Update src/caching.py

src/caching.py

@@ -174,6 +174,155 @@ def apply_cache_on_transformer(
 def apply_cache_on_pipe(pipe: DiffusionPipeline, *, shallow_patch: bool = False, **kwargs):
     original_call = pipe.__class__.__call__
 
+    if not getattr(original_call, "_is_cached", False):
+        @functools.wraps(original_call)
+        def new_call(self, *args, **kwargs):
+            with cache_context(create_cache_context()):
+                return original_call(self, *args, **kwargs)
+
+        pipe.__class__.__call__ = new_call
+        new_call._is_cached = True
+
+    if not shallow_patch:
+        apply_cache_on_transformer(pipe.transformer, **kwargs)
+
+    pipe._is_cached = True
+    return pipe
+
+@dataclasses.dataclass
+class CacheContext:
+    buffers: Dict[str, torch.Tensor] = dataclasses.field(default_factory=dict)
+    incremental_name_counters: DefaultDict[str, int] = dataclasses.field(default_factory=lambda: defaultdict(int))
+
+    def get_buffer(self, name):
+        return self.buffers.get(name)
+
+    def set_buffer(self, name, buffer):
+        self.buffers[name] = buffer
+
+    def clear_buffers(self):
+        self.buffers.clear()
+
+_current_cache_context = None
+
+def create_cache_context():
+    return CacheContext()
+
+def get_current_cache_context():
+    return _current_cache_context
+
+def set_current_cache_context(cache_context=None):
+    global _current_cache_context
+    _current_cache_context = cache_context
+
+@contextlib.contextmanager
+def cache_context(cache_context):
+    global _current_cache_context
+    old_cache_context = _current_cache_context
+    _current_cache_context = cache_context
+    try:
+        yield
+    finally:
+        _current_cache_context = old_cache_context
+
+def are_two_tensors_similar(t1, t2, *, threshold=0.85):
+    mean_diff = (t1 - t2).abs().mean()
+    mean_t1 = t1.abs().mean()
+    diff = mean_diff / mean_t1
+    return diff.item() < threshold
+
+class CachedTransformerBlocks(torch.nn.Module):
+    def __init__(
+        self,
+        transformer_blocks,
+        single_transformer_blocks=None,
+        *,
+        transformer=None,
+        residual_diff_threshold=0.05,
+    ):
+        super().__init__()
+        self.transformer = transformer
+        self.transformer_blocks = transformer_blocks
+        self.single_transformer_blocks = single_transformer_blocks
+        self.residual_diff_threshold = residual_diff_threshold
+
+    def forward(self, encoder_hidden_states, hidden_states, *args, **kwargs):
+        # Important: For Flux, the order is encoder_hidden_states, hidden_states
+        original_encoder_states = encoder_hidden_states
+
+        # Process first block
+        encoder_hidden_states, hidden_states = self.transformer_blocks[0](
+            encoder_hidden_states, hidden_states, *args, **kwargs
+        )
+
+        # Calculate residual for encoder states
+        first_residual = encoder_hidden_states - original_encoder_states
+
+        cache_context = get_current_cache_context()
+        prev_residual = cache_context.get_buffer("first_residual")
+        can_use_cache = prev_residual is not None and are_two_tensors_similar(
+            prev_residual,
+            first_residual,
+            threshold=self.residual_diff_threshold
+        )
+
+        if can_use_cache:
+            residual = cache_context.get_buffer("residual")
+            encoder_hidden_states = encoder_hidden_states + residual
+        else:
+            cache_context.set_buffer("first_residual", first_residual)
+
+            # Process remaining blocks
+            for block in self.transformer_blocks[1:]:
+                encoder_hidden_states, hidden_states = block(
+                    encoder_hidden_states, hidden_states, *args, **kwargs
+                )
+
+            cache_context.set_buffer("residual", encoder_hidden_states - original_encoder_states)
+
+        return encoder_hidden_states, hidden_states
+
+def apply_cache_on_transformer(
+    transformer: FluxTransformer2DModel,
+    *,
+    residual_diff_threshold=0.05,
+):
+    cached_transformer_blocks = torch.nn.ModuleList([
+        CachedTransformerBlocks(
+            transformer.transformer_blocks,
+            transformer.single_transformer_blocks if hasattr(transformer, 'single_transformer_blocks') else None,
+            transformer=transformer,
+            residual_diff_threshold=residual_diff_threshold,
+        )
+    ])
+    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
+
+def apply_cache_on_pipe(
+    pipe: DiffusionPipeline,
+    *,
+    shallow_patch: bool = False,
+    **kwargs,
+):
+    original_call = pipe.__class__.__call__
+
     if not getattr(original_call, "_is_cached", False):
         @functools.wraps(original_call)
         def new_call(self, *args, **kwargs):