From c31eed855116605d3bbcaff76cec33b5bf41d5d9 Mon Sep 17 00:00:00 2001
From: Sergey Penkovsky <sergey.penkovsky@gmail.com>
Date: Sun, 5 Oct 2025 20:47:36 +0300
Subject: [PATCH] fix(hf-integration): handle logits as tuple in hf_adapter,
 convert torch.Tensor to list in hf_tokenizer.decode for decoding
 compatibility

---
 hf-proxy/src/hf_proxy/hf_adapter.py   |  6 +++++-
 hf-proxy/src/hf_proxy/hf_tokenizer.py | 22 ++++++++++++++++++----
 2 files changed, 23 insertions(+), 5 deletions(-)

diff --git a/hf-proxy/src/hf_proxy/hf_adapter.py b/hf-proxy/src/hf_proxy/hf_adapter.py
index fa89ac4..3b12124 100644
--- a/hf-proxy/src/hf_proxy/hf_adapter.py
+++ b/hf-proxy/src/hf_proxy/hf_adapter.py
@@ -99,7 +99,11 @@ class HFGPTAdapter(PreTrainedModel):
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
         
         # Основной forward pass
-        logits = self.llm_model(input_ids)
+        outputs = self.llm_model(input_ids)
+        if isinstance(outputs, tuple):
+            logits = outputs[0]
+        else:
+            logits = outputs
         
         loss = None
         if labels is not None:
diff --git a/hf-proxy/src/hf_proxy/hf_tokenizer.py b/hf-proxy/src/hf_proxy/hf_tokenizer.py
index 3a1df1a..34d66aa 100644
--- a/hf-proxy/src/hf_proxy/hf_tokenizer.py
+++ b/hf-proxy/src/hf_proxy/hf_tokenizer.py
@@ -56,10 +56,24 @@ class HFTokenizerAdapter:
         add_special_tokens = kwargs.get('add_special_tokens', True)
         
         # Кодируем текст
-        input_ids = self.llm_tokenizer.encode(
-            text, 
-            add_special_tokens=add_special_tokens
-        )
+        #input_ids = self.llm_tokenizer.encode(
+        #    text, 
+        #    add_special_tokens=add_special_tokens
+        #)
+        if isinstance(text, str):
+            input_ids = self.llm_tokenizer.encode(
+                text, 
+                add_special_tokens=add_special_tokens
+            )
+            input_ids = [input_ids]  # <-- оборачиваем в batch
+        else:
+            # Список строк, батч-режим!
+            input_ids = [
+                self.llm_tokenizer.encode(
+                    t,
+                    add_special_tokens=add_special_tokens
+                ) for t in text
+            ]
         
         # Применяем truncation
         if truncation and max_length is not None and len(input_ids) > max_length: