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
96 lines
3.7 KiB
Swift
96 lines
3.7 KiB
Swift
import XCTest
|
|
@testable import MarkBase
|
|
|
|
final class MoEMinimalForwardTest: XCTestCase {
|
|
|
|
func testMinimalMoEForwardPass() throws {
|
|
print("\n═══════════════════════════════════════")
|
|
print(" Minimal MoE Forward Pass Test")
|
|
print("═══════════════════════════════════════\n")
|
|
|
|
let modelDir = "/Users/accusys/MarkBaseEngine/models/gemma-4-26b-a4b-it-4bit"
|
|
|
|
print("Step 1: Load model...")
|
|
let engine = try MarkBaseEngine(autoCompile: true)
|
|
let model = try E4BModel(modelDir: modelDir, engine: engine, maxContextLength: 32)
|
|
print(" ✓ Model loaded (30 layers, MoE enabled)")
|
|
|
|
print("\nStep 2: Get layer 0...")
|
|
let layer0 = model.layers[0]
|
|
print(" ✓ Layer 0 retrieved")
|
|
print(" useMoE: \(layer0.useMoE)")
|
|
print(" routerScale: \(layer0.routerScale)")
|
|
|
|
if !layer0.useMoE {
|
|
XCTFail("Layer 0 should use MoE")
|
|
return
|
|
}
|
|
|
|
print("\nStep 3: Create minimal buffers and KV cache...")
|
|
let hs = model.hiddenSize
|
|
|
|
// Create input buffer (single token embedding)
|
|
let inputBuffer = engine.device.makeBuffer(length: hs * MemoryLayout<Float>.size)!
|
|
var inputData = [Float](repeating: 0.5, count: hs)
|
|
memcpy(inputBuffer.contents(), &inputData, inputData.count * MemoryLayout<Float>.size)
|
|
print(" ✓ Input buffer created (size: \(hs))")
|
|
|
|
// Create minimal KV cache for testing
|
|
let kvCache = try KVCache(
|
|
device: engine.device,
|
|
isSliding: false,
|
|
maxContextLength: 32,
|
|
nKvHeads: 8,
|
|
headDim: 256
|
|
)
|
|
print(" ✓ KV cache created")
|
|
|
|
// Create temps for forward pass
|
|
let temps = model.temps
|
|
print(" ✓ Forward temps available")
|
|
|
|
print("\nStep 4: Test layer 0 forward pass (minimal)...")
|
|
let cmdBuf = engine.commandQueue.makeCommandBuffer()!
|
|
|
|
let start = Date()
|
|
print(" Calling layer0.forward...")
|
|
fflush(stdout)
|
|
|
|
do {
|
|
// Try forward pass for position 0
|
|
try layer0.forward(
|
|
input: inputBuffer,
|
|
position: 0,
|
|
kvCache: kvCache,
|
|
shouldStoreKV: false,
|
|
temps: temps,
|
|
engine: engine
|
|
)
|
|
|
|
cmdBuf.commit()
|
|
cmdBuf.waitUntilCompleted()
|
|
|
|
let elapsed = Date().timeIntervalSince(start)
|
|
print(" ✓ Forward pass completed in \(String(format: "%.3f", elapsed))s")
|
|
print(" Command buffer status: \(cmdBuf.status.rawValue)")
|
|
|
|
if elapsed > 5.0 {
|
|
print(" ⚠️ Forward pass took >5s - might indicate issue")
|
|
}
|
|
|
|
if cmdBuf.status != .completed {
|
|
print(" ❌ Command buffer did not complete (status: \(cmdBuf.status.rawValue))")
|
|
XCTFail("Forward pass command buffer failed")
|
|
}
|
|
|
|
} catch {
|
|
print(" ❌ Forward pass failed: \(error)")
|
|
throw error
|
|
}
|
|
|
|
print("\n═══════════════════════════════════════")
|
|
print("✓ Minimal MoE forward pass works!")
|
|
print("═══════════════════════════════════════\n")
|
|
}
|
|
}
|