Add TeaCache and FBCache

README.md

@@ -84,6 +84,20 @@ We also have implemented the following parallel strategies for reference:
 1. Tensor Parallelism
 2. [DistriFusion](https://arxiv.org/abs/2402.19481)
 
+<h3 id="meet-xdit-cache">Cache Acceleration</h3>
+
+Cache method is inspired by work from TeaCache(https://github.com/ali-vilab/TeaCache.git) and ParaAttn(https://github.com/chengzeyi/ParaAttention.git); We adapted the TeaCache and First-Block-Cache in xDiT.
+
+This method is not orthogonal to parallel in xDiT. Only when SP or no parrallelism can activate the cache function.
+
+To use this functionality, you can activate it by `--use_teacache` or `--use_fbcache`, which activate TeaCache and First-Block-Cache respectively. Right now, this repo only supports FLUX model.
+
+The Performance shown as below, tested on 4 H20 with SP=4:
+| 方法           | 性能   |
+|----------------|--------|
+| 原始           | 2.02s  |
+| use_teacache   | 1.58s  |
+| use_fbcache    | 0.93s  |
 
 <h3 id="meet-xdit-perf">Computing Acceleration</h3>
 

examples/flux_example.py

@@ -11,6 +11,10 @@
     get_data_parallel_world_size,
     get_runtime_state,
     is_dp_last_group,
+    get_pipeline_parallel_world_size,
+    get_classifier_free_guidance_world_size,
+    get_tensor_model_parallel_world_size,
+    get_data_parallel_world_size,
 )
 
 
@@ -45,7 +49,27 @@ def main():
     parameter_peak_memory = torch.cuda.max_memory_allocated(device=f"cuda:{local_rank}")
 
     pipe.prepare_run(input_config, steps=1)
+    from xfuser.model_executor.plugins.cache_.diffusers_adapters import apply_cache_on_transformer
+    use_cache = engine_args.use_teacache or engine_args.use_fbcache
+    if (use_cache
+        and get_pipeline_parallel_world_size() == 1
+        and get_classifier_free_guidance_world_size() == 1
+        and get_tensor_model_parallel_world_size() == 1
+    ):
+        cache_args = {
+            "rel_l1_thresh": 0.6,
+            "return_hidden_states_first": False,
+            "num_steps": input_config.num_inference_steps,
+        }
+
+        if engine_args.use_fbcache and engine_args.use_teacache:
+            cache_args["use_cache"] = "Fb"
+        elif engine_args.use_teacache:
+            cache_args["use_cache"] = "Tea"
+        elif engine_args.use_fbcache:
+            cache_args["use_cache"] = "Fb"
 
+        pipe.transformer = apply_cache_on_transformer(pipe.transformer, **cache_args)
     torch.cuda.reset_peak_memory_stats()
     start_time = time.time()
     output = pipe(

examples/run.sh

@@ -10,7 +10,7 @@ declare -A MODEL_CONFIGS=(
     ["Pixart-alpha"]="pixartalpha_example.py /cfs/dit/PixArt-XL-2-1024-MS 20"
     ["Pixart-sigma"]="pixartsigma_example.py /cfs/dit/PixArt-Sigma-XL-2-2K-MS 20"
     ["Sd3"]="sd3_example.py /cfs/dit/stable-diffusion-3-medium-diffusers 20"
-    ["Flux"]="flux_example.py /cfs/dit/FLUX.1-dev 28"
+    ["Flux"]="flux_example.py /cfs/dit/FLUX.1-dev/ 28"
     ["HunyuanDiT"]="hunyuandit_example.py /cfs/dit/HunyuanDiT-v1.2-Diffusers 50"
 )
 
@@ -27,6 +27,9 @@ mkdir -p ./results
 # task args
 TASK_ARGS="--height 1024 --width 1024 --no_use_resolution_binning"
 
+# cache args
+# CACHE_ARGS="--use_teacache"
+# CACHE_ARGS="--use_fbcache"
 
