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
120 lines
5.6 KiB
Swift
120 lines
5.6 KiB
Swift
import XCTest
|
|
@testable import MarkBase
|
|
|
|
final class G12BPerformanceTests: XCTestCase {
|
|
|
|
func testInferenceSpeed() throws {
|
|
print("\n═══════════════════════════════════════")
|
|
print(" 12B Performance Benchmark")
|
|
print("═══════════════════════════════════════\n")
|
|
|
|
let modelDir = "/Users/accusys/.cache/huggingface/hub/models--mlx-community--gemma-4-12B-it-4bit/snapshots/73bcf09092aa277861d5a191b989b666f7f32e8f"
|
|
|
|
print("Step 1: Initialize Metal engine...")
|
|
let engine = try MarkBaseEngine(autoCompile: true)
|
|
print(" ✓ Engine ready\n")
|
|
|
|
print("Step 2: Load 12B model...")
|
|
let startLoad = Date()
|
|
let model = try E4BModel(modelDir: modelDir, engine: engine, maxContextLength: 512)
|
|
let loadTime = Date().timeIntervalSince(startLoad)
|
|
print(" ✓ Model loaded in \(loadTime) seconds\n")
|
|
|
|
print("Step 3: Warm-up inference (5 tokens)...")
|
|
for i in 0..<5 {
|
|
_ = try model.forward(tokenId: i, position: i)
|
|
}
|
|
print(" ✓ Warm-up completed\n")
|
|
|
|
print("Step 4: Benchmark inference (20 tokens)...")
|
|
let startInference = Date()
|
|
for i in 0..<20 {
|
|
let tokenStart = Date()
|
|
_ = try model.forward(tokenId: i, position: i)
|
|
let tokenTime = Date().timeIntervalSince(tokenStart)
|
|
if i >= 5 { // Skip warm-up in stats
|
|
print(" Token \(i): \(tokenTime) seconds")
|
|
}
|
|
}
|
|
let totalInferenceTime = Date().timeIntervalSince(startInference)
|
|
let avgTokenTime = totalInferenceTime / 20.0
|
|
let tokensPerSecond = 20.0 / totalInferenceTime
|
|
|
|
print("\nPerformance Summary:")
|
|
print(" Total time (20 tokens): \(totalInferenceTime) seconds")
|
|
print(" Average per token: \(avgTokenTime) seconds")
|
|
print(" Tokens per second: \(tokensPerSecond) tok/s")
|
|
print(" Comparison: E4B achieves ~0.051s/token (~19.7 tok/s)")
|
|
print(" 12B ratio: \(avgTokenTime / 0.051)x slower than E4B\n")
|
|
|
|
print("═══════════════════════════════════════")
|
|
print("✓ Performance benchmark completed")
|
|
print("═══════════════════════════════════════\n")
|
|
|
|
// Log results for tracking
|
|
print("RESULTS:")
|
|
print(" avg_token_time=\(avgTokenTime)")
|
|
print(" tokens_per_second=\(tokensPerSecond)")
|
|
print(" model_load_time=\(loadTime)")
|
|
}
|
|
|
|
func test31BPerformance() throws {
|
|
print("\n═══════════════════════════════════════")
|
|
print(" 31B Dense Performance Benchmark")
|
|
print("═══════════════════════════════════════\n")
|
|
|
|
let modelDir = "/Users/accusys/MarkBaseEngine/models/gemma-4-31b-it-4bit"
|
|
|
|
guard FileManager.default.fileExists(atPath: modelDir + "/config.json") else {
|
|
print("✗ 31B model not found")
|
|
return
|
|
}
|
|
|
|
print("Step 1: Initialize Metal engine...")
|
|
let engine = try MarkBaseEngine(autoCompile: true)
|
|
print(" ✓ Engine ready\n")
|
|
|
|
print("Step 2: Load 31B model (~18 GB)...")
|
|
let startLoad = Date()
|
|
let model = try E4BModel(modelDir: modelDir, engine: engine, maxContextLength: 128)
|
|
let loadTime = Date().timeIntervalSince(startLoad)
|
|
print(" ✓ Model loaded in \(String(format: "%.1f", loadTime))s")
|
|
print(" Layers: \(model.numHiddenLayers)")
|
|
print(" Hidden: \(model.hiddenSize)\n")
|
|
|
|
print("Step 3: Warm-up inference (3 tokens)...")
|
|
for i in 0..<3 {
|
|
_ = try model.forward(tokenId: i, position: i)
|
|
}
|
|
print(" ✓ Warm-up completed\n")
|
|
|
|
print("Step 4: Benchmark inference (10 tokens)...")
|
|
let startInference = Date()
|
|
var tokenTimes: [TimeInterval] = []
|
|
for i in 0..<10 {
|
|
let tokenStart = Date()
|
|
_ = try model.forward(tokenId: 2, position: i)
|
|
let tokenTime = Date().timeIntervalSince(tokenStart)
|
|
tokenTimes.append(tokenTime)
|
|
print(" Token \(i): \(String(format: "%.3f", tokenTime))s")
|
|
}
|
|
let totalInferenceTime = Date().timeIntervalSince(startInference)
|
|
let avgTokenTime = totalInferenceTime / 10.0
|
|
let tokensPerSecond = 10.0 / totalInferenceTime
|
|
|
|
print("\nPerformance Summary:")
|
|
print(" Model load time: \(String(format: "%.1f", loadTime))s")
|
|
print(" Total time (10 tokens): \(String(format: "%.3f", totalInferenceTime))s")
|
|
print(" Average per token: \(String(format: "%.3f", avgTokenTime))s")
|
|
print(" Tokens per second: \(String(format: "%.1f", tokensPerSecond)) tok/s")
|
|
|
|
print("\n═══════════════════════════════════════")
|
|
print("✓ 31B Performance benchmark completed")
|
|
print("═══════════════════════════════════════\n")
|
|
|
|
print("RESULTS:")
|
|
print(" avg_token_time=\(avgTokenTime)")
|
|
print(" tokens_per_second=\(tokensPerSecond)")
|
|
print(" model_load_time=\(loadTime)")
|
|
}
|
|
} |