augllm 2.3.2__tar.gz → 2.3.4__tar.gz

{augllm-2.3.2/src/augllm.egg-info → augllm-2.3.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: augllm
-Version: 2.3.2
+Version: 2.3.4
 Summary: This is LLM interface library.
 Requires-Python: >=3.13
 Description-Content-Type: text/markdown
@@ -10,7 +10,7 @@ Requires-Dist: av>=17.0.0
 Requires-Dist: huggingface-hub>=1.4.1
 Requires-Dist: librosa>=0.11.0
 Requires-Dist: llama-cpp-python[server]>=0.3.16
-Requires-Dist: mlx-lm>=0.30.7
+Requires-Dist: mlx-lm>=0.31.2
 Requires-Dist: mlx-vlm>=0.3.12
 Requires-Dist: ollama>=0.6.1
 Requires-Dist: openai>=2.24.0

{augllm-2.3.2 → augllm-2.3.4}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "augllm"
-version = "2.3.2"
+version = "2.3.4"
 description = "This is LLM interface library."
 readme = "README.md"
 requires-python = ">=3.13"
@@ -10,7 +10,7 @@ dependencies = [
     "huggingface-hub>=1.4.1",
     "librosa>=0.11.0",
     "llama-cpp-python[server]>=0.3.16",
-    "mlx-lm>=0.30.7",
+    "mlx-lm>=0.31.2",
     "mlx-vlm>=0.3.12",
     "ollama>=0.6.1",
     "openai>=2.24.0",

{augllm-2.3.2 → augllm-2.3.4}/src/augllm/mlx/mlx_interface.py

@@ -24,11 +24,13 @@ class MlxLLMInterface:
         top_p: float = 1.0, 
         min_p: float = 0.0, 
         max_tokens: int = 8192,
-        enable_thinking: bool = False
+        enable_thinking: bool = False,
+        keep_alive: bool = False  # モデルをメモリに保持するか
     ):
         self.model_path = model_path
         self.use_vision = use_vision
         self.enable_thinking = enable_thinking
+        self.keep_alive = keep_alive
         
         # サンプリングパラメータの保持
         self.temp = temp
@@ -47,6 +49,11 @@ class MlxLLMInterface:
 
     #---------------------------------------------------------
     def _load_model(self):
+
+        # モデルが既にロードされている場合はスキップ
+        if self.model is not None:
+            return
+        
         # 指定されたパラメータでサンプラーを固定
         self.sampler = make_sampler(
             temp=self.temp,
@@ -64,7 +71,7 @@ class MlxLLMInterface:
     
     #---------------------------------------------------------
     def _free_model(self):
-        # メモリの明示的な解放（M4 Maxのユニファイドメモリをクリーンアップ）
+        # メモリの明示的な解放
         self.model = None
         self.tokenizer = None
         self.processor = None
@@ -146,5 +153,6 @@ class MlxLLMInterface:
                     yield SimpleNamespace(text=full_text)
         
         finally:
-            # 2. 推論終了後に必ずモデルを解放
-            self._free_model()
+            # 2. keep_alive が False の場合のみ、推論終了後に解放
+            if not self.keep_alive:
+                self._free_model()

{augllm-2.3.2 → augllm-2.3.4}/src/augllm/mlx/mlx_llm.py

@@ -16,7 +16,8 @@ class MlxLLM:
             top_p: float = 1.0, 
             min_p: float = 0.0, 
             max_tokens: int = 8192,
-            enable_thinking: bool = False
+            enable_thinking: bool = False,
+            keep_alive: bool = False  # ← 追加：モデルをメモリに保持するか
     ):
         # LLMの初期化
         self.llm = MlxLLMInterface(
@@ -27,7 +28,8 @@ class MlxLLM:
             top_p=top_p,
             min_p=min_p,
             max_tokens=max_tokens,
-            enable_thinking=enable_thinking
+            enable_thinking=enable_thinking,
+            keep_alive=False  
         )
     
     #-----------------------------------------------------------------------------
@@ -40,10 +42,14 @@ class MlxLLM:
             images: Optional[List] = None
     ) -> Dict:
         
-        # (元の実装をそのまま維持。VLMのパッチやタグ付け処理など)
+        # 画像が含まれる場合
         if images and self.llm.use_vision:
-            pass # ※元のコードのタグ付与処理などがここに入ります
-            
+            content = [{"type": "text", "text": text}]
+            for _ in images:
+                content.append({"type": "image"})
+            return {"role": role, "content": content}
+
+        # テキストのみの場合    
         return {"role": role, "content": text}
 
     #-----------------------------------------------------------------------------
@@ -52,10 +58,9 @@ class MlxLLM:
     def respond(
         self,
         user_text: str, 
-        media_paths: Union[str, List[str]] = None, # 変更: user_images -> media_paths
+        media_paths: Union[str, List[str]] = None,
         system_prompt: str = None,
         stream: bool = False,
-        think: Optional[bool] = None,          # thinkモデル用のフラグ
         save_frames: bool = False,             # フレーム保存フラグ
         save_dir: str = "output/frames"        # 保存先フォルダ
     ):
@@ -67,7 +72,7 @@ class MlxLLM:
         if system_prompt:
             messages.append(self._build_message(role="system", text=system_prompt))
         