 # On 8 gpus, pp=2, ulysses=2, ring=1, cfg_parallel=2 (split batch)
 N_GPUS=8
@@ -64,3 +67,4 @@ $CFG_ARGS \
 $PARALLLEL_VAE \
 $COMPILE_FLAG \
 $QUANTIZE_FLAG \
+$CACHE_ARGS \

xfuser/config/args.py

@@ -111,6 +111,8 @@ class xFuserArgs:
     window_size: int = 64
     coco_path: Optional[str] = None
     use_cache: bool = False
+    use_teacache: bool = False
+    use_fbcache: bool = False
     use_fp8_t5_encoder: bool = False
 
     @staticmethod
@@ -154,6 +156,16 @@ def add_cli_args(parser: FlexibleArgumentParser):
             action="store_true",
             help="Enable onediff to accelerate inference in a single card",
         )
+        runtime_group.add_argument(
+            "--use_teacache",
+            action="store_true",
+            help="Enable teacache to accelerate inference in a single card",
+        )
+        runtime_group.add_argument(
+            "--use_fbcache",
+            action="store_true",
+            help="Enable teacache to accelerate inference in a single card",
+        )
 
         # Parallel arguments
         parallel_group = parser.add_argument_group("Parallel Processing Options")

xfuser/config/config.py

@@ -60,6 +60,8 @@ class RuntimeConfig:
     use_torch_compile: bool = False
     use_onediff: bool = False
     use_fp8_t5_encoder: bool = False
+    use_teacache: bool = False
+    use_fbcache: bool = False
 
     def __post_init__(self):
         check_packages()

xfuser/core/distributed/group_coordinator.py

@@ -196,7 +196,7 @@ def group_skip_rank(self):
         world_size = self.world_size
         return (world_size - rank_in_group - 1) % world_size
 
-    def all_reduce(self, input_: torch.Tensor) -> torch.Tensor:
+    def all_reduce(self, input_: torch.Tensor, op=torch._C._distributed_c10d.ReduceOp.SUM) -> torch.Tensor:
         """
         NOTE: This operation will be applied in-place or out-of-place.
         Always assume this function modifies its input, but use the return
@@ -206,7 +206,7 @@ def all_reduce(self, input_: torch.Tensor) -> torch.Tensor:
         if self.world_size == 1:
             return input_
         else:
-            torch.distributed.all_reduce(input_, group=self.device_group)
+            torch.distributed.all_reduce(input_, op=op, group=self.device_group)
         return input_
 
     def all_gather(

xfuser/core/distributed/parallel_state.py

@@ -217,6 +217,7 @@ def init_distributed_environment(
             world_size=world_size,
             rank=rank,
         )
+        torch.cuda.set_device(torch.distributed.get_rank() % torch.cuda.device_count())
     # set the local rank
     # local_rank is not available in torch ProcessGroup,
     # see https://github.com/pytorch/pytorch/issues/122816

xfuser/model_executor/plugins/__init__.py

@@ -0,0 +1,4 @@
+"""
+adapted from https://github.com/ali-vilab/TeaCache.git
+adapted from https://github.com/chengzeyi/ParaAttention.git
+"""

xfuser/model_executor/plugins/cache_/__init__.py

@@ -0,0 +1,4 @@
+"""
+adapted from https://github.com/ali-vilab/TeaCache.git
+adapted from https://github.com/chengzeyi/ParaAttention.git
+"""

xfuser/model_executor/plugins/cache_/diffusers_adapters/__init__.py

@@ -0,0 +1,17 @@
+"""
+adapted from https://github.com/ali-vilab/TeaCache.git
+adapted from https://github.com/chengzeyi/ParaAttention.git
+"""
+import importlib
+from typing import Type, Dict, TypeVar
+from xfuser.model_executor.plugins.cache_.diffusers_adapters.registry import TRANSFORMER_ADAPTER_REGISTRY
+
+
+def apply_cache_on_transformer(transformer, *args, **kwargs):
+    adapter_name = TRANSFORMER_ADAPTER_REGISTRY.get(type(transformer))
+    if not adapter_name:
+        raise ValueError(f"Unknown transformer class: {transformer.__class__.__name__}")
+
+    adapter_module = importlib.import_module(f".{adapter_name}", __package__)
+    apply_cache_on_transformer_fn = getattr(adapter_module, "apply_cache_on_transformer")
+    return apply_cache_on_transformer_fn(transformer, *args, **kwargs)

