Support fp16 model to weight-only quantization for PyTorch framework (#1387)

Signed-off-by: Cheng, Penghui <[email protected]>

Author: PenghuiCheng

neural_compressor/adaptor/torch_utils/weight_only.py

@@ -399,6 +399,9 @@ def rtn_quantize(
     for name, m in model.named_modules():
         if m.__class__.__name__ not in supported_layers:
             continue
+        orig_dtype = next(m.parameters()).dtype
+        if orig_dtype != torch.float:
+            m = m.float()
         if name in weight_config:  # pragma: no cover
             num_bits = weight_config[name]["bits"]
             group_size = weight_config[name]["group_size"]
@@ -466,6 +469,8 @@ def rtn_quantize(
             )
             q_weight = q_weight.T if group_dim == 0 else q_weight
             m.weight.data.copy_(q_weight)
+        if orig_dtype != torch.float:
+            m = m.to(orig_dtype)
     return model
 
 

test/quantization/test_weight_only_quantization.py

@@ -54,6 +54,7 @@ def test_trace(self):
 
     def test_rtn(self):
         fp32_model = copy.deepcopy(self.model)
+        fp16_model = copy.deepcopy(self.model).to(torch.float16)
         model1 = rtn_quantize(fp32_model, num_bits=3, group_size=-1)
         self.assertTrue(isinstance(model1.fc1, torch.nn.Linear))
         weight_config = {
@@ -67,7 +68,7 @@ def test_rtn(self):
             },
         }
         model2 = rtn_quantize(fp32_model, weight_config=weight_config)
-        model2 = rtn_quantize(fp32_model, weight_config=weight_config, return_int=True)
+        model2 = rtn_quantize(fp16_model, weight_config=weight_config, return_int=True)
         self.assertTrue(isinstance(model2.fc1, WeightOnlyLinear))
 
     def test_awq(self):