[torchao safetensors] integrate torchao safetensors support with transformers (#40735)

* enable torchao safetensors

* enable torchao safetensors support

* add more version checking

Author: liangel-02

src/transformers/modeling_utils.py

@@ -727,11 +727,12 @@ def _load_state_dict_into_meta_model(
         device_map_regex = "|".join([re.escape(k) for k in sorted(device_map.keys(), reverse=True)])
 
     is_quantized = hf_quantizer is not None
-    is_hqq_or_bnb = is_quantized and hf_quantizer.quantization_config.quant_method in {
+    is_hqq_or_bnb_or_ao = is_quantized and hf_quantizer.quantization_config.quant_method in {
         QuantizationMethod.HQQ,
         QuantizationMethod.BITS_AND_BYTES,
+        QuantizationMethod.TORCHAO,
     }
-    is_meta_state_dict = shard_file.endswith(".safetensors") and not is_hqq_or_bnb
+    is_meta_state_dict = shard_file.endswith(".safetensors") and not is_hqq_or_bnb_or_ao
     file_pointer = None
     if is_meta_state_dict:
         file_pointer = safe_open(shard_file, framework="pt", device=tensor_device)
@@ -873,7 +874,7 @@ def load_shard_file(args):
         shard_file,
         state_dict,
         disk_only_shard_files,
-        is_hqq_or_bnb,
+        is_hqq_or_bnb_or_ao,
         is_quantized,
         device_map,
         hf_quantizer,
@@ -899,7 +900,7 @@ def load_shard_file(args):
     map_location = "cpu"
     if (
         shard_file.endswith(".safetensors")
-        and not is_hqq_or_bnb
+        and not is_hqq_or_bnb_or_ao
         and not (is_deepspeed_zero3_enabled() and not is_quantized)
     ):
         map_location = "meta"
@@ -922,6 +923,13 @@ def load_shard_file(args):
 
     # Fix the key names
     state_dict = {key_renaming_mapping[k]: v for k, v in state_dict.items() if k in key_renaming_mapping}
+    metadata = None
+    if shard_file.endswith(".safetensors") and is_safetensors_available():
+        with safe_open(shard_file, framework="pt") as f:
+            metadata = f.metadata()
+
+    if hf_quantizer:
+        state_dict = hf_quantizer.update_state_dict_with_metadata(state_dict, metadata)
 
     error_msgs = []
 
@@ -5277,9 +5285,10 @@ def _load_pretrained_model(
             QuantizationMethod.HQQ,
             QuantizationMethod.QUARK,
         }
-        is_hqq_or_bnb = is_quantized and hf_quantizer.quantization_config.quant_method in {
+        is_hqq_or_bnb_or_ao = is_quantized and hf_quantizer.quantization_config.quant_method in {
             QuantizationMethod.HQQ,
             QuantizationMethod.BITS_AND_BYTES,
+            QuantizationMethod.TORCHAO,
         }
 
         # Get all the keys of the state dicts that we have to initialize the model
@@ -5451,7 +5460,7 @@ def _load_pretrained_model(
                 shard_file,
                 state_dict,
                 disk_only_shard_files,
-                is_hqq_or_bnb,
+                is_hqq_or_bnb_or_ao,
                 is_quantized,
                 device_map,
                 hf_quantizer,

src/transformers/quantizers/base.py

@@ -342,6 +342,10 @@ def get_state_dict_and_metadata(self, model, safe_serialization=False):
         """Get state dict and metadata. Useful when we need to modify a bit the state dict due to quantization"""
         return None, {}
 
+    def update_state_dict_with_metadata(self, state_dict, metadata):
+        """Update state dict with metadata. Default behaviour returns state_dict"""
+        return state_dict
+
     @abstractmethod
     def _process_model_before_weight_loading(self, model, **kwargs): ...
 

src/transformers/quantizers/quantizer_torchao.py

@@ -35,6 +35,17 @@
     import torch
     import torch.nn as nn
 
+if is_torchao_available():
+    import torchao
+
+    if version.parse(importlib.metadata.version("torchao")) >= version.parse("0.14.0"):
+        from torchao.prototype.safetensors.safetensors_support import (
+            flatten_tensor_state_dict,
+            unflatten_tensor_state_dict,
+        )
+        from torchao.prototype.safetensors.safetensors_utils import is_metadata_torchao
+
+
 logger = logging.get_logger(__name__)
 
 
@@ -81,6 +92,15 @@ def _linear_extra_repr(self):
         return f"in_features={self.weight.shape[1]}, out_features={self.weight.shape[0]}, weight={weight}"
 
