support HyperTile optimization

modules/hypertile.py

@@ -0,0 +1,371 @@
+"""
+Hypertile module for splitting attention layers in SD-1.5 U-Net and SD-1.5 VAE
+Warn : The patch works well only if the input image has a width and height that are multiples of 128
+Author : @tfernd Github : https://github.com/tfernd/HyperTile
+"""
+
+from __future__ import annotations
+from typing import Callable
+from typing_extensions import Literal
+
+import logging
+from functools import wraps, cache
+from contextlib import contextmanager
+
+import math
+import torch.nn as nn
+import random
+
+from einops import rearrange
+
+# TODO add SD-XL layers
+DEPTH_LAYERS = {
+    0: [
+        # SD 1.5 U-Net (diffusers)
+        "down_blocks.0.attentions.0.transformer_blocks.0.attn1",
+        "down_blocks.0.attentions.1.transformer_blocks.0.attn1",
+        "up_blocks.3.attentions.0.transformer_blocks.0.attn1",
+        "up_blocks.3.attentions.1.transformer_blocks.0.attn1",
+        "up_blocks.3.attentions.2.transformer_blocks.0.attn1",
+        # SD 1.5 U-Net (ldm)
+        "input_blocks.1.1.transformer_blocks.0.attn1",
+        "input_blocks.2.1.transformer_blocks.0.attn1",
+        "output_blocks.9.1.transformer_blocks.0.attn1",
+        "output_blocks.10.1.transformer_blocks.0.attn1",
+        "output_blocks.11.1.transformer_blocks.0.attn1",
+        # SD 1.5 VAE
+        "decoder.mid_block.attentions.0",
+        "decoder.mid.attn_1",
+    ],
+    1: [
+        # SD 1.5 U-Net (diffusers)
+        "down_blocks.1.attentions.0.transformer_blocks.0.attn1",
+        "down_blocks.1.attentions.1.transformer_blocks.0.attn1",
+        "up_blocks.2.attentions.0.transformer_blocks.0.attn1",
+        "up_blocks.2.attentions.1.transformer_blocks.0.attn1",
+        "up_blocks.2.attentions.2.transformer_blocks.0.attn1",
+        # SD 1.5 U-Net (ldm)
+        "input_blocks.4.1.transformer_blocks.0.attn1",
+        "input_blocks.5.1.transformer_blocks.0.attn1",
+        "output_blocks.6.1.transformer_blocks.0.attn1",
+        "output_blocks.7.1.transformer_blocks.0.attn1",
+        "output_blocks.8.1.transformer_blocks.0.attn1",
+    ],
+    2: [
+        # SD 1.5 U-Net (diffusers)
+        "down_blocks.2.attentions.0.transformer_blocks.0.attn1",
+        "down_blocks.2.attentions.1.transformer_blocks.0.attn1",
+        "up_blocks.1.attentions.0.transformer_blocks.0.attn1",
+        "up_blocks.1.attentions.1.transformer_blocks.0.attn1",
+        "up_blocks.1.attentions.2.transformer_blocks.0.attn1",
+        # SD 1.5 U-Net (ldm)
+        "input_blocks.7.1.transformer_blocks.0.attn1",
+        "input_blocks.8.1.transformer_blocks.0.attn1",
+        "output_blocks.3.1.transformer_blocks.0.attn1",
+        "output_blocks.4.1.transformer_blocks.0.attn1",
+        "output_blocks.5.1.transformer_blocks.0.attn1",
+    ],
+    3: [
+        # SD 1.5 U-Net (diffusers)
+        "mid_block.attentions.0.transformer_blocks.0.attn1",
+        # SD 1.5 U-Net (ldm)
+        "middle_block.1.transformer_blocks.0.attn1",
+    ],
+}
+# XL layers, thanks for GitHub@gel-crabs for the help
+DEPTH_LAYERS_XL = {
+    0: [
+        # SD 1.5 U-Net (diffusers)
+        "down_blocks.0.attentions.0.transformer_blocks.0.attn1",
+        "down_blocks.0.attentions.1.transformer_blocks.0.attn1",
+        "up_blocks.3.attentions.0.transformer_blocks.0.attn1",
+        "up_blocks.3.attentions.1.transformer_blocks.0.attn1",
+        "up_blocks.3.attentions.2.transformer_blocks.0.attn1",
+        # SD 1.5 U-Net (ldm)
+        "input_blocks.4.1.transformer_blocks.0.attn1",
+        "input_blocks.5.1.transformer_blocks.0.attn1",
+        "output_blocks.3.1.transformer_blocks.0.attn1",
+        "output_blocks.4.1.transformer_blocks.0.attn1",
+        "output_blocks.5.1.transformer_blocks.0.attn1",
+        # SD 1.5 VAE
+        "decoder.mid_block.attentions.0",
+        "decoder.mid.attn_1",
+    ],
+    1: [
+        # SD 1.5 U-Net (diffusers)
