enable torchao quantization on CPU

docs/source/en/quantization/overview.md

@@ -59,7 +59,7 @@ Use the table below to help you decide which quantization method to use.
 | [HQQ](./hqq.md)                               | 🟢                   | 🟢              | 🟢        | 🔴        | 🔴                                 | 🔴              | 🟢              | 1/8         | 🟢               | 🔴                          | 🟢                      | https://github.com/mobiusml/hqq/            |
 | [optimum-quanto](./quanto.md)                 | 🟢                   | 🟢              | 🟢        | 🔴        | 🟢                                 | 🔴              | 🟢              | 2/4/8     | 🔴               | 🔴                          | 🟢                      | https://github.com/huggingface/optimum-quanto       |
 | [FBGEMM_FP8](./fbgemm_fp8.md)                 | 🟢                   | 🔴              | 🟢        | 🔴        | 🔴                                 | 🔴              | 🔴              | 8             | 🔴               | 🟢                          | 🟢                      | https://github.com/pytorch/FBGEMM       |
-| [torchao](./torchao.md)                       | 🟢                   |                 | 🟢        | 🔴        | 🟡 <sub>5</sub> | 🔴              |                 | 4/8         |                  | 🟢🔴                        | 🟢                      | https://github.com/pytorch/ao       |
+| [torchao](./torchao.md)                       | 🟢                   | 🟢               | 🟢        | 🔴        | 🟡 <sub>5</sub> | 🔴              |                 | 4/8         |                  | 🟢🔴                        | 🟢                      | https://github.com/pytorch/ao       |
 | [VPTQ](./vptq.md)                             | 🔴                   | 🔴              |     🟢     | 🟡        | 🔴                                 | 🔴              | 🟢              | 1/8         | 🔴               | 🟢                          | 🟢                      | https://github.com/microsoft/VPTQ            |
 | [SpQR](./spqr.md)                          | 🔴                       |  🔴   | 🟢        | 🔴              |    🔴    | 🔴         |         🟢              | 3              |              🔴                     | 🟢           | 🟢                      | https://github.com/Vahe1994/SpQR/       |
 | [FINEGRAINED_FP8](./finegrained_fp8.md)                 | 🟢                   | 🔴              | 🟢        | 🔴        | 🔴                                 | 🔴              | 🔴              | 8             | 🔴               | 🟢                          | 🟢                      |        |

docs/source/en/quantization/torchao.md

@@ -22,9 +22,11 @@ pip install --upgrade torch torchao transformers
 
 By default, the weights are loaded in full precision (torch.float32) regardless of the actual data type the weights are stored in such as torch.float16. Set `torch_dtype="auto"` to load the weights in the data type defined in a model's `config.json` file to automatically load the most memory-optimal data type.
 
+
 ## Manually Choose Quantization Types and Settings
 
 `torchao` Provides many commonly used types of quantization, including different dtypes like int4, float8 and different flavors like weight only, dynamic quantization etc., only `int4_weight_only`, `int8_weight_only` and `int8_dynamic_activation_int8_weight` are integrated into hugigngface transformers currently, but we can add more when needed.
+If you want to run the following codes on CPU even with GPU available, just change `device_map="cpu"` and `quantization_config = TorchAoConfig("int4_weight_only", group_size=128, layout=Int4CPULayout())` where `layout` comes from `from torchao.dtypes import Int4CPULayout` which is only available from torchao 0.8.0 and higher.
 
 Users can manually specify the quantization types and settings they want to use:
 
@@ -40,7 +42,7 @@ quantized_model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="
 
 tokenizer = AutoTokenizer.from_pretrained(model_name)
 input_text = "What are we having for dinner?"
-input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+input_ids = tokenizer(input_text, return_tensors="pt").to(quantized_model.device)
 
 # auto-compile the quantized model with `cache_implementation="static"` to get speedup
 output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
@@ -59,7 +61,7 @@ def benchmark_fn(func: Callable, *args, **kwargs) -> float:
 MAX_NEW_TOKENS = 1000
 print("int4wo-128 model:", benchmark_fn(quantized_model.generate, **input_ids, max_new_tokens=MAX_NEW_TOKENS, cache_implementation="static"))
 
-bf16_model = AutoModelForCausalLM.from_pretrained(model_name, device_map="cuda", torch_dtype=torch.bfloat16)
+bf16_model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", torch_dtype=torch.bfloat16)
 output = bf16_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static") # auto-compile
 print("bf16 model:", benchmark_fn(bf16_model.generate, **input_ids, max_new_tokens=MAX_NEW_TOKENS, cache_implementation="static"))
 
@@ -122,7 +124,7 @@ quantized_model.save_pretrained(output_dir, safe_serialization=False)
 
 # load quantized model
 ckpt_id = "llama3-8b-int4wo-128"  # or huggingface hub model id
-loaded_quantized_model = AutoModelForCausalLM.from_pretrained(ckpt_id, device_map="cuda")
+loaded_quantized_model = AutoModelForCausalLM.from_pretrained(ckpt_id, device_map="auto")
 
 
 # confirm the speedup

src/transformers/testing_utils.py

@@ -45,6 +45,7 @@
 import huggingface_hub.utils
 import urllib3
 from huggingface_hub import delete_repo
+from packaging import version
 
 from transformers import logging as transformers_logging
 
@@ -963,6 +964,18 @@ def require_torchao(test_case):
     return unittest.skipUnless(is_torchao_available(), "test requires torchao")(test_case)
 
 
+def require_torchao_version_greater_or_equal(torchao_version):
+    def decorator(test_case):
+        correct_torchao_version = is_torchao_available() and version.parse(
+            version.parse(importlib.metadata.version("torchao")).base_version
+        ) >= version.parse(torchao_version)
+        return unittest.skipUnless(
+            correct_torchao_version, f"Test requires torchao with the version greater than {torchao_version}."
+        )(test_case)
+
+    return decorator
+
+
 def require_torch_tensorrt_fx(test_case):
     """Decorator marking a test that requires Torch-TensorRT FX"""
     return unittest.skipUnless(is_torch_tensorrt_fx_available(), "test requires Torch-TensorRT FX")(test_case)

src/transformers/utils/quantization_config.py

@@ -1558,7 +1558,17 @@ def _get_torchao_quant_type_to_method(self):
 
     def get_apply_tensor_subclass(self):
         _STR_TO_METHOD = self._get_torchao_quant_type_to_method()
-        return _STR_TO_METHOD[self.quant_type](**self.quant_type_kwargs)
+        quant_type_kwargs = self.quant_type_kwargs.copy()
+        if (
+            not torch.cuda.is_available()
+            and is_torchao_available()
+            and self.quant_type == "int4_weight_only"
+            and version.parse(importlib.metadata.version("torchao")) >= version.parse("0.8.0")
+        ):
+            from torchao.dtypes import Int4CPULayout
+
+            quant_type_kwargs["layout"] = Int4CPULayout()
+        return _STR_TO_METHOD[self.quant_type](**quant_type_kwargs)
 
     def __repr__(self):
         config_dict = self.to_dict()