768
|
|
var patches: [Float] = []
|
|
for i in 0..<numPatches * patchDim {
|
|
patches.append(Float.random(in: -0.5...0.5))
|
|
}
|
|
print(" ✓ Patch shape: [\(numPatches), \(patchDim)]")
|
|
|
|
print("\nStep 4: Process vision...")
|
|
let patchBuffer = engine.device.makeBuffer(bytes: patches, length: patches.count * 4)!
|
|
let outputBuffer = engine.device.makeBuffer(length: numPatches * visionTower.config.outputProjDims * 4)!
|
|
|
|
let forwardStart = Date()
|
|
try visionTower.forward(patchEmbeddings: patchBuffer, numPatches: numPatches, outputBuffer: outputBuffer)
|
|
let forwardTime = Date().timeIntervalSince(forwardStart) * 1000
|
|
|
|
let ptr = outputBuffer.contents().assumingMemoryBound(to: Float.self)
|
|
let output = Array(UnsafeBufferPointer(start: ptr, count: numPatches * visionTower.config.outputProjDims))
|
|
|
|
print(" Output shape: [\(numPatches), \(visionTower.config.outputProjDims)]")
|
|
print(" Range: [\(output.min() ?? 0), \(output.max() ?? 0)]")
|
|
print(" NaN count: \(output.filter { $0.isNaN }.count)")
|
|
print(" Mean: \(output.reduce(0, +) / Float(output.count))")
|
|
print(" Time: \(forwardTime) ms")
|
|
|
|
XCTAssertFalse(output.contains { $0.isNaN }, "Vision output should not have NaN")
|
|
|
|
print("\n═══════════════════════════════════════")
|
|
print("✓ E4B Vision INDEPENDENT test passed")
|
|
print(" Load time: \(loadTime) ms")
|
|
print(" Forward time: \(forwardTime) ms")
|
|
print("═══════════════════════════════════════\n")
|
|
}
|
|
} |