Enable InputRecorder to run in offline mode and cuda device (#2067)

Summary:
Pull Request resolved: #2067

Enable `InputRecorder` to run in offline mode and cuda device. Specificlaly `InputRecorder` is called during Sandcastle CI tests, which do not allow us to access HuggingFace website.

Additionally, there is a bug in BUCK file for missing `lm_eval` dependency, which was called by `torchao`.

Differential Revision: D73179650

Author: Chia-Jung Chang

torchao/_models/_eval.py

@@ -183,12 +183,20 @@ def __init__(
         vocab_size=32000,
         pad_token=0,
         device="cpu",
+        online_access=True,
     ):
-        try:
-            super().__init__()
-        except TypeError:
-            # lm_eval 0.4.2 removed the default init
-            super().__init__("gpt2", device="cpu")
+        if online_access:
+            try:
+                super().__init__()
+            except TypeError:
+                # lm_eval 0.4.2 removed the default init
+                super().__init__("gpt2", device=device)
+        else:
+            # Create a minimal implementation when run offline
+            self._device = torch.device(device)
+            self._max_length = 2048
+            self._max_gen_toks = 256
+            self._batch_size = 1
 
         self.tokenizer = tokenizer
         self._device = torch.device(device)
@@ -257,10 +265,11 @@ def record_inputs(
         calibration_tasks,
         calibration_limit,
     ):
-        try:
-            lm_eval.tasks.initialize_tasks()
-        except:
-            pass
+        if self.online_access:
+            try:
+                lm_eval.tasks.initialize_tasks()
+            except:
+                pass
 
         task_dict = get_task_dict(calibration_tasks)
         print("Obtaining GPTQ calibration inputs on: ", calibration_tasks)