Initial commit: E4B-MarkBase model integration with passing tests

- 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

compare_e4b_logits.py

@@ -0,0 +1,84 @@
+#!/usr/bin/env python3
+"""
+Compare logits from original E4B model vs our Swift implementation.
+Uses transformers library since MLX doesn't support Gemma-4 yet.
+"""
+
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import json
+import numpy as np
+
+# Load original E4B model from HuggingFace
+model_path = "google/gemma-4-4b-it"
+
+print("Loading original E4B model from HuggingFace...")
+print("This may take a moment to download...")
+
+try:
+    tokenizer = AutoTokenizer.from_pretrained(model_path)
+    model = AutoModelForCausalLM.from_pretrained(
+        model_path,
+        device_map="auto",
+        torch_dtype=torch.bfloat16
+    )
+    print("Model loaded successfully!")
+except Exception as e:
+    print(f"Error loading model: {e}")
+    print("\nFalling back to checking config from converted model...")
+    
+    # Check config from converted model
+    conv_path = "/Users/accusys/MarkBase12B/models/E4B-MarkBase/config.json"
+    with open(conv_path) as f:
+        cfg = json.load(f)
+    
+    print("\nConverted model config:")
+    for key in ["hidden_size", "num_hidden_layers", "vocab_size", 
+                "num_attention_heads", "num_key_value_heads",
+                "hidden_size_per_layer_input"]:
+        print(f"  {key}: {cfg.get('text_config', {}).get(key, cfg.get(key))}")
+    
+    exit(1)
+
+# Get BOS token
+bos_token_id = tokenizer.bos_token_id if tokenizer.bos_token_id else 2
+print(f"\nBOS token ID: {bos_token_id}")
+
+# Run forward pass for position 0
+input_ids = torch.tensor([[bos_token_id]])
+print(f"Input IDs: {input_ids}")
+
+with torch.no_grad():
+    outputs = model(input_ids)
+    logits = outputs.logits[0, 0].float().numpy()  # [vocab_size]
+
+print(f"\nLogits shape: {logits.shape}")
+print(f"Logits range: min={logits.min():.2f}, max={logits.max():.2f}")
+
+# Get top 10 tokens
+top_indices = np.argsort(logits)[-10:][::-1]
+print("\nTop 10 tokens (position 0, BOS):")
+for i, idx in enumerate(top_indices):
+    token = tokenizer.decode([idx])
+    print(f"  {i+1}. token {idx} '{token}': {logits[idx]:.2f}")
+
+# Test with actual prompt
+prompt = "<start_of_turn>user\nThe capital of France is<end_of_turn>\n<start_of_turn>model\n"
+print(f"\nTesting prompt: '{prompt[:50]}...'")
+
+# Tokenize
+inputs = tokenizer(prompt, return_tensors="pt")
+print(f"Tokens: {inputs['input_ids'][0][:20].tolist()}...")
+
+# Generate
+print("\nGenerating response...")
+with torch.no_grad():
+    outputs = model.generate(
+        **inputs,
+        max_new_tokens=30,
+        do_sample=True,
+        temperature=1.0,
+        top_k=40
+    )
+response = tokenizer.decode(outputs[0], skip_special_tokens=True)
+print(f"Response: '{response}'")