+        #"down_blocks.1.attentions.0.transformer_blocks.0.attn1",
+        #"down_blocks.1.attentions.1.transformer_blocks.0.attn1",
+        #"up_blocks.2.attentions.0.transformer_blocks.0.attn1",
+        #"up_blocks.2.attentions.1.transformer_blocks.0.attn1",
+        #"up_blocks.2.attentions.2.transformer_blocks.0.attn1",
+        # SD 1.5 U-Net (ldm)
+        "input_blocks.4.1.transformer_blocks.1.attn1",
+        "input_blocks.5.1.transformer_blocks.1.attn1",
+        "output_blocks.3.1.transformer_blocks.1.attn1",
+        "output_blocks.4.1.transformer_blocks.1.attn1",
+        "output_blocks.5.1.transformer_blocks.1.attn1",
+        "input_blocks.7.1.transformer_blocks.0.attn1",
+        "input_blocks.8.1.transformer_blocks.0.attn1",
+        "output_blocks.0.1.transformer_blocks.0.attn1",
+        "output_blocks.1.1.transformer_blocks.0.attn1",
+        "output_blocks.2.1.transformer_blocks.0.attn1",
+        "input_blocks.7.1.transformer_blocks.1.attn1",
+        "input_blocks.8.1.transformer_blocks.1.attn1",
+        "output_blocks.0.1.transformer_blocks.1.attn1",
+        "output_blocks.1.1.transformer_blocks.1.attn1",
+        "output_blocks.2.1.transformer_blocks.1.attn1",
+        "input_blocks.7.1.transformer_blocks.2.attn1",
+        "input_blocks.8.1.transformer_blocks.2.attn1",
+        "output_blocks.0.1.transformer_blocks.2.attn1",
+        "output_blocks.1.1.transformer_blocks.2.attn1",
+        "output_blocks.2.1.transformer_blocks.2.attn1",
+        "input_blocks.7.1.transformer_blocks.3.attn1",
+        "input_blocks.8.1.transformer_blocks.3.attn1",
+        "output_blocks.0.1.transformer_blocks.3.attn1",
+        "output_blocks.1.1.transformer_blocks.3.attn1",
+        "output_blocks.2.1.transformer_blocks.3.attn1",
+        "input_blocks.7.1.transformer_blocks.4.attn1",
+        "input_blocks.8.1.transformer_blocks.4.attn1",
+        "output_blocks.0.1.transformer_blocks.4.attn1",
+        "output_blocks.1.1.transformer_blocks.4.attn1",
+        "output_blocks.2.1.transformer_blocks.4.attn1",
+        "input_blocks.7.1.transformer_blocks.5.attn1",
+        "input_blocks.8.1.transformer_blocks.5.attn1",
+        "output_blocks.0.1.transformer_blocks.5.attn1",
+        "output_blocks.1.1.transformer_blocks.5.attn1",
+        "output_blocks.2.1.transformer_blocks.5.attn1",
+        "input_blocks.7.1.transformer_blocks.6.attn1",
+        "input_blocks.8.1.transformer_blocks.6.attn1",
+        "output_blocks.0.1.transformer_blocks.6.attn1",
+        "output_blocks.1.1.transformer_blocks.6.attn1",
+        "output_blocks.2.1.transformer_blocks.6.attn1",
+        "input_blocks.7.1.transformer_blocks.7.attn1",
+        "input_blocks.8.1.transformer_blocks.7.attn1",
+        "output_blocks.0.1.transformer_blocks.7.attn1",
+        "output_blocks.1.1.transformer_blocks.7.attn1",
+        "output_blocks.2.1.transformer_blocks.7.attn1",
+        "input_blocks.7.1.transformer_blocks.8.attn1",
+        "input_blocks.8.1.transformer_blocks.8.attn1",
+        "output_blocks.0.1.transformer_blocks.8.attn1",
+        "output_blocks.1.1.transformer_blocks.8.attn1",
+        "output_blocks.2.1.transformer_blocks.8.attn1",
+        "input_blocks.7.1.transformer_blocks.9.attn1",
+        "input_blocks.8.1.transformer_blocks.9.attn1",
+        "output_blocks.0.1.transformer_blocks.9.attn1",
+        "output_blocks.1.1.transformer_blocks.9.attn1",
+        "output_blocks.2.1.transformer_blocks.9.attn1",
+    ],
+    2: [
+        # SD 1.5 U-Net (diffusers)
+        "mid_block.attentions.0.transformer_blocks.0.attn1",
+        # SD 1.5 U-Net (ldm)
+        "middle_block.1.transformer_blocks.0.attn1",
+        "middle_block.1.transformer_blocks.1.attn1",
+        "middle_block.1.transformer_blocks.2.attn1",
+        "middle_block.1.transformer_blocks.3.attn1",
+        "middle_block.1.transformer_blocks.4.attn1",
+        "middle_block.1.transformer_blocks.5.attn1",
