Replace torch.empty with torch.zeros (pytorch#1157)

Summary:
X-link: pytorch/executorch#6478


It turns out that it is unsafe to use `torch.empty` in oss environment because `torch.empty` creates tensor with uninitialized data. That means the buffer could be initialized with random values depends on what is left on that piece of memory. This causes code to generate inconsistent behavior.

This PR replaces `torch.empty` with `torch.zeros` to make sure that they are properly initialized and avoid inconsistent behaviors.

Reviewed By: msaroufim

Differential Revision: D64875312

Author: Lunwen He

torchao/quantization/GPTQ.py

@@ -583,12 +583,12 @@ def __init__(
         assert in_features % (inner_k_tiles * 16) == 0, "require in_features % (innerKTiles * 16) == 0"
         self.register_buffer(
             "weight",
-            torch.empty((out_features // 8, in_features // (inner_k_tiles * 16), 32, inner_k_tiles // 2), dtype=torch.int32, device=device)
+            torch.zeros((out_features // 8, in_features // (inner_k_tiles * 16), 32, inner_k_tiles // 2), dtype=torch.int32, device=device)
         )
         self.dtype = dtype
         self.register_buffer(
             "scales_and_zeros",
-            torch.empty((in_features // groupsize, out_features, 2), dtype=self.scales_precision, device=device)
+            torch.zeros((in_features // groupsize, out_features, 2), dtype=self.scales_precision, device=device)
         )
 
     def forward(self, input: torch.Tensor) -> torch.Tensor:
@@ -935,18 +935,18 @@ def __init__(
         # currently storing unpacked int8 weights
         self.register_buffer(
             "weight",
-            torch.empty((out_features, in_features), dtype=torch.int8),
+            torch.zeros((out_features, in_features), dtype=torch.int8),
         )
         self.register_buffer(
             "scales",
-            torch.empty(
+            torch.zeros(
                 (out_features, in_features // groupsize),
                 dtype=scales_precision,
             ),
         )
         self.register_buffer(
             "zeros",
-            torch.empty(
+            torch.zeros(
                 (out_features, in_features // groupsize),
                 dtype=scales_precision,
             ),