rename

Signed-off-by: changwangss <[email protected]>

neural_compressor/transformers/models/__init__.py

@@ -12,4 +12,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .modeling_auto import _BaseQBitsAutoModelClass
+from .modeling_auto import _BaseINCAutoModelClass

neural_compressor/transformers/models/modeling_auto.py

@@ -76,7 +76,7 @@ def build_woq_model(model, quantization_config):
     return model
 
 
-class _BaseQBitsAutoModelClass:
+class _BaseINCAutoModelClass:
     ORIG_MODEL = None
 
     @classmethod
@@ -632,13 +632,13 @@ def load_low_bit(cls, pretrained_model_name_or_path, *model_args, **kwargs):
         return model
 
 
-class AutoModelForCausalLM(_BaseQBitsAutoModelClass):
+class AutoModelForCausalLM(_BaseINCAutoModelClass):
     ORIG_MODEL = transformers.AutoModelForCausalLM
 
 
-class AutoModel(_BaseQBitsAutoModelClass):
+class AutoModel(_BaseINCAutoModelClass):
     ORIG_MODEL = transformers.AutoModel
 
 
-class AutoModelForSeq2SeqLM(_BaseQBitsAutoModelClass):
+class AutoModelForSeq2SeqLM(_BaseINCAutoModelClass):
     ORIG_MODEL = transformers.AutoModelForSeq2SeqLM

neural_compressor/transformers/quantization/utils.py

@@ -330,6 +330,10 @@ def convert_to_quantized_model(model, config, device="cpu"):
         import intel_extension_for_pytorch
 
         assert hasattr(torch, "xpu") and torch.xpu.is_available(), "There is no xpu device in this system!"
+        os.environ["FORCE_DEVICE"] = "cpu"
+        logger.info(
+            "Set the environment variable FORCE_DEVICE='cpu' to ensure the quantization process occurs on the CPU."
+        )
 
     orig_dtype = torch.float32
     for param in model.parameters():
@@ -403,31 +407,6 @@ def convert_to_quantized_model(model, config, device="cpu"):
     return q_model.to(device)
 
 
-def pack_tensor_with_torch(raw_tensor, bits, compression_dtype=torch.int32):
-    """Pack the tensor with torch.
-
-    Args:
-        raw_tensor (tensor): raw tensor.
-
-    Returns:
-        tensor: packed tensor.
-    """
-    n_pack = 32 // bits
-    target_len = math.ceil(raw_tensor.shape[1] / n_pack)
-    packed_tensor = torch.zeros(raw_tensor.shape[0], target_len, dtype=compression_dtype).to(raw_tensor.device)
-    mask = torch.tensor(2**bits - 1, dtype=compression_dtype).to(raw_tensor.device)
-    for j in range(packed_tensor.shape[1]):
-        start = n_pack * j
-        end = n_pack * (j + 1)
-        tmp = raw_tensor[:, start:end].type(compression_dtype)
-        tmp &= mask
-        for e in range(tmp.shape[1]):
-            tmp[:, e] = tmp[:, e] << (bits * e)
-            packed_tensor[:, j] |= tmp[:, e]
-
-    return packed_tensor
-
-
 def convert_to_GPTQ_checkpoints(model, quantization_config):
     from intel_extension_for_pytorch.nn.modules import WeightOnlyQuantizedLinear as ipex_cpu_linear