modules/sd_hijack_optimizations.py

import math
import sys
import traceback
import psutil

import torch
from torch import einsum

from ldm.util import default
from einops import rearrange

from modules import shared, errors, devices
from modules.hypernetworks import hypernetwork

from .sub_quadratic_attention import efficient_dot_product_attention


if shared.cmd_opts.xformers or shared.cmd_opts.force_enable_xformers:
    try:
        import xformers.ops
        shared.xformers_available = True
    except Exception:
        print("Cannot import xformers", file=sys.stderr)
        print(traceback.format_exc(), file=sys.stderr)


def get_available_vram():
    if shared.device.type == 'cuda':
        stats = torch.cuda.memory_stats(shared.device)
        mem_active = stats['active_bytes.all.current']
        mem_reserved = stats['reserved_bytes.all.current']
        mem_free_cuda, _ = torch.cuda.mem_get_info(torch.cuda.current_device())
        mem_free_torch = mem_reserved - mem_active
        mem_free_total = mem_free_cuda + mem_free_torch
        return mem_free_total
    else:
        return psutil.virtual_memory().available


# see https://github.com/basujindal/stable-diffusion/pull/117 for discussion
def split_cross_attention_forward_v1(self, x, context=None, mask=None):
    h = self.heads

    q_in = self.to_q(x)
    context = default(context, x)

    context_k, context_v = hypernetwork.apply_hypernetworks(shared.loaded_hypernetworks, context)
    k_in = self.to_k(context_k)
    v_in = self.to_v(context_v)
    del context, context_k, context_v, x

    q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q_in, k_in, v_in))
    del q_in, k_in, v_in

    dtype = q.dtype
    if shared.opts.upcast_attn:
        q, k, v = q.float(), k.float(), v.float()

    with devices.without_autocast(disable=not shared.opts.upcast_attn):
        r1 = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)
        for i in range(0, q.shape[0], 2):
            end = i + 2
            s1 = einsum('b i d, b j d -> b i j', q[i:end], k[i:end])
            s1 *= self.scale
    
            s2 = s1.softmax(dim=-1)
            del s1
    
            r1[i:end] = einsum('b i j, b j d -> b i d', s2, v[i:end])
            del s2
        del q, k, v

    r1 = r1.to(dtype)

    r2 = rearrange(r1, '(b h) n d -> b n (h d)', h=h)
    del r1

    return self.to_out(r2)


# taken from https://github.com/Doggettx/stable-diffusion and modified
def split_cross_attention_forward(self, x, context=None, mask=None):
    h = self.heads

    q_in = self.to_q(x)
    context = default(context, x)

    context_k, context_v = hypernetwork.apply_hypernetworks(shared.loaded_hypernetworks, context)
    k_in = self.to_k(context_k)
    v_in = self.to_v(context_v)

    dtype = q_in.dtype
    if shared.opts.upcast_attn:
        q_in, k_in, v_in = q_in.float(), k_in.float(), v_in if v_in.device.type == 'mps' else v_in.float()

    with devices.without_autocast(disable=not shared.opts.upcast_attn):
        k_in = k_in * self.scale
    
        del context, x
    
        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q_in, k_in, v_in))
        del q_in, k_in, v_in
    
        r1 = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)
    
        mem_free_total = get_available_vram()
    
        gb = 1024 ** 3
        tensor_size = q.shape[0] * q.shape[1] * k.shape[1] * q.element_size()
        modifier = 3 if q.element_size() == 2 else 2.5
        mem_required = tensor_size * modifier
        steps = 1
    
        if mem_required > mem_free_total:
            steps = 2 ** (math.ceil(math.log(mem_required / mem_free_total, 2)))
            # print(f"Expected tensor size:{tensor_size/gb:0.1f}GB, cuda free:{mem_free_cuda/gb:0.1f}GB "
            #       f"torch free:{mem_free_torch/gb:0.1f} total:{mem_free_total/gb:0.1f} steps:{steps}")
    
        if steps > 64:
            max_res = math.floor(math.sqrt(math.sqrt(mem_free_total / 2.5)) / 8) * 64
            raise RuntimeError(f'Not enough memory, use lower resolution (max approx. {max_res}x{max_res}). '
                               f'Need: {mem_required / 64 / gb:0.1f}GB free, Have:{mem_free_total / gb:0.1f}GB free')
    
