[GLM4.6v] Required changes for bumping up to transformer 5.x (#13229)

python/sglang/bench_serving.py

@@ -995,7 +995,10 @@ async def get_mooncake_request_over_time(
             # Form the full prompt from history
             try:
                 full_prompt_text = tokenizer.apply_chat_template(
-                    chat_history, tokenize=False, add_generation_prompt=True
+                    chat_history,
+                    tokenize=False,
+                    add_generation_prompt=True,
+                    return_dict=False,
                 )
             except Exception:
                 full_prompt_text = "\n".join(
@@ -1161,6 +1164,7 @@ def sample_sharegpt_requests(
                 [{"role": "user", "content": prompt}],
                 add_generation_prompt=True,
                 tokenize=False,
+                return_dict=False,
             )
             if tokenizer.bos_token:
                 prompt = prompt.replace(tokenizer.bos_token, "")

python/sglang/multimodal_gen/runtime/loader/weight_utils.py

@@ -46,7 +46,8 @@ enable_hf_transfer()
 class DisabledTqdm(tqdm):
 
     def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs, disable=True)
+        kwargs["disable"] = True
+        super().__init__(*args, **kwargs)
 
 
 def get_lock(model_name_or_path: str | Path, cache_dir: str | None = None):

python/sglang/srt/entrypoints/openai/serving_chat.py

@@ -324,6 +324,7 @@ class OpenAIServingChat(OpenAIServingBase):
                 **(
                     request.chat_template_kwargs if request.chat_template_kwargs else {}
                 ),
+                return_dict=False,
             )
         except Exception:
             # This except branch will be triggered when the chosen model
@@ -343,6 +344,7 @@ class OpenAIServingChat(OpenAIServingBase):
                 **(
                     request.chat_template_kwargs if request.chat_template_kwargs else {}
                 ),
+                return_dict=False,
             )
 
         if assistant_prefix:

python/sglang/srt/model_loader/weight_utils.py

@@ -70,7 +70,8 @@ enable_hf_transfer()
 
 class DisabledTqdm(tqdm):
     def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs, disable=True)
+        kwargs["disable"] = True
+        super().__init__(*args, **kwargs)
 
 
 def get_lock(model_name_or_path: str, cache_dir: Optional[str] = None):

python/sglang/test/runners.py

@@ -353,7 +353,7 @@ class HFRunner:
                     scores = []
                     for conv in prompts:
                         conv_formatted = self.tokenizer.apply_chat_template(
-                            conv, tokenize=False
+                            conv, tokenize=False, return_dict=False
                         )
                         conv_tokenized = self.tokenizer(
                             conv_formatted, return_tensors="pt"