 
+if is_torchao_available():
+    SUPPORTED_SAFE_SERIALIZATION_CONFIGS = [
+        torchao.quantization.Float8WeightOnlyConfig,
+        torchao.quantization.Float8DynamicActivationFloat8WeightConfig,
+    ]
+
+    TORCHAO_VERSION = version.parse(importlib.metadata.version("torchao"))
+
+
 class TorchAoHfQuantizer(HfQuantizer):
     """
     Quantizer for torchao: https://github.com/pytorch/ao/
@@ -137,6 +157,21 @@ def update_dtype(self, dtype):
                 dtype = torch.float32
         return dtype
 
+    def get_state_dict_and_metadata(self, model, safe_serialization: Optional[bool] = False):
+        """
+        If the model is safe serializable, we flatten the state dict of tensor subclasses so that it is compatible with
+        the safetensors format.
+        """
+        if type(self.quantization_config.quant_type) in SUPPORTED_SAFE_SERIALIZATION_CONFIGS and safe_serialization:
+            if TORCHAO_VERSION >= version.parse("0.14.0"):
+                return flatten_tensor_state_dict(model.state_dict())
+            else:
+                raise RuntimeError(
+                    f"In order to use safetensors with torchao, please use torchao version >= 0.14.0. Current version: {TORCHAO_VERSION}"
+                )
+        else:
+            return super().get_state_dict_and_metadata(model)
+
     def adjust_target_dtype(self, dtype: "torch.dtype") -> "torch.dtype":
         if version.parse(importlib.metadata.version("accelerate")) > version.parse("0.19.0"):
             from accelerate.utils import CustomDtype
@@ -279,6 +314,16 @@ def create_quantized_param(
 
             quantize_(module, self.quantization_config.get_apply_tensor_subclass())
 
+    def update_state_dict_with_metadata(self, state_dict, metadata):
+        """
+        If the metadata contains torchao tensor subclass information, we reconstruct the tensor subclass state dict
+        from the provided state_dict and metadata.
+        """
+        if TORCHAO_VERSION >= version.parse("0.14.0") and is_metadata_torchao(metadata):
+            return unflatten_tensor_state_dict(state_dict, metadata)
+        else:
+            return super().update_state_dict_with_metadata(state_dict, metadata)
+
     def _process_model_after_weight_loading(self, model, **kwargs):
         """No process required for torchao quantized model"""
         if self.quantization_config.quant_type == "autoquant":
@@ -297,10 +342,17 @@ def _process_model_after_weight_loading(self, model, **kwargs):
 
     def is_serializable(self, safe_serialization=None) -> bool:
         if safe_serialization:
-            logger.warning(
-                "torchao quantized model does not support safe serialization, please set `safe_serialization` to False"
-            )
-            return False
+            _is_torchao_serializable = type(
+                self.quantization_config.quant_type
+            ) in SUPPORTED_SAFE_SERIALIZATION_CONFIGS and TORCHAO_VERSION >= version.parse("0.14.0")
+            if not _is_torchao_serializable:
+                logger.warning(
+                    f"torchao quantized model only supports safe serialization for {SUPPORTED_SAFE_SERIALIZATION_CONFIGS}, \
+                    and torchao version >= 0.14.0, please set `safe_serialization` to False for \
+                    {type(self.quantization_config.quant_type)} and {TORCHAO_VERSION}."
+                )
+            return _is_torchao_serializable
+
         _is_torchao_serializable = version.parse(importlib.metadata.version("huggingface_hub")) >= version.parse(
             "0.25.0"
         )

tests/quantization/torchao_integration/test_torchao.py

@@ -18,6 +18,7 @@
 import unittest
 
 from packaging import version
+from parameterized import parameterized
 
 from transformers import AutoModelForCausalLM, AutoTokenizer, TorchAoConfig
 from transformers.testing_utils import (
@@ -37,6 +38,8 @@
     import torch
 
 if is_torchao_available():
+    import torchao
+
     # renamed in torchao 0.7.0, please install the latest torchao
     from torchao.dtypes import (
         AffineQuantizedTensor,
@@ -135,7 +138,7 @@ class TorchAoTest(unittest.TestCase):
     model_name = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
     device = "cpu"
     quant_scheme_kwargs = (
-        {"group_size": 32, "layout": Int4CPULayout()}
+        {"group_size": 32, "layout": Int4CPULayout(), "version": 1}
         if is_torchao_available() and version.parse(importlib.metadata.version("torchao")) >= version.parse("0.8.0")
         else {"group_size": 32}
     )
@@ -225,6 +228,7 @@ def test_include_input_output_embeddings(self):
             weight_dtype=weight_dtype,
             granularity=granularity,
             mapping_type=mapping_type,
+            version=1,
         )
         config = ModuleFqnToConfig(
             {"_default": None, "model.embed_tokens": embedding_config, "lm_head": embedding_config}
@@ -277,7 +281,7 @@ def test_per_module_config_skip(self):
 @require_torch_accelerator
 class TorchAoAcceleratorTest(TorchAoTest):
     device = torch_device
-    quant_scheme_kwargs = {"group_size": 32}
+    quant_scheme_kwargs = {"group_size": 32, "version": 1}
 
