Fix bugs in executorch package (#10251)

Summary:

There were bugs in `executorch`, which did not update `InputRecorder` import from `torchao` given changes in the path, which also required adding dependency to import `lm_eval`.

Differential Revision: D73166222

Author: Chia-Jung Chang

examples/models/llama/source_transformation/quantize.py

@@ -51,6 +51,7 @@ def quantize(  # noqa C901
     blocksize: int = 128,
     tokenizer_path: Optional[Path] = None,
     verbose: bool = False,
+    online_access: bool = True,
 ) -> torch.nn.Module:
     """
     Quantizes a model by converting all weights to int8.
@@ -62,6 +63,7 @@ def quantize(  # noqa C901
             Also the dtype of the rest of the non-quantized compoents of the model.
         checkpoint_dtype: The dtype of the checkpoint, this arg exists since it is more accurate to
             quantize the weight in its original dtype.
+        online_access: whether the execution environment has internet access. 
 
     Returns:
         A quantized model.
@@ -164,7 +166,7 @@ def quantize(  # noqa C901
 
         try:
             # torchao 0.3+
-            from torchao._eval import InputRecorder  # pyre-fixme[21]
+            from torchao._models._eval import InputRecorder
         except ImportError:
             from torchao.quantization.GPTQ import InputRecorder  # pyre-ignore
 
@@ -185,6 +187,7 @@ def quantize(  # noqa C901
                 None,  # input_prep_func
                 pad_calibration_inputs,
                 model.vocab_size,
+                online_access=online_access,
             )
             .record_inputs(
                 calibration_tasks,