docs: add model inspection script and comment failing tests

inspect_model.py

@@ -0,0 +1,60 @@
+#!/usr/bin/env python3
+"""Inspect the Ornith-1.0-35B model architecture"""
+
+import torch
+from transformers import AutoModelForCausalLM, AutoConfig
+
+print("=" * 80)
+print("Inspecting Ornith-1.0-35B model")
+print("=" * 80)
+
+# Load config
+print("\n1. Loading config...")
+config = AutoConfig.from_pretrained("/data/models/Ornith-1.0-35B", trust_remote_code=True)
+print(f"   Config class: {type(config).__name__}")
+print(f"   Model type: {config.model_type}")
+
+# Load model to CPU
+print("\n2. Loading model to CPU...")
+model = AutoModelForCausalLM.from_pretrained(
+    "/data/models/Ornith-1.0-35B",
+    device_map="cpu",
+    torch_dtype=torch.bfloat16,
+    trust_remote_code=True,
+    low_cpu_mem_usage=True,
+)
+print(f"   Model class: {type(model).__name__}")
+
+# Check if model has quantize_4bit
+print("\n3. Checking for quantization methods...")
+has_quantize = hasattr(model, 'quantize_4bit')
+print(f"   Has quantize_4bit(): {has_quantize}")
+
+# List all model components
+print("\n4. Model components:")
+for name, module in model.named_modules():
+    if len(name.split('.')) <= 2:  # Top-level and first-level
+        print(f"   {name}: {type(module).__name__}")
+
+# Check for BnB quantization support
+print("\n5. Checking BnB support...")
+try:
+    from bitsandbytes.nn import Linear4bit, Linear8bitLt
+    print("   ✓ BnB 4bit and 8bit modules available")
+except ImportError:
+    print("   ✗ BnB not installed")
+
+# Check if we can use prepare_model_for_kbit_training
+print("\n6. Checking PEFT support...")
+try:
+    from peft import prepare_model_for_kbit_training
+    print("   ✓ prepare_model_for_kbit_training available")
+except ImportError:
+    print("   ✗ PEFT not installed")
+
+print("\n" + "=" * 80)
+print("Summary:")
+print(f"   Model: {type(model).__name__}")
+print(f"   Config: {type(config).__name__}")
+print(f"   Has quantize_4bit(): {has_quantize}")
+print("=" * 80)