Remove legacy dispatcher (superseded by moe_orchestrator.py)

frankenmoe_dispatcher.py

@@ -1,161 +0,0 @@
-"""
-FrankenMoE Dispatcher — Rebuild-on-Swap Architecture (v4: Handle MoE Switch Layers)
-แก้ปัญหา 'Can't convert layer of type LoRASwitchLinear to LoRA'
-โดยตรวจจับทั้ง LoRALinear (Attention) และ LoRASwitchLinear (MoE Expert MLP)
-"""
-import json
-from mlx.utils import tree_unflatten
-from mlx_lm import load, generate
-from mlx_lm.tuner.utils import linear_to_lora_layers
-
-# 🔑 Import ทั้ง 2 Class ที่ mlx_lm ใช้ห่อ LoRA
-# LoRALinear      -> ใช้กับ Attention Layer ปกติ (q,k,v,o_proj)
-# LoRASwitchLinear -> ใช้กับ MoE Switch/Expert MLP Layer โดยเฉพาะ
-LORA_WRAPPER_TYPES = []
-try:
-    from mlx_lm.tuner.lora import LoRALinear
-    LORA_WRAPPER_TYPES.append(LoRALinear)
-except ImportError:
-    pass
-try:
-    from mlx_lm.tuner.lora import LoRASwitchLinear
-    LORA_WRAPPER_TYPES.append(LoRASwitchLinear)
-except ImportError:
-    pass
-
-LORA_WRAPPER_TYPES = tuple(LORA_WRAPPER_TYPES)
-print(f"🔎 ตรวจพบ LoRA Wrapper Types ที่ต้องจัดการ: {[t.__name__ for t in LORA_WRAPPER_TYPES]}")
-
-
-class FrankenMoEDispatcher:
-    def __init__(self, base_model_path="output-moe-mlx-4bit", max_num_layers=16):
-        print(f"🔄 กำลังโหลด Base Model: {base_model_path} ...")
-        self.model, self.tokenizer = load(base_model_path)
-        self.max_num_layers = max_num_layers
-        self.current_expert = None
-
-        self.experts = {
-            "python": {
-                "config_path": "adapters/expert-1-python-v2/adapter_config.json",
-                "weights_path": "adapters/expert-1-python-v2/0000500_adapters.safetensors",
-            },
-            "typescript": {
-                "config_path": "adapters/expert-2-typescript/adapter_config.json",
-                "weights_path": "adapters/expert-2-typescript/0000100_adapters.safetensors",
-            },
-        }
-
-        # Snapshot ทุก Module ใน Layer เป้าหมาย (Attention + MoE Switch + อะไรก็ตาม)
-        self._original_layers = {}
-        total_layers = len(self.model.layers)
-        self._layer_range = range(total_layers - max_num_layers, total_layers)
-        for real_idx in self._layer_range:
-            l = self.model.layers[real_idx]
-            for k, m in l.named_modules():
-                self._original_layers[(real_idx, k)] = m
-
-        print(f"✅ Base Model พร้อมใช้งาน (Snapshot {len(self._original_layers)} modules)\n")
-
-    def _restore_original_layers(self):
-        """สแกนหา Module ที่เป็น LoRALinear หรือ LoRASwitchLinear แล้ว Restore กลับเป็นของเดิม"""
-        restored_count = 0
-        restored_types = {}
-        for real_idx in self._layer_range:
-            l = self.model.layers[real_idx]
-            restore_pairs = []
-            for k, m in l.named_modules():
-                if isinstance(m, LORA_WRAPPER_TYPES):
-                    orig = self._original_layers.get((real_idx, k))
-                    if orig is not None:
-                        restore_pairs.append((k, orig))
-                        restored_count += 1
-                        type_name = type(m).__name__
-                        restored_types[type_name] = restored_types.get(type_name, 0) + 1
-            if restore_pairs:
-                l.update_modules(tree_unflatten(restore_pairs))
-        return restored_count, restored_types
-
-    def load_expert(self, expert_name: str):
-        if expert_name == self.current_expert:
-            return
-
-        if expert_name not in self.experts:
-            raise ValueError(f"ไม่รู้จัก Expert: {expert_name}")
-
-        info = self.experts[expert_name]
-        with open(info["config_path"]) as f:
-            adapter_config = json.load(f)
-
-        num_layers = adapter_config.get("num_layers", 16)
-        lora_parameters = adapter_config.get("lora_parameters", {})
-
-        print(f"🔧 กำลังสลับไปใช้ Expert: '{expert_name}' "
-              f"(rank={lora_parameters.get('rank')}, "
-              f"scale={lora_parameters.get('scale')}, "
-              f"keys={lora_parameters.get('keys', 'default')})")
-
-        n, types_breakdown = self._restore_original_layers()
-        print(f"   ↳ Restored {n} module(s) กลับเป็นของเดิม | breakdown: {types_breakdown}")
-
-        linear_to_lora_layers(self.model, num_layers, lora_parameters, use_dora=False)
-        self.model.load_weights(info["weights_path"], strict=False)
-        self.model.eval()
-
-        self.current_expert = expert_name
-        print(f"✅ พร้อมใช้งาน Expert: '{expert_name}'\n")
-
-    def route(self, prompt: str) -> str:
-        ts_signals = [
-            "typescript", "interface", "type guard", "generic", ": string",
-            ": number", "utility type", "conditional type", "mapped type",
-        ]
-        py_signals = [
-            "python", "def ", "decorator", "list comprehension",
-            "generator", "context manager",
-        ]
-        prompt_lower = prompt.lower()
-        ts_score = sum(1 for kw in ts_signals if kw in prompt_lower)
-        py_score = sum(1 for kw in py_signals if kw in prompt_lower)
-        return "typescript" if ts_score > py_score else "python"
-
-    def generate_response(self, prompt: str, max_tokens: int = 300):
-        expert_name = self.route(prompt)
-        self.load_expert(expert_name)
-
-        messages = [{"role": "user", "content": prompt}]
-        formatted_prompt = self.tokenizer.apply_chat_template(
-            messages, add_generation_prompt=True, tokenize=False
-        )
-        response = generate(
-            self.model, self.tokenizer,
-            prompt=formatted_prompt, max_tokens=max_tokens, verbose=False
-        )
-        return expert_name, response
-
-
-if __name__ == "__main__":
-    dispatcher = FrankenMoEDispatcher()
-
-    test_cases = [
-        "Write a Python function to check if a number is prime.",
-        "Write a TypeScript type guard function called isArray.",
-        "Write a Python decorator that logs execution time.",
-        "Create a TypeScript utility type Readonly that makes properties readonly recursively.",
-        "Write a Python generator function that yields Fibonacci numbers.",
-        "Write a TypeScript function isString that acts as a type guard.",
-        "Write a Python function using list comprehension to flatten a nested list.",
-        "Write a TypeScript debounce function using closures.",
-    ]
-
-    for i, prompt in enumerate(test_cases, 1):
-        print("=" * 70)
-        print(f"โจทย์ที่ {i}: {prompt}")
-        print("=" * 70)
-        expert_used, response = dispatcher.generate_response(prompt)
-        print(f"[Router เลือก Expert: {expert_used}]")
-        print(response)
-        print()
-
-    print("=" * 70)
-    print("🎉 สลับ Expert สำเร็จทั้งหมด 8 รอบ โดยไม่มี Error")
-    print("=" * 70)