xfuser/model_executor/plugins/cache_/diffusers_adapters/flux.py

@@ -0,0 +1,74 @@
+"""
+adapted from https://github.com/ali-vilab/TeaCache.git
+adapted from https://github.com/chengzeyi/ParaAttention.git
+"""
+import functools
+import unittest
+
+import torch
+from torch import nn
+from diffusers import DiffusionPipeline, FluxTransformer2DModel
+from xfuser.model_executor.plugins.cache_.diffusers_adapters.registry import TRANSFORMER_ADAPTER_REGISTRY
+
+from xfuser.model_executor.plugins.cache_ import utils
+
+def create_cached_transformer_blocks(use_cache, transformer, rel_l1_thresh, return_hidden_states_first, num_steps):
+    cached_transformer_class = {
+        "Fb": utils.FBCachedTransformerBlocks,
+        "Tea": utils.TeaCachedTransformerBlocks,
+    }.get(use_cache)
+
+    if not cached_transformer_class:
+        raise ValueError(f"Unsupported use_cache value: {use_cache}")
+
+    return cached_transformer_class(
+        transformer.transformer_blocks,
+        transformer.single_transformer_blocks,
+        transformer=transformer,
+        rel_l1_thresh=rel_l1_thresh,
+        return_hidden_states_first=return_hidden_states_first,
+        num_steps=num_steps,
+        name=TRANSFORMER_ADAPTER_REGISTRY.get(type(transformer)),
+    )
+
+
+def apply_cache_on_transformer(
+    transformer: FluxTransformer2DModel,
+    *,
+    rel_l1_thresh=0.6,
+    return_hidden_states_first=False,
+    num_steps=8,
+    use_cache="Fb",
+):
+    cached_transformer_blocks = nn.ModuleList([
+        create_cached_transformer_blocks(use_cache, transformer, rel_l1_thresh, return_hidden_states_first, num_steps)
+    ])
+
+    dummy_single_transformer_blocks = torch.nn.ModuleList()
+
+    original_forward = transformer.forward
+
+    @functools.wraps(original_forward)
+    def new_forward(
+        self,
+        *args,
+        **kwargs,
+    ):
+        with unittest.mock.patch.object(
+            self,
+            "transformer_blocks",
+            cached_transformer_blocks,
+        ), unittest.mock.patch.object(
+            self,
+            "single_transformer_blocks",
+            dummy_single_transformer_blocks,
+        ):
+            return original_forward(
+                *args,
+                **kwargs,
+            )
+
+    transformer.forward = new_forward.__get__(transformer)
+
+    return transformer
+

xfuser/model_executor/plugins/cache_/diffusers_adapters/registry.py

@@ -0,0 +1,16 @@
+"""
+adapted from https://github.com/ali-vilab/TeaCache.git
+adapted from https://github.com/chengzeyi/ParaAttention.git
+"""
+from typing import Type, Dict
+from diffusers.models.transformers.transformer_flux import FluxTransformer2DModel
+from xfuser.model_executor.models.transformers.transformer_flux import xFuserFluxTransformer2DWrapper
+
+TRANSFORMER_ADAPTER_REGISTRY: Dict[Type, str] = {}
+
+def register_transformer_adapter(transformer_class: Type, adapter_name: str):
+    TRANSFORMER_ADAPTER_REGISTRY[transformer_class] = adapter_name
+
+register_transformer_adapter(FluxTransformer2DModel, "flux")
+register_transformer_adapter(xFuserFluxTransformer2DWrapper, "flux")
+