+        "middle_block.1.transformer_blocks.6.attn1",
+        "middle_block.1.transformer_blocks.7.attn1",
+        "middle_block.1.transformer_blocks.8.attn1",
+        "middle_block.1.transformer_blocks.9.attn1",
+    ],
+    3 : [] # TODO - separate layers for SD-XL
+}
+
+
+RNG_INSTANCE = random.Random()
+
+def random_divisor(value: int, min_value: int, /, max_options: int = 1) -> int:
+    """
+    Returns a random divisor of value that
+        x * min_value <= value
+    if max_options is 1, the behavior is deterministic
+    """
+    min_value = min(min_value, value)
+
+    # All big divisors of value (inclusive)
+    divisors = [i for i in range(min_value, value + 1) if value % i == 0] # divisors in small -> big order
+
+    ns = [value // i for i in divisors[:max_options]]  # has at least 1 element # big -> small order
+
+    idx = RNG_INSTANCE.randint(0, len(ns) - 1)
+
+    return ns[idx]
+
+def set_hypertile_seed(seed: int) -> None:
+    RNG_INSTANCE.seed(seed)
+
+def largest_tile_size_available(width:int, height:int) -> int:
+    """
+    Calculates the largest tile size available for a given width and height
+    Tile size is always a power of 2
+    """
+    gcd = math.gcd(width, height)
+    largest_tile_size_available = 1
+    while gcd % (largest_tile_size_available * 2) == 0:
+        largest_tile_size_available *= 2
+    return largest_tile_size_available
+
+def iterative_closest_divisors(hw:int, aspect_ratio:float) -> tuple[int, int]:
+    """
+    Finds h and w such that h*w = hw and h/w = aspect_ratio
+    We check all possible divisors of hw and return the closest to the aspect ratio
+    """
+    divisors = [i for i in range(2, hw + 1) if hw % i == 0] # all divisors of hw
+    pairs = [(i, hw // i) for i in divisors] # all pairs of divisors of hw
+    ratios = [w/h for h, w in pairs] # all ratios of pairs of divisors of hw
+    closest_ratio = min(ratios, key=lambda x: abs(x - aspect_ratio)) # closest ratio to aspect_ratio
+    closest_pair = pairs[ratios.index(closest_ratio)] # closest pair of divisors to aspect_ratio
+    return closest_pair
+
+@cache
+def find_hw_candidates(hw:int, aspect_ratio:float) -> tuple[int, int]:
+    """
+    Finds h and w such that h*w = hw and h/w = aspect_ratio
+    """
+    h, w = round(math.sqrt(hw * aspect_ratio)), round(math.sqrt(hw / aspect_ratio))
+    # find h and w such that h*w = hw and h/w = aspect_ratio
+    if h * w != hw:
+        w_candidate = hw / h
+        # check if w is an integer
+        if not w_candidate.is_integer():
+            h_candidate = hw / w
+            # check if h is an integer
+            if not h_candidate.is_integer():
+                return iterative_closest_divisors(hw, aspect_ratio)
+            else:
+                h = int(h_candidate)
+        else:
+            w = int(w_candidate)
+    return h, w
+
+@contextmanager
+def split_attention(
+    layer: nn.Module,
+    /,
+    aspect_ratio: float,  # width/height
+    tile_size: int = 128,  # 128 for VAE
+    swap_size: int = 1,  # 1 for VAE
+    *,
+    disable: bool = False,
+    max_depth: Literal[0, 1, 2, 3] = 0,  # ! Try 0 or 1
+    scale_depth: bool = True,  # scale the tile-size depending on the depth
+    is_sdxl: bool = False,  # is the model SD-XL
+):
+    # Hijacks AttnBlock from ldm and Attention from diffusers
+
+    if disable:
+        logging.info(f"Attention for {layer.__class__.__qualname__} not splitted")
+        yield
+        return
+
+    latent_tile_size = max(128, tile_size) // 8
+
+    def self_attn_forward(forward: Callable, depth: int, layer_name: str, module: nn.Module) -> Callable:
+        @wraps(forward)
+        def wrapper(*args, **kwargs):
+            x = args[0]
+
+            # VAE
+            if x.ndim == 4:
+                b, c, h, w = x.shape
+
+                nh = random_divisor(h, latent_tile_size, swap_size)
+                nw = random_divisor(w, latent_tile_size, swap_size)
