Upload 2 files

optimization.py

@@ -0,0 +1,130 @@
+"""
+"""
+
+from typing import Any
+from typing import Callable
+from typing import ParamSpec
+
+import spaces
+import torch
+from torch.utils._pytree import tree_map_only
+from torchao.quantization import quantize_
+from torchao.quantization import Float8DynamicActivationFloat8WeightConfig
+from torchao.quantization import Int8WeightOnlyConfig
+
+from optimization_utils import capture_component_call
+from optimization_utils import aoti_compile
+from optimization_utils import ZeroGPUCompiledModel
+from optimization_utils import drain_module_parameters
+
+
+P = ParamSpec('P')
+
+
+TRANSFORMER_NUM_FRAMES_DIM = torch.export.Dim('num_frames', min=3, max=21)
+
+TRANSFORMER_DYNAMIC_SHAPES = {
+    'hidden_states': {
+        2: TRANSFORMER_NUM_FRAMES_DIM,
+    },
+}
+
+INDUCTOR_CONFIGS = {
+    'conv_1x1_as_mm': True,
+    'epilogue_fusion': False,
+    'coordinate_descent_tuning': True,
+    'coordinate_descent_check_all_directions': True,
+    'max_autotune': True,
+    'triton.cudagraphs': True,
+}
+
+
+def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kwargs):
+
+    @spaces.GPU(duration=1500)
+    def compile_transformer():
+        
+        pipeline.load_lora_weights(
+            "Kijai/WanVideo_comfy", 
+            weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors", 
+            adapter_name="lightx2v"
+        )
+        kwargs_lora = {}
+        kwargs_lora["load_into_transformer_2"] = True
+        pipeline.load_lora_weights(
+            "Kijai/WanVideo_comfy", 
+            weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors", 
+            adapter_name="lightx2v_2", **kwargs_lora
+        )
+        pipeline.set_adapters(["lightx2v", "lightx2v_2"], adapter_weights=[1., 1.])
+        pipeline.fuse_lora(adapter_names=["lightx2v"], lora_scale=3., components=["transformer"])
+        pipeline.fuse_lora(adapter_names=["lightx2v_2"], lora_scale=1., components=["transformer_2"])
+        pipeline.unload_lora_weights()
+        
+        with capture_component_call(pipeline, 'transformer') as call:
+            pipeline(*args, **kwargs)
+        
+        dynamic_shapes = tree_map_only((torch.Tensor, bool), lambda t: None, call.kwargs)
+        dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES
+
+        quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
+        quantize_(pipeline.transformer_2, Float8DynamicActivationFloat8WeightConfig())
+        
+        hidden_states: torch.Tensor = call.kwargs['hidden_states']
+        hidden_states_transposed = hidden_states.transpose(-1, -2).contiguous()
+        if hidden_states.shape[-1] > hidden_states.shape[-2]:
+            hidden_states_landscape = hidden_states
+            hidden_states_portrait = hidden_states_transposed
+        else:
+            hidden_states_landscape = hidden_states_transposed
+            hidden_states_portrait = hidden_states
+
+        exported_landscape_1 = torch.export.export(
+            mod=pipeline.transformer,
+            args=call.args,
+            kwargs=call.kwargs | {'hidden_states': hidden_states_landscape},
+            dynamic_shapes=dynamic_shapes,
+        )
+        
+        exported_portrait_2 = torch.export.export(
+            mod=pipeline.transformer_2,
+            args=call.args,
+            kwargs=call.kwargs | {'hidden_states': hidden_states_portrait},
+            dynamic_shapes=dynamic_shapes,
+        )
+
+        compiled_landscape_1 = aoti_compile(exported_landscape_1, INDUCTOR_CONFIGS)
+        compiled_portrait_2 = aoti_compile(exported_portrait_2, INDUCTOR_CONFIGS)
+
+        compiled_landscape_2 = ZeroGPUCompiledModel(compiled_landscape_1.archive_file, compiled_portrait_2.weights)
+        compiled_portrait_1 = ZeroGPUCompiledModel(compiled_portrait_2.archive_file, compiled_landscape_1.weights)
+
+        return (
+            compiled_landscape_1,
+            compiled_landscape_2,
+            compiled_portrait_1,
+            compiled_portrait_2,
+        )
+
+    quantize_(pipeline.text_encoder, Int8WeightOnlyConfig())
+    cl1, cl2, cp1, cp2 = compile_transformer()
+
+    def combined_transformer_1(*args, **kwargs):
+        hidden_states: torch.Tensor = kwargs['hidden_states']
+        if hidden_states.shape[-1] > hidden_states.shape[-2]:
+            return cl1(*args, **kwargs)
+        else:
+            return cp1(*args, **kwargs)
+
+    def combined_transformer_2(*args, **kwargs):
+        hidden_states: torch.Tensor = kwargs['hidden_states']
+        if hidden_states.shape[-1] > hidden_states.shape[-2]:
+            return cl2(*args, **kwargs)
+        else:
+            return cp2(*args, **kwargs)
+
+    pipeline.transformer.forward = combined_transformer_1
+    drain_module_parameters(pipeline.transformer)
+
+    pipeline.transformer_2.forward = combined_transformer_2
+    drain_module_parameters(pipeline.transformer_2)

