Fix woq autoround last layer quant issue (#1419)

Signed-off-by: changwangss <chang1.wang@intel.com>

Author: changwangss

intel_extension_for_transformers/transformers/llm/quantization/utils.py

@@ -108,7 +108,9 @@ def replace_linear(
     empty_weights=False,
 ):
     if modules_to_not_convert is None:
-        modules_to_not_convert = ["lm_head"]
+        # output_layer is chatglm last layer name
+        # embed_out is dolly_v2 last layer name
+        modules_to_not_convert = ["lm_head", "output_layer", "embed_out"]
     if quantization_config.llm_int8_skip_modules:
         modules_to_not_convert = modules_to_not_convert.extend(
             quantization_config.llm_int8_skip_modules
@@ -518,6 +520,12 @@ def default_calib_func(model):
                 ".*lm_head": {  # re.match
                     "weight": {"dtype": "fp32"},
                 },
+                ".*output_layer": {  # re.match
+                    "weight": {"dtype": "fp32"},
+                },
+                ".*embed_out": {  # re.match
+                    "weight": {"dtype": "fp32"},
+                },
             },
             recipes=recipes,
         )
@@ -532,7 +540,6 @@ def default_calib_func(model):
             if orig_dtype != torch.float32:
                 model.to(dtype=torch.float32)
             break
-
         inc_model = quantization.fit(
             model, conf, calib_func=calib_func, calib_dataloader=calib_dataloader
         )

intel_extension_for_transformers/transformers/modeling/modeling_auto.py

@@ -144,7 +144,7 @@ def build_woq_model(model, quantization_config):
     from neural_compressor.adaptor.torch_utils.util import set_module
 
     for n, m in model.named_modules():
-        if "lm_head" in n:
+        if "lm_head" in n or "output_layer" in n or "embed_out" in n:
             continue
         if isinstance(m, torch.nn.Linear):
             zp = (

intel_extension_for_transformers/transformers/utils/config.py

@@ -940,7 +940,6 @@ class AutoRoundConfig(ITREXQuantizationConfigMixin):
     def __init__(
         self,
         bits: int = 8,
-        dtype: str = "int",
         tokenizer: Any = None,
         dataset: str = "NeelNanda/pile-10k",
         group_size: int = 32,
@@ -955,7 +954,6 @@ def __init__(
         use_quant_input: bool = True,
         nsamples: int = 128,
         iters: int = 200,
-        static_groups: bool = False,
         use_ggml: bool = False,
         use_neural_speed: bool = False,
         llm_int8_skip_modules=None,