-        # 変更: メディア処理 (PIL Imageのリストへ変換)
+        # メディア処理 (PIL Imageのリストへ変換)
         pil_user_images = []
         if media_paths:
             pil_user_images = process_media_to_pil(

{augllm-2.3.2 → augllm-2.3.4/src/augllm.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: augllm
-Version: 2.3.2
+Version: 2.3.4
 Summary: This is LLM interface library.
 Requires-Python: >=3.13
 Description-Content-Type: text/markdown
@@ -10,7 +10,7 @@ Requires-Dist: av>=17.0.0
 Requires-Dist: huggingface-hub>=1.4.1
 Requires-Dist: librosa>=0.11.0
 Requires-Dist: llama-cpp-python[server]>=0.3.16
-Requires-Dist: mlx-lm>=0.30.7
+Requires-Dist: mlx-lm>=0.31.2
 Requires-Dist: mlx-vlm>=0.3.12
 Requires-Dist: ollama>=0.6.1
 Requires-Dist: openai>=2.24.0

{augllm-2.3.2 → augllm-2.3.4}/src/augllm.egg-info/SOURCES.txt

@@ -15,7 +15,6 @@ src/augllm/llm/llm.py
 src/augllm/mlx/media_controller_pil.py
 src/augllm/mlx/mlx_interface.py
 src/augllm/mlx/mlx_like_stream.py
-src/augllm/mlx/mlx_llm copy.py
 src/augllm/mlx/mlx_llm.py
 src/augllm/mlx/prompt_builder.py
 src/augllm/mlx/toggle_hf_progress_bar.py

{augllm-2.3.2 → augllm-2.3.4}/src/augllm.egg-info/requires.txt

@@ -3,7 +3,7 @@ av>=17.0.0
 huggingface-hub>=1.4.1
 librosa>=0.11.0
 llama-cpp-python[server]>=0.3.16
-mlx-lm>=0.30.7
+mlx-lm>=0.31.2
 mlx-vlm>=0.3.12
 ollama>=0.6.1
 openai>=2.24.0

augllm-2.3.2/src/augllm/mlx/mlx_llm copy.py

@@ -1,100 +0,0 @@
-
-from typing import List, Optional, Dict
-from mlx_vlm.utils import load_image
-from .mlx_interface import MlxLLMInterface
-#--------------------------------------------------------------------------------------------
-# LLM推論を制御するメインクラス
-#--------------------------------------------------------------------------------------------
-class MlxLLM:
-
-    def __init__(
-            self,
-            model_path: str, 
-            use_vision: bool = False, 
-            temp: float = 0.0,
-            top_k: int = -1, 
-            top_p: float = 1.0, 
-            min_p: float = 0.0, 
-            max_tokens: int = 8192,
-            enable_thinking: bool = False
-    ):
-        # LLMの初期化
-        self.llm = MlxLLMInterface(
-            model_path=model_path,
-            use_vision=use_vision,
-            temp=temp,
-            top_k=top_k,
-            top_p=top_p,
-            min_p=min_p,
-            max_tokens=max_tokens,
-            enable_thinking=enable_thinking
-        )
-    
-    #-----------------------------------------------------------------------------
-    # MLX/VLM形式のメッセージ構造を構築する共通メソッド
-    #-----------------------------------------------------------------------------
-    def _build_message(
-            self, 
-            role: str, 
-            text: str, 
-            images: Optional[List] = None
-    ) -> Dict:
-        
-        # 画像があり、かつVLMインターフェースが有効な場合
-        if images and self.llm.use_vision:
-            content = [{"type": "text", "text": text}]
-            for _ in images:
-                content.append({"type": "image"})
-            return {"role": role, "content": content}
-        
-        # 通常のテキストのみの場合
-        return {"role": role, "content": text}
-
-    #-----------------------------------------------------------------------------
-    # ユーザー入力に対してツール実行と回答生成を行うメインメソッド
-    #-----------------------------------------------------------------------------
-    def respond(
-        self,
-        user_text: str, 
-        user_images: List = None,
-        system_prompt: str = None,
-        stream: bool = False,
-        think: Optional[bool] = None,          # thinkモデル用のフラグ
-        save_frames: bool = False,             # フレーム保存フラグ
-        save_dir: str = "output/frames"        # 保存先フォルダ
-    ):
-        
-        # 初期化
-        messages = []
-        
-        # システムプロンプト
-        if system_prompt:
-            messages.append(self._build_message(role="system", text=system_prompt))
-        
-        # ユーザー画像の処理
-        user_images = user_images or []
-        pil_user_images = [load_image(str(p)) for p in user_images]
-        
-        # ユーザー入力
-        messages.append(self._build_message(role="user", text=user_text, images=pil_user_images))
-
-        # 回答生成フェーズ
-        full_reply = ""
-        generator = self.llm.generate(
-            messages=messages, 
-            images=pil_user_images if pil_user_images else None,
-            stream=stream
-        )
-
-        # ストリーミング応答
-        if stream:
-            for response in generator:
-                chunk = response.text
-                full_reply += chunk
-                yield chunk
-        
-        # 一括応答
-        else:
-            for response in generator:
-                full_reply += response.text
-            yield full_reply