Fix LSTM int8 quantization model size issue (pytorch#23577)

Summary:
Pull Request resolved: pytorch#23577

This diff is fixing a model size issue introduced in pytorch#23291. After that PR, the model size after in8 quantization is the same as that of the original unquantized model. The reason is that we save original weight for int8 quantization even when that's not needed anymore. This diff fixes that by only saving original weight for fp16 quantization path.

Reviewed By: llyfacebook

Differential Revision: D16557619

fbshipit-source-id: f924ae8d155a0d525b86a7440b3c7147d5bead0a

torch/jit/quantized.py

@@ -290,9 +290,6 @@ def process_weights(ihhh, layer, suffix, dtype):
                     weight = getattr(other, weight_name)
                     bias = getattr(other, bias_name)
 
-                    orig_weights.append(weight_name)
-                    self.register_buffer(weight_name, weight)
-
                     if dtype == torch.int8: 
                         # for each layer, for each direction we need to quantize and pack
                         # weights and pack parameters in this order:
@@ -318,6 +315,8 @@ def process_weights(ihhh, layer, suffix, dtype):
                         packed_weight = torch.fbgemm_pack_gemm_matrix_fp16(
                             weight.clone().float())
 
+                        orig_weights.append(weight_name)
+                        self.register_buffer(weight_name, weight)
                         params = [packed_weight, bias]
                         pos_names = ['packed', 'b']
                         ret_name = ['{}_{}_l{}{}'.format(name, ihhh, layer, suffix) for name in pos_names]
@@ -336,7 +335,13 @@ def process_weights(ihhh, layer, suffix, dtype):
 
         self._packed_weights = packed_weights
         self._quantized_weights = quantized_weights
-        self._orig_weights = orig_weights
+        # For int8 quantization, _orig_weights is not needed in the quantization logic, 
+        # however there is a JIT compilation error without it. This is just used to 
+        # workaround that error. 
+        if dtype == torch.int8: 
+            self._orig_weights = self._packed_weights
+        else: 
+            self._orig_weights = orig_weights
 
     @torch.jit.script_method
     def check_input(self, input, batch_sizes):