v0.2.0 (#330)

Co-authored-by: jinz2014 <7799920+jinz2014@users.noreply.github.com>
Co-authored-by: Jin Z <5zj@cousteau.ftpn.ornl.gov>

awq/utils/module.py

 import torch.nn as nn
 
+
 def get_named_linears(module):
     return {name: m for name, m in module.named_modules() if isinstance(m, nn.Linear)}
 
+
 def get_op_by_name(module, op_name):
     # get the op by its name relative to the module
     for name, m in module.named_modules():
@@ -12,10 +14,10 @@ def get_op_by_name(module, op_name):
 
 
 def set_op_by_name(layer, name, new_module):
-    levels = name.split('.')
+    levels = name.split(".")
     if len(levels) > 1:
         mod_ = layer
-        for l_idx in range(len(levels)-1):
+        for l_idx in range(len(levels) - 1):
             if levels[l_idx].isdigit():
                 mod_ = mod_[int(levels[l_idx])]
             else:
@@ -43,6 +45,7 @@ def append_str_prefix(x, prefix):
     else:
         return x
 
+
 def exclude_layers_to_not_quantize(linear_layers, modules_to_not_convert):
     if modules_to_not_convert is None:
         return linear_layers
@@ -51,4 +54,4 @@ def exclude_layers_to_not_quantize(linear_layers, modules_to_not_convert):
     for name, linear_layer in linear_layers.items():
         if not any(key in name for key in modules_to_not_convert):
             filtered_layers[name] = linear_layer
-    return filtered_layers
+    return filtered_layers
\ No newline at end of file

awq/utils/packing_utils.py

@@ -79,6 +79,7 @@ def unpack_reorder_pack(qweight, qzeros, bits):
 
     return qweight, qzeros
 
+
 def dequantize_gemm(qweight, qzeros, scales, bits, group_size):
     # Unpack the qweight and qzeros tensors
     iweight, izeros = unpack_awq(qweight, qzeros, bits)
@@ -94,4 +95,4 @@ def dequantize_gemm(qweight, qzeros, scales, bits, group_size):
     izeros = izeros.repeat_interleave(group_size, dim=0)
     iweight = (iweight - izeros) * scales
 
-    return iweight
+    return iweight
\ No newline at end of file

awq/utils/parallel.py

@@ -23,6 +23,7 @@ def auto_parallel(args):
     else:
         cuda_visible_devices = list(range(8))
     os.environ["CUDA_VISIBLE_DEVICES"] = ",".join(
-        [str(dev) for dev in cuda_visible_devices[:n_gpu]])
+        [str(dev) for dev in cuda_visible_devices[:n_gpu]]
+    )
     logging.debug("CUDA_VISIBLE_DEVICES: ", os.environ["CUDA_VISIBLE_DEVICES"])
     return cuda_visible_devices

awq/utils/quant_utils.py

@@ -115,7 +115,7 @@ def dequantize(imatrix, scales, zeros, group_size):
     ) * scales.repeat_interleave(group_size, dim=0)
 
     fmatrix = fmatrix.to(torch.float16)
-    
+
     return fmatrix
 
 

docs/reference/index.md

+# Auto and Base model classes in AutoAWQ
+
+View the documentation of the main classes of AutoAWQ models below.
+
+::: awq.models.auto.AutoAWQForCausalLM
+::: awq.models.base.BaseAWQForCausalLM