     # called only once for all test in this class
     @classmethod
@@ -327,7 +331,7 @@ def test_int4wo_offload(self):
             "lm_head": 0,
         }
 
-        quant_config = TorchAoConfig("int4_weight_only", group_size=32)
+        quant_config = TorchAoConfig("int4_weight_only", **self.quant_scheme_kwargs)
 
         quantized_model = AutoModelForCausalLM.from_pretrained(
             self.model_name,
@@ -399,7 +403,7 @@ def test_autoquant(self):
 
         check_autoquantized(self, quantized_model.model.layers[0].self_attn.v_proj)
 
-        EXPECTED_OUTPUT = "What are we having for dinner?\n\nJane: (sighs)"
+        EXPECTED_OUTPUT = "What are we having for dinner?\n\nJessica: (smiling)"
         output = quantized_model.generate(
             **input_ids, max_new_tokens=self.max_new_tokens, cache_implementation="static"
         )
@@ -414,7 +418,7 @@ class TorchAoSerializationTest(unittest.TestCase):
     model_name = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
     quant_scheme = "int4_weight_only"
     quant_scheme_kwargs = (
-        {"group_size": 32, "layout": Int4CPULayout()}
+        {"group_size": 32, "layout": Int4CPULayout(), "version": 1}
         if is_torchao_available() and version.parse(importlib.metadata.version("torchao")) >= version.parse("0.8.0")
         else {"group_size": 32}
     )
@@ -447,13 +451,13 @@ def test_original_model_expected_output(self):
 
         self.assertEqual(self.tokenizer.decode(output[0], skip_special_tokens=True), self.EXPECTED_OUTPUT)
 
-    def check_serialization_expected_output(self, device, expected_output):
+    def check_serialization_expected_output(self, device, expected_output, safe_serialization=False):
         """
         Test if we can serialize and load/infer the model again on the same device
         """
         dtype = torch.bfloat16 if self.quant_scheme == "int4_weight_only" else "auto"
         with tempfile.TemporaryDirectory() as tmpdirname:
-            self.quantized_model.save_pretrained(tmpdirname, safe_serialization=False)
+            self.quantized_model.save_pretrained(tmpdirname, safe_serialization=safe_serialization)
             loaded_quantized_model = AutoModelForCausalLM.from_pretrained(tmpdirname, dtype=dtype, device_map=device)
             input_ids = self.tokenizer(self.input_text, return_tensors="pt").to(device)
 
@@ -464,6 +468,48 @@ def test_serialization_expected_output(self):
         self.check_serialization_expected_output(self.device, self.EXPECTED_OUTPUT)
 
 
+@require_torchao
+@require_torchao_version_greater_or_equal("0.14.0")
+class TorchAoSafeSerializationTest(TorchAoSerializationTest):
+    # called only once for all test in this class
+    @classmethod
+    def setUpClass(cls):
+        cls.tokenizer = AutoTokenizer.from_pretrained(cls.model_name)
+        cls.EXPECTED_OUTPUT = "What are we having for dinner?\n- 1. What is the temperature outside"
+
+    def tearDown(self):
+        gc.collect()
+        backend_empty_cache(torch_device)
+        gc.collect()
+        if hasattr(self, "quantized_model"):
+            del self.quantized_model
+        gc.collect()
+
+    test_params = (
+        [
+            (
+                torchao.quantization.Float8DynamicActivationFloat8WeightConfig(),
+                "What are we having for dinner?\n\nJess: (smiling) I",
+            ),
+            (torchao.quantization.Float8WeightOnlyConfig(), "What are we having for dinner?\n\nJessica: (smiling)"),
+        ]
+        if is_torchao_available()
+        else []
+    )
+
+    @parameterized.expand(test_params, skip_on_empty=True)
+    def test_serialization_expected_output(self, config, expected_output):
+        device = "cuda"
+        self.quant_config = TorchAoConfig(config)
+        self.quantized_model = AutoModelForCausalLM.from_pretrained(
+            self.model_name,
+            dtype=torch.bfloat16,
+            device_map=device,
+            quantization_config=self.quant_config,
+        )
+        self.check_serialization_expected_output(device, expected_output, safe_serialization=True)
+
+
 class TorchAoSerializationW8A8CPUTest(TorchAoSerializationTest):
     quant_scheme, quant_scheme_kwargs = "int8_dynamic_activation_int8_weight", {}
 
@@ -500,7 +546,7 @@ def test_serialization_expected_output_on_accelerator(self):
 
 @require_torch_accelerator
 class TorchAoSerializationAcceleratorTest(TorchAoSerializationTest):
-    quant_scheme, quant_scheme_kwargs = "int4_weight_only", {"group_size": 32}
+    quant_scheme, quant_scheme_kwargs = "int4_weight_only", {"group_size": 32, "version": 1}
     device = f"{torch_device}:0"
 
     # called only once for all test in this class