backend/headless/fcbh_extras/nodes_tomesd.py

#Taken from: https://github.com/dbolya/tomesd

import torch
from typing import Tuple, Callable
import math

def do_nothing(x: torch.Tensor, mode:str=None):
    return x


def mps_gather_workaround(input, dim, index):
    if input.shape[-1] == 1:
        return torch.gather(
            input.unsqueeze(-1),
            dim - 1 if dim < 0 else dim,
            index.unsqueeze(-1)
        ).squeeze(-1)
    else:
        return torch.gather(input, dim, index)


def bipartite_soft_matching_random2d(metric: torch.Tensor,
                                     w: int, h: int, sx: int, sy: int, r: int,
                                     no_rand: bool = False) -> Tuple[Callable, Callable]:
    """
    Partitions the tokens into src and dst and merges r tokens from src to dst.
    Dst tokens are partitioned by choosing one randomy in each (sx, sy) region.
    Args:
     - metric [B, N, C]: metric to use for similarity
     - w: image width in tokens
     - h: image height in tokens
     - sx: stride in the x dimension for dst, must divide w
     - sy: stride in the y dimension for dst, must divide h
     - r: number of tokens to remove (by merging)
     - no_rand: if true, disable randomness (use top left corner only)
    """
    B, N, _ = metric.shape

    if r <= 0 or w == 1 or h == 1:
        return do_nothing, do_nothing

    gather = mps_gather_workaround if metric.device.type == "mps" else torch.gather
    
    with torch.no_grad():
        
        hsy, wsx = h // sy, w // sx

        # For each sy by sx kernel, randomly assign one token to be dst and the rest src
        if no_rand:
            rand_idx = torch.zeros(hsy, wsx, 1, device=metric.device, dtype=torch.int64)
        else:
            rand_idx = torch.randint(sy*sx, size=(hsy, wsx, 1), device=metric.device)
        
        # The image might not divide sx and sy, so we need to work on a view of the top left if the idx buffer instead
        idx_buffer_view = torch.zeros(hsy, wsx, sy*sx, device=metric.device, dtype=torch.int64)
        idx_buffer_view.scatter_(dim=2, index=rand_idx, src=-torch.ones_like(rand_idx, dtype=rand_idx.dtype))
        idx_buffer_view = idx_buffer_view.view(hsy, wsx, sy, sx).transpose(1, 2).reshape(hsy * sy, wsx * sx)

        # Image is not divisible by sx or sy so we need to move it into a new buffer
        if (hsy * sy) < h or (wsx * sx) < w:
            idx_buffer = torch.zeros(h, w, device=metric.device, dtype=torch.int64)
            idx_buffer[:(hsy * sy), :(wsx * sx)] = idx_buffer_view
        else:
            idx_buffer = idx_buffer_view

        # We set dst tokens to be -1 and src to be 0, so an argsort gives us dst|src indices
        rand_idx = idx_buffer.reshape(1, -1, 1).argsort(dim=1)

        # We're finished with these
        del idx_buffer, idx_buffer_view

        # rand_idx is currently dst|src, so split them
        num_dst = hsy * wsx
        a_idx = rand_idx[:, num_dst:, :] # src
        b_idx = rand_idx[:, :num_dst, :] # dst

        def split(x):
            C = x.shape[-1]
            src = gather(x, dim=1, index=a_idx.expand(B, N - num_dst, C))
            dst = gather(x, dim=1, index=b_idx.expand(B, num_dst, C))
            return src, dst

        # Cosine similarity between A and B
        metric = metric / metric.norm(dim=-1, keepdim=True)
        a, b = split(metric)
        scores = a @ b.transpose(-1, -2)

        # Can't reduce more than the # tokens in src
        r = min(a.shape[1], r)

        # Find the most similar greedily
        node_max, node_idx = scores.max(dim=-1)
        edge_idx = node_max.argsort(dim=-1, descending=True)[..., None]

        unm_idx = edge_idx[..., r:, :]  # Unmerged Tokens
        src_idx = edge_idx[..., :r, :]  # Merged Tokens
        dst_idx = gather(node_idx[..., None], dim=-2, index=src_idx)

    def merge(x: torch.Tensor, mode="mean") -> torch.Tensor:
        src, dst = split(x)
        n, t1, c = src.shape
        
        unm = gather(src, dim=-2, index=unm_idx.expand(n, t1 - r, c))
        src = gather(src, dim=-2, index=src_idx.expand(n, r, c))
        dst = dst.scatter_reduce(-2, dst_idx.expand(n, r, c), src, reduce=mode)

        return torch.cat([unm, dst], dim=1)

    def unmerge(x: torch.Tensor) -> torch.Tensor:
        unm_len = unm_idx.shape[1]
        unm, dst = x[..., :unm_len, :], x[..., unm_len:, :]
        _, _, c = unm.shape

        src = gather(dst, dim=-2, index=dst_idx.expand(B, r, c))

        # Combine back to the original shape
        out = torch.zeros(B, N, c, device=x.device, dtype=x.dtype)
        out.scatter_(dim=-2, index=b_idx.expand(B, num_dst, c), src=dst)
        out.scatter_(dim=-2, index=gather(a_idx.expand(B, a_idx.shape[1], 1), dim=1, index=unm_idx).expand(B, unm_len, c), src=unm)
        out.scatter_(dim=-2, index=gather(a_idx.expand(B, a_idx.shape[1], 1), dim=1, index=src_idx).expand(B, r, c), src=src)

        return out

    return merge, unmerge


def get_functions(x, ratio, original_shape):
    b, c, original_h, original_w = original_shape
    original_tokens = original_h * original_w
    downsample = int(math.ceil(math.sqrt(original_tokens // x.shape[1])))
    stride_x = 2
    stride_y = 2
    max_downsample = 1

    if downsample <= max_downsample:
        w = int(math.ceil(original_w / downsample))
        h = int(math.ceil(original_h / downsample))
        r = int(x.shape[1] * ratio)
        no_rand = False
        m, u = bipartite_soft_matching_random2d(x, w, h, stride_x, stride_y, r, no_rand)
        return m, u

    nothing = lambda y: y
    return nothing, nothing



class TomePatchModel:
    @classmethod
    def INPUT_TYPES(s):
        return {"required": { "model": ("MODEL",),
                              "ratio": ("FLOAT", {"default": 0.3, "min": 0.0, "max": 1.0, "step": 0.01}),
                              }}
    RETURN_TYPES = ("MODEL",)
    FUNCTION = "patch"

    CATEGORY = "_for_testing"

    def patch(self, model, ratio):
        self.u = None
        def tomesd_m(q, k, v, extra_options):
            #NOTE: In the reference code get_functions takes x (input of the transformer block) as the argument instead of q
            #however from my basic testing it seems that using q instead gives better results
            m, self.u = get_functions(q, ratio, extra_options["original_shape"])
            return m(q), k, v
        def tomesd_u(n, extra_options):
            return self.u(n)

        m = model.clone()
        m.set_model_attn1_patch(tomesd_m)
        m.set_model_attn1_output_patch(tomesd_u)
        return (m, )


NODE_CLASS_MAPPINGS = {
    "TomePatchModel": TomePatchModel,
}