Added cache_block_outputs option to enable GPTQ for non-regular models (#27032)

* Added cache_block_outputs option to enable GPTQ for non-regular models

* Update src/transformers/utils/quantization_config.py

Co-authored-by: Marc Sun <[email protected]>

* Update src/transformers/utils/quantization_config.py

Co-authored-by: Marc Sun <[email protected]>

* Fixed style

* Update src/transformers/utils/quantization_config.py

Co-authored-by: amyeroberts <[email protected]>

---------

Co-authored-by: Marc Sun <[email protected]>
Co-authored-by: amyeroberts <[email protected]>

Author: 3 people

src/transformers/utils/quantization_config.py

@@ -360,6 +360,8 @@ class GPTQConfig(QuantizationConfigMixin):
         max_input_length (`int`, *optional*):
             The maximum input length. This is needed to initialize a buffer that depends on the maximum expected input
             length. It is specific to the exllama backend with act-order.
+        cache_block_outputs (`bool`, *optional*, defaults to `True`):
+                Whether to cache block outputs to reuse as inputs for the succeeding block.
     """
 
     def __init__(
@@ -380,6 +382,7 @@ def __init__(
         pad_token_id: Optional[int] = None,
         disable_exllama: bool = False,
         max_input_length: Optional[int] = None,
+        cache_block_outputs: bool = True,
         **kwargs,
     ):
         self.quant_method = QuantizationMethod.GPTQ
@@ -399,6 +402,7 @@ def __init__(
         self.pad_token_id = pad_token_id
         self.disable_exllama = disable_exllama
         self.max_input_length = max_input_length
+        self.cache_block_outputs = cache_block_outputs
         self.post_init()
 
     def get_loading_attributes(self):