Update modeling_internlm2.py

modeling_internlm2.py

@@ -46,8 +46,11 @@ logger = logging.get_logger(__name__)
 
 _CONFIG_FOR_DOC = 'InternLM2Config'
 
+
+# --- PATCH: Safe FlashAttention import ---
 flash_attn_func, flash_attn_varlen_func = None, None
 pad_input, index_first_axis, unpad_input = None, None, None
+
 try:
     from flash_attn import flash_attn_func as _flash_attn_func
     from flash_attn import flash_attn_varlen_func as _flash_attn_varlen_func
@@ -58,25 +61,29 @@ try:
     flash_attn_func, flash_attn_varlen_func = _flash_attn_func, _flash_attn_varlen_func
     pad_input, index_first_axis, unpad_input = _pad_input, _index_first_axis, _unpad_input
     has_flash_attn = True
-except:
+    print("[INFO] FlashAttention detected and enabled.")
+except Exception as e:
     has_flash_attn = False
-
+    print(f"[WARNING] FlashAttention not available ({e}). Using PyTorch scaled_dot_product_attention instead.")
 
 def _import_flash_attn():
+    """Safe import for FlashAttention; if not available, fallback to torch attention."""
     global flash_attn_func, flash_attn_varlen_func
     global pad_input, index_first_axis, unpad_input
     try:
         from flash_attn import flash_attn_func as _flash_attn_func
-        from flash_attn import \
-            flash_attn_varlen_func as _flash_attn_varlen_func
-        from flash_attn.bert_padding import \
-            index_first_axis as _index_first_axis
+        from flash_attn import flash_attn_varlen_func as _flash_attn_varlen_func
+        from flash_attn.bert_padding import index_first_axis as _index_first_axis
         from flash_attn.bert_padding import pad_input as _pad_input
         from flash_attn.bert_padding import unpad_input as _unpad_input
         flash_attn_func, flash_attn_varlen_func = _flash_attn_func, _flash_attn_varlen_func
         pad_input, index_first_axis, unpad_input = _pad_input, _index_first_axis, _unpad_input
+        print("[INFO] FlashAttention successfully imported.")
     except ImportError:
-        raise ImportError('flash_attn is not installed.')
+        print("[WARNING] flash_attn is not installed. Continuing with standard attention.")
+        flash_attn_func = None
+        flash_attn_varlen_func = None
+        pad_input = index_first_axis = unpad_input = None
 
 
 # Copied from transformers.models.llama.modeling_llama._get_unpad_data