optimization_utils.py

@@ -0,0 +1,107 @@
+"""
+"""
+import contextlib
+from contextvars import ContextVar
+from io import BytesIO
+from typing import Any
+from typing import cast
+from unittest.mock import patch
+
+import torch
+from torch._inductor.package.package import package_aoti
+from torch.export.pt2_archive._package import AOTICompiledModel
+from torch.export.pt2_archive._package_weights import Weights
+
+
+INDUCTOR_CONFIGS_OVERRIDES = {
+    'aot_inductor.package_constants_in_so': False,
+    'aot_inductor.package_constants_on_disk': True,
+    'aot_inductor.package': True,
+}
+
+
+class ZeroGPUWeights:
+    def __init__(self, constants_map: dict[str, torch.Tensor], to_cuda: bool = False):
+        if to_cuda:
+            self.constants_map = {name: tensor.to('cuda') for name, tensor in constants_map.items()}
+        else:
+            self.constants_map = constants_map
+    def __reduce__(self):
+        constants_map: dict[str, torch.Tensor] = {}
+        for name, tensor in self.constants_map.items():
+            tensor_ = torch.empty_like(tensor, device='cpu').pin_memory()
+            constants_map[name] = tensor_.copy_(tensor).detach().share_memory_()
+        return ZeroGPUWeights, (constants_map, True)
+
+
+class ZeroGPUCompiledModel:
+    def __init__(self, archive_file: torch.types.FileLike, weights: ZeroGPUWeights):
+        self.archive_file = archive_file
+        self.weights = weights
+        self.compiled_model: ContextVar[AOTICompiledModel | None] = ContextVar('compiled_model', default=None)
+    def __call__(self, *args, **kwargs):
+        if (compiled_model := self.compiled_model.get()) is None:
+            compiled_model = cast(AOTICompiledModel, torch._inductor.aoti_load_package(self.archive_file))
+            compiled_model.load_constants(self.weights.constants_map, check_full_update=True, user_managed=True)
+            self.compiled_model.set(compiled_model)
+        return compiled_model(*args, **kwargs)
+    def __reduce__(self):
+        return ZeroGPUCompiledModel, (self.archive_file, self.weights)
+
+
+def aoti_compile(
+    exported_program: torch.export.ExportedProgram,
+    inductor_configs: dict[str, Any] | None = None,
+):
+    inductor_configs = (inductor_configs or {}) | INDUCTOR_CONFIGS_OVERRIDES
+    gm = cast(torch.fx.GraphModule, exported_program.module())
+    assert exported_program.example_inputs is not None
+    args, kwargs = exported_program.example_inputs
+    artifacts = torch._inductor.aot_compile(gm, args, kwargs, options=inductor_configs)
+    archive_file = BytesIO()
+    files: list[str | Weights] = [file for file in artifacts if isinstance(file, str)]
+    package_aoti(archive_file, files)
+    weights, = (artifact for artifact in artifacts if isinstance(artifact, Weights))
+    zerogpu_weights = ZeroGPUWeights({name: weights.get_weight(name)[0] for name in weights})
+    return ZeroGPUCompiledModel(archive_file, zerogpu_weights)
+
+
+@contextlib.contextmanager
+def capture_component_call(
+    pipeline: Any,
+    component_name: str,
+    component_method='forward',
+):
+
+    class CapturedCallException(Exception):
+        def __init__(self, *args, **kwargs):
+            super().__init__()
+            self.args = args
+            self.kwargs = kwargs
+
+    class CapturedCall:
+        def __init__(self):
+            self.args: tuple[Any, ...] = ()
+            self.kwargs: dict[str, Any] = {}
+
+    component = getattr(pipeline, component_name)
+    captured_call = CapturedCall()
+
+    def capture_call(*args, **kwargs):
+        raise CapturedCallException(*args, **kwargs)
+
+    with patch.object(component, component_method, new=capture_call):
+        try:
+            yield captured_call
+        except CapturedCallException as e:
+            captured_call.args = e.args
+            captured_call.kwargs = e.kwargs
+
+
+def drain_module_parameters(module: torch.nn.Module):
+    state_dict_meta = {name: {'device': tensor.device, 'dtype': tensor.dtype} for name, tensor in module.state_dict().items()}
+    state_dict = {name: torch.nn.Parameter(torch.empty_like(tensor, device='cpu')) for name, tensor in module.state_dict().items()}
+    module.load_state_dict(state_dict, assign=True)
+    for name, param in state_dict.items():
+        meta = state_dict_meta[name]
+        param.data = torch.Tensor([]).to(**meta)