refactor(models): unify generate() signatures across all LLM architectures\n\n- Unified method signature: (x, max_new_tokens, do_sample, temperature, top_k, top_p, use_cache, attention_mask, **kwargs)\n- Added del attention_mask, kwargs in every generate() for compatibility and clean API\n- Prepared for drop-in replacement and ease of future batching/serving\n\nNo changes to core model logic or sampling algorithms.

llm/src/llm/models/gemma/gemma.py

@@ -209,14 +209,17 @@ class Gemma(BaseModel):
         else:
             return (logits, None)
 
-    def generate(self,
-        x: torch.Tensor, 
-        max_new_tokens: int, 
+    def generate(
+        self,
+        x: torch.Tensor,
+        max_new_tokens: int,
         do_sample: bool,
         temperature: float = 1.0,
         top_k: int = None,
         top_p: float = None,
-        use_cache: bool = True
+        use_cache: bool = True,
+        attention_mask: torch.Tensor = None,
+        **kwargs
     ) -> torch.Tensor:
         """
         Авторегрессивная генерация токенов с использованием greedy, temperature, top-k и top-p sampling.

llm/src/llm/models/gpt/gpt.py

@@ -193,8 +193,9 @@ class GPT(BaseModel):
         temperature: float = 1.0,
         top_k: int = None,
         top_p: float = None,
-        attention_mask: torch.Tensor = None,  # Добавляем для совместимости с HF
-        **kwargs,  # Игнорируем остальные параметры
+        use_cache: bool = True,
+        attention_mask: torch.Tensor = None,
+        **kwargs
     ) -> torch.Tensor:
         """
         Авторегрессивная генерация текста с поддержкой жадного поиска (greedy), вероятностного сэмплирования с температурой,

llm/src/llm/models/gpt/gpt2.py

@@ -214,6 +214,8 @@ class GPT2(BaseModel):
         top_k: int = None,
         top_p: float = None,
         use_cache: bool = True,
+        attention_mask: torch.Tensor = None,
+        **kwargs
     ) -> torch.Tensor:
         """
         Авторегрессивная генерация токенов с поддержкой greedy, temperature, top-k, top-p sampling и KV-кэша.

llm/src/llm/models/llama/llama.py

@@ -176,6 +176,8 @@ class Llama(BaseModel):
         top_k: int = None,
         top_p: float = None,
         use_cache: bool = True,
+        attention_mask: torch.Tensor = None,
+        **kwargs
     ) -> torch.Tensor:
         """
         Авторегрессивная генерация последовательностей на основе LLaMA (greedy, temperature, top-k, top-p/nucleus, поддержка KV-кэша).

llm/src/llm/models/mistral/mistral.py

@@ -140,14 +140,17 @@ class Mistral(BaseModel):
         else:
             return (logits, None)
 
-    def generate(self,
-        x: torch.Tensor, 
-        max_new_tokens: int, 
+    def generate(
+        self,
+        x: torch.Tensor,
+        max_new_tokens: int,
         do_sample: bool,
         temperature: float = 1.0,
         top_k: int = None,
         top_p: float = None,
-        use_cache: bool = True
+        use_cache: bool = True,
+        attention_mask: torch.Tensor = None,
+        **kwargs
     ) -> torch.Tensor:
         """
         Авторегрессивная генерация токенов с поддержкой greedy, temperature, top-k/top-p sampling

llm/src/llm/models/mixtral/mixtral.py

@@ -222,14 +222,17 @@ class Mixtral(BaseModel):
         else:
             return (logits, None)
 
-    def generate(self,
-        x: torch.Tensor, 
-        max_new_tokens: int, 
+    def generate(
+        self,
+        x: torch.Tensor,
+        max_new_tokens: int,
         do_sample: bool,
         temperature: float = 1.0,
         top_k: int = None,
         top_p: float = None,
-        use_cache: bool = True
+        use_cache: bool = True,
+        attention_mask: torch.Tensor = None,
+        **kwargs
     ) -> torch.Tensor:
         """
         Авторегрессивная генерация токенов с поддержкой greedy, temperature, top-k/top-p sampling