+
+                if nh * nw > 1:
+                    x = rearrange(x, "b c (nh h) (nw w) -> (b nh nw) c h w", nh=nh, nw=nw) # split into nh * nw tiles
+
+                out = forward(x, *args[1:], **kwargs)
+
+                if nh * nw > 1:
+                    out = rearrange(out, "(b nh nw) c h w -> b c (nh h) (nw w)", nh=nh, nw=nw)
+
+            # U-Net
+            else:
+                hw: int = x.size(1)
+                h, w = find_hw_candidates(hw, aspect_ratio)
+                assert h * w == hw, f"Invalid aspect ratio {aspect_ratio} for input of shape {x.shape}, hw={hw}, h={h}, w={w}"
+
+                factor = 2**depth if scale_depth else 1
+                nh = random_divisor(h, latent_tile_size * factor, swap_size)
+                nw = random_divisor(w, latent_tile_size * factor, swap_size)
+
+                module._split_sizes_hypertile.append((nh, nw))  # type: ignore
+
+                if nh * nw > 1:
+                    x = rearrange(x, "b (nh h nw w) c -> (b nh nw) (h w) c", h=h // nh, w=w // nw, nh=nh, nw=nw)
+
+                out = forward(x, *args[1:], **kwargs)
+
+                if nh * nw > 1:
+                    out = rearrange(out, "(b nh nw) hw c -> b nh nw hw c", nh=nh, nw=nw)
+                    out = rearrange(out, "b nh nw (h w) c -> b (nh h nw w) c", h=h // nh, w=w // nw)
+
+            return out
+
+        return wrapper
+
+    # Handle hijacking the forward method and recovering afterwards
+    try:
+        if is_sdxl:
+            layers = DEPTH_LAYERS_XL
+        else:
+            layers = DEPTH_LAYERS
+        for depth in range(max_depth + 1):
+            for layer_name, module in layer.named_modules():
+                if any(layer_name.endswith(try_name) for try_name in layers[depth]):
+                    # print input shape for debugging
+                    logging.debug(f"HyperTile hijacking attention layer at depth {depth}: {layer_name}")
+                    # hijack
+                    module._original_forward_hypertile = module.forward
+                    module.forward = self_attn_forward(module.forward, depth, layer_name, module)
+                    module._split_sizes_hypertile = []
+        yield
+    finally:
+        for layer_name, module in layer.named_modules():
+            # remove hijack
+            if hasattr(module, "_original_forward_hypertile"):
+                if module._split_sizes_hypertile:
+                    logging.debug(f"layer {layer_name} splitted with ({module._split_sizes_hypertile})")
+                # recover
+                module.forward = module._original_forward_hypertile
+                del module._original_forward_hypertile
+                del module._split_sizes_hypertile
+
+def hypertile_context_vae(model:nn.Module, aspect_ratio:float, tile_size:int, opts):
+    """
+    Returns context manager for VAE
+    """
+    enabled = opts.hypertile_split_vae_attn
+    swap_size = opts.hypertile_swap_size_vae
+    max_depth = opts.hypertile_max_depth_vae
+    tile_size_max = opts.hypertile_max_tile_vae
+    return split_attention(
+        model,
+        aspect_ratio=aspect_ratio,
+        tile_size=min(tile_size, tile_size_max),
+        swap_size=swap_size,
+        disable=not enabled,
+        max_depth=max_depth,
+        is_sdxl=False,
+    )
+
+def hypertile_context_unet(model:nn.Module, aspect_ratio:float, tile_size:int, opts, is_sdxl:bool):
+    """
+    Returns context manager for U-Net
+    """
+    enabled = opts.hypertile_split_unet_attn
+    swap_size = opts.hypertile_swap_size_unet
+    max_depth = opts.hypertile_max_depth_unet
+    tile_size_max = opts.hypertile_max_tile_unet
+    return split_attention(
+        model,
+        aspect_ratio=aspect_ratio,
+        tile_size=min(tile_size, tile_size_max),
+        swap_size=swap_size,
+        disable=not enabled,
+        max_depth=max_depth,
+        is_sdxl=is_sdxl,
+    )