model: adapt mllama4 to VisionAttention (#8512)

Co-authored-by: root <mickjagger19@icloud.com>

python/sglang/srt/hf_transformers_utils.py

@@ -14,7 +14,6 @@
 """Utilities for Huggingface Transformers."""
 
 import contextlib
-import logging
 import os
 import warnings
 from pathlib import Path
@@ -45,7 +44,7 @@ from sglang.srt.configs import (
 )
 from sglang.srt.configs.internvl import InternVLChatConfig
 from sglang.srt.connector import create_remote_connector
-from sglang.srt.utils import is_remote_url, lru_cache_frozenset
+from sglang.srt.utils import is_remote_url, logger, lru_cache_frozenset
 
 _CONFIG_REGISTRY: Dict[str, Type[PretrainedConfig]] = {
     ChatGLMConfig.model_type: ChatGLMConfig,
@@ -317,15 +316,31 @@ def get_processor(
 
     if config.model_type not in {"llava", "clip"}:
         kwargs["use_fast"] = use_fast
+    try:
+        processor = AutoProcessor.from_pretrained(
+            tokenizer_name,
+            *args,
+            trust_remote_code=trust_remote_code,
+            revision=revision,
+            **kwargs,
+        )
 
-    processor = AutoProcessor.from_pretrained(
-        tokenizer_name,
-        *args,
-        trust_remote_code=trust_remote_code,
-        revision=revision,
-        **kwargs,
-    )
-
+    except ValueError as e:
+        error_message = str(e)
+        if "does not have a slow version" in error_message:
+            logger.info(
+                f"Processor {tokenizer_name} does not have a slow version. Automatically use fast version"
+            )
+            kwargs["use_fast"] = True
+            processor = AutoProcessor.from_pretrained(
+                tokenizer_name,
+                *args,
+                trust_remote_code=trust_remote_code,
+                revision=revision,
+                **kwargs,
+            )
+        else:
+            raise e
     tokenizer = get_tokenizer_from_processor(processor)
 
     attach_additional_stop_token_ids(tokenizer)