Update brevitas requirement (#1677)

also clean up useless args

Co-authored-by: powderluv <powderluv@users.noreply.github.com>

apps/language_models/scripts/vicuna.py

@@ -706,20 +706,13 @@ class ShardedVicuna(SharkLLMBase):
             quantize_model(
                 get_model_impl(vicuna_model).layers,
                 dtype=torch.float32,
-                weight_quant_type="asym",
                 weight_bit_width=weight_bit_width,
                 weight_param_method="stats",
                 weight_scale_precision="float",
+                weight_quant_type="asym",
                 weight_quant_granularity="per_group",
                 weight_group_size=self.weight_group_size,
                 quantize_weight_zero_point=False,
-                input_bit_width=None,
-                input_scale_type="float",
-                input_param_method="stats",
-                input_quant_type="asym",
-                input_quant_granularity="per_tensor",
-                quantize_input_zero_point=False,
-                seqlen=2048,
             )
             print("Weight quantization applied.")
 

apps/language_models/src/model_wrappers/vicuna_model.py

@@ -26,20 +26,13 @@ class FirstVicuna(torch.nn.Module):
             quantize_model(
                 get_model_impl(self.model).layers,
                 dtype=torch.float32,
-                weight_quant_type="asym",
                 weight_bit_width=weight_bit_width,
                 weight_param_method="stats",
                 weight_scale_precision="float",
+                weight_quant_type="asym",
                 weight_quant_granularity="per_group",
                 weight_group_size=weight_group_size,
                 quantize_weight_zero_point=False,
-                input_bit_width=None,
-                input_scale_type="float",
-                input_param_method="stats",
-                input_quant_type="asym",
-                input_quant_granularity="per_tensor",
-                quantize_input_zero_point=False,
-                seqlen=2048,
             )
             print("Weight quantization applied.")
 
@@ -75,20 +68,13 @@ class SecondVicuna(torch.nn.Module):
             quantize_model(
                 get_model_impl(self.model).layers,
                 dtype=torch.float32,
-                weight_quant_type="asym",
                 weight_bit_width=weight_bit_width,
                 weight_param_method="stats",
                 weight_scale_precision="float",
+                weight_quant_type="asym",
                 weight_quant_granularity="per_group",
                 weight_group_size=weight_group_size,
                 quantize_weight_zero_point=False,
-                input_bit_width=None,
-                input_scale_type="float",
-                input_param_method="stats",
-                input_quant_type="asym",
-                input_quant_granularity="per_tensor",
-                quantize_input_zero_point=False,
-                seqlen=2048,
             )
             print("Weight quantization applied.")
 

requirements.txt

@@ -39,5 +39,5 @@ joblib # for langchain
 pefile
 pyinstaller
 
-# low precision vicuna
-brevitas @ git+https://github.com/Xilinx/brevitas.git@llm
+# vicuna quantization
+brevitas @ git+https://github.com/Xilinx/brevitas.git@dev

setup_venv.sh

@@ -159,5 +159,3 @@ if [[ -z "${CONDA_PREFIX}" && "$SKIP_VENV" != "1" ]]; then
   echo "${Green}Before running examples activate venv with:"
   echo "  ${Green}source $VENV_DIR/bin/activate"
 fi
-
-$PYTHON -m pip install git+https://github.com/Xilinx/brevitas.git@llm