bugfix AWQ with Llama models and python 3.9 (#1384)

SUMMARY:
LlamaAttention.forward has an optional `attention_mask` field that has
no default (see
[here](https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py#L246)).
So `attention_mask=None` must be passed in, otherwise AWQ will error
out.

The previous check only worked for Python 3.10 and 3.11. This fixes it
to be a more general recommended solution that works with Python 3.9

```python
from transformers.models.llama.modeling_llama import LlamaAttention
import inspect
import typing

params = inspect.signature(LlamaAttention.forward).parameters

#old check
old_check = (params["attention_mask"].annotation._name == "Optional")
#new check
new_check = (params["attention_mask"].default is inspect.Parameter.empty)

print(f"OLD {old_check}, NEW {new_check}")
# Python 3.9: OLD False, NEW True
# Python 3.11: OLD True, NEW True
```


TEST PLAN:
This will resolve the failing e2e test at
https://github.com/neuralmagic/llm-compressor-testing/actions/runs/14654995202/job/41128588916#step:15:33208

---------

Signed-off-by: Brian Dellabetta <[email protected]>

Author: brian-dellabetta

src/llmcompressor/modifiers/awq/base.py

@@ -617,12 +617,12 @@ def _sanitize_kwargs(inputs_kwargs, module):
     # In case forward pass has optional dependencies that don't default to None.
     # This is the case for `LlamaAttention.forward` which has input
     #  `attention_mask: Optional[torch.Tensor],` (with no `= None` default)
-    # https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py#L269
+    # https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py#L246
     for k, v in params.items():
         if (
             k not in sanitized_kwargs
             and k != "use_cache"
-            and getattr(v.annotation, "_name", "") == "Optional"
+            and v.default is inspect.Parameter.empty
         ):
             sanitized_kwargs[k] = None