diff --git a/tests/quantization/bnb/test_4bit.py b/tests/quantization/bnb/test_4bit.py index 2d8b9f698bfe..c6d59e8b71ed 100644 --- a/tests/quantization/bnb/test_4bit.py +++ b/tests/quantization/bnb/test_4bit.py @@ -881,4 +881,4 @@ def test_torch_compile_with_cpu_offload(self): super()._test_torch_compile_with_cpu_offload(quantization_config=self.quantization_config) def test_torch_compile_with_group_offload(self): - super()._test_torch_compile_with_group_offload(quantization_config=self.quantization_config) + super()._test_torch_compile_with_group_offload_leaf_stream(quantization_config=self.quantization_config) diff --git a/tests/quantization/bnb/test_mixed_int8.py b/tests/quantization/bnb/test_mixed_int8.py index b15a9f72a8f6..fc4d6127fef9 100644 --- a/tests/quantization/bnb/test_mixed_int8.py +++ b/tests/quantization/bnb/test_mixed_int8.py @@ -845,6 +845,6 @@ def test_torch_compile_with_cpu_offload(self): @pytest.mark.xfail(reason="Test fails because of an offloading problem from Accelerate with confusion in hooks.") def test_torch_compile_with_group_offload(self): - super()._test_torch_compile_with_group_offload( + super()._test_torch_compile_with_group_offload_leaf_stream( quantization_config=self.quantization_config, torch_dtype=torch.float16 ) diff --git a/tests/quantization/test_torch_compile_utils.py b/tests/quantization/test_torch_compile_utils.py index 1ae77b27d7cd..1205d0baf93e 100644 --- a/tests/quantization/test_torch_compile_utils.py +++ b/tests/quantization/test_torch_compile_utils.py @@ -64,7 +64,29 @@ def _test_torch_compile_with_cpu_offload(self, quantization_config, torch_dtype= # small resolutions to ensure speedy execution. pipe("a dog", num_inference_steps=3, max_sequence_length=16, height=256, width=256) - def _test_torch_compile_with_group_offload(self, quantization_config, torch_dtype=torch.bfloat16): + def _test_torch_compile_with_group_offload_leaf(self, quantization_config, torch_dtype=torch.bfloat16): + torch._dynamo.config.cache_size_limit = 10000 + + pipe = self._init_pipeline(quantization_config, torch_dtype) + group_offload_kwargs = { + "onload_device": torch.device("cuda"), + "offload_device": torch.device("cpu"), + "offload_type": "leaf_level", + "num_blocks_per_group": 1, + "use_stream": False, + } + pipe.transformer.enable_group_offload(**group_offload_kwargs) + pipe.transformer.compile() + for name, component in pipe.components.items(): + if name != "transformer" and isinstance(component, torch.nn.Module): + if torch.device(component.device).type == "cpu": + component.to("cuda") + + for _ in range(2): + # small resolutions to ensure speedy execution. + pipe("a dog", num_inference_steps=3, max_sequence_length=16, height=256, width=256) + + def _test_torch_compile_with_group_offload_leaf_stream(self, quantization_config, torch_dtype=torch.bfloat16): torch._dynamo.config.cache_size_limit = 10000 pipe = self._init_pipeline(quantization_config, torch_dtype) @@ -73,7 +95,6 @@ def _test_torch_compile_with_group_offload(self, quantization_config, torch_dtyp "offload_device": torch.device("cpu"), "offload_type": "leaf_level", "use_stream": True, - "non_blocking": True, } pipe.transformer.enable_group_offload(**group_offload_kwargs) pipe.transformer.compile() diff --git a/tests/quantization/torchao/test_torchao.py b/tests/quantization/torchao/test_torchao.py index 743da17356f7..e708fbbbb3ae 100644 --- a/tests/quantization/torchao/test_torchao.py +++ b/tests/quantization/torchao/test_torchao.py @@ -29,6 +29,7 @@ TorchAoConfig, ) from diffusers.models.attention_processor import Attention +from diffusers.quantizers import PipelineQuantizationConfig from diffusers.utils.testing_utils import ( backend_empty_cache, backend_synchronize, @@ -44,6 +45,8 @@ torch_device, ) +from ..test_torch_compile_utils import QuantCompileTests + enable_full_determinism() @@ -625,6 +628,48 @@ def test_int_a16w8_cpu(self): self._check_serialization_expected_slice(quant_method, quant_method_kwargs, expected_slice, device) +@require_torchao_version_greater_or_equal("0.7.0") +class TorchAoCompileTest(QuantCompileTests): + quantization_config = PipelineQuantizationConfig( + quant_mapping={ + "transformer": TorchAoConfig(quant_type="int8_weight_only"), + }, + ) + + def test_torch_compile(self): + super()._test_torch_compile(quantization_config=self.quantization_config) + + @unittest.skip( + "Changing the device of AQT tensor with module._apply (called from doing module.to() in accelerate) does not work " + "when compiling." + ) + def test_torch_compile_with_cpu_offload(self): + # RuntimeError: _apply(): Couldn't swap Linear.weight + super()._test_torch_compile_with_cpu_offload(quantization_config=self.quantization_config) + + @unittest.skip( + "Changing the device of AQT tensor, with `param.data = param.data.to(device)` as done in group offloading implementation " + "is unsupported in TorchAO. When compiling, FakeTensor device mismatch causes failure." + ) + def test_torch_compile_with_group_offload_leaf(self): + # If we run group offloading without compilation, we will see: + # RuntimeError: Attempted to set the storage of a tensor on device "cpu" to a storage on different device "cuda:0". This is no longer allowed; the devices must match. + # When running with compilation, the error ends up being different: + # Dynamo failed to run FX node with fake tensors: call_function (*(FakeTensor(..., device='cuda:0', size=(s0, 256), dtype=torch.bfloat16), AffineQuantizedTensor(tensor_impl=PlainAQTTensorImpl(data=FakeTensor(..., size=(1536, 256), dtype=torch.int8)... , scale=FakeTensor(..., size=(1536,), dtype=torch.bfloat16)... , zero_point=FakeTensor(..., size=(1536,), dtype=torch.int64)... , _layout=PlainLayout()), block_size=(1, 256), shape=torch.Size([1536, 256]), device=cpu, dtype=torch.bfloat16, requires_grad=False), Parameter(FakeTensor(..., device='cuda:0', size=(1536,), dtype=torch.bfloat16, + # requires_grad=True))), **{}): got RuntimeError('Unhandled FakeTensor Device Propagation for aten.mm.default, found two different devices cuda:0, cpu') + # Looks like something that will have to be looked into upstream. + # for linear layers, weight.tensor_impl shows cuda... but: + # weight.tensor_impl.{data,scale,zero_point}.device will be cpu + super()._test_torch_compile_with_group_offload_leaf(quantization_config=self.quantization_config) + + @unittest.skip( + "Using non-default stream requires ability to pin tensors. AQT does not seem to support this yet in TorchAO." + ) + def test_torch_compile_with_group_offload_leaf_stream(self): + # NotImplementedError: AffineQuantizedTensor dispatch: attempting to run unimplemented operator/function: func=, types=(,), arg_types=(,), kwarg_types={} + super()._test_torch_compile_with_group_offload_leaf_stream(quantization_config=self.quantization_config) + + # Slices for these tests have been obtained on our aws-g6e-xlarge-plus runners @require_torch @require_torch_accelerator