        slice_size = q.shape[1] // steps if (q.shape[1] % steps) == 0 else q.shape[1]
        for i in range(0, q.shape[1], slice_size):
            end = i + slice_size
            s1 = einsum('b i d, b j d -> b i j', q[:, i:end], k)
    
            s2 = s1.softmax(dim=-1, dtype=q.dtype)
            del s1
    
            r1[:, i:end] = einsum('b i j, b j d -> b i d', s2, v)
            del s2
    
        del q, k, v

    r1 = r1.to(dtype)

    r2 = rearrange(r1, '(b h) n d -> b n (h d)', h=h)
    del r1

    return self.to_out(r2)


# -- Taken from https://github.com/invoke-ai/InvokeAI and modified --
mem_total_gb = psutil.virtual_memory().total // (1 << 30)

def einsum_op_compvis(q, k, v):
    s = einsum('b i d, b j d -> b i j', q, k)
    s = s.softmax(dim=-1, dtype=s.dtype)
    return einsum('b i j, b j d -> b i d', s, v)

def einsum_op_slice_0(q, k, v, slice_size):
    r = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)
    for i in range(0, q.shape[0], slice_size):
        end = i + slice_size
        r[i:end] = einsum_op_compvis(q[i:end], k[i:end], v[i:end])
    return r

def einsum_op_slice_1(q, k, v, slice_size):
    r = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)
    for i in range(0, q.shape[1], slice_size):
        end = i + slice_size
        r[:, i:end] = einsum_op_compvis(q[:, i:end], k, v)
    return r

def einsum_op_mps_v1(q, k, v):
    if q.shape[0] * q.shape[1] <= 2**16: # (512x512) max q.shape[1]: 4096
        return einsum_op_compvis(q, k, v)
    else:
        slice_size = math.floor(2**30 / (q.shape[0] * q.shape[1]))
        if slice_size % 4096 == 0:
            slice_size -= 1
        return einsum_op_slice_1(q, k, v, slice_size)

def einsum_op_mps_v2(q, k, v):
    if mem_total_gb > 8 and q.shape[0] * q.shape[1] <= 2**16:
        return einsum_op_compvis(q, k, v)
    else:
        return einsum_op_slice_0(q, k, v, 1)

def einsum_op_tensor_mem(q, k, v, max_tensor_mb):
    size_mb = q.shape[0] * q.shape[1] * k.shape[1] * q.element_size() // (1 << 20)
    if size_mb <= max_tensor_mb:
        return einsum_op_compvis(q, k, v)
    div = 1 << int((size_mb - 1) / max_tensor_mb).bit_length()
    if div <= q.shape[0]:
        return einsum_op_slice_0(q, k, v, q.shape[0] // div)
    return einsum_op_slice_1(q, k, v, max(q.shape[1] // div, 1))

def einsum_op_cuda(q, k, v):
    stats = torch.cuda.memory_stats(q.device)
    mem_active = stats['active_bytes.all.current']
    mem_reserved = stats['reserved_bytes.all.current']
    mem_free_cuda, _ = torch.cuda.mem_get_info(q.device)
    mem_free_torch = mem_reserved - mem_active
    mem_free_total = mem_free_cuda + mem_free_torch
    # Divide factor of safety as there's copying and fragmentation
    return einsum_op_tensor_mem(q, k, v, mem_free_total / 3.3 / (1 << 20))

def einsum_op(q, k, v):
    if q.device.type == 'cuda':
        return einsum_op_cuda(q, k, v)

    if q.device.type == 'mps':
        if mem_total_gb >= 32 and q.shape[0] % 32 != 0 and q.shape[0] * q.shape[1] < 2**18:
            return einsum_op_mps_v1(q, k, v)
        return einsum_op_mps_v2(q, k, v)

    # Smaller slices are faster due to L2/L3/SLC caches.
    # Tested on i7 with 8MB L3 cache.
    return einsum_op_tensor_mem(q, k, v, 32)

def split_cross_attention_forward_invokeAI(self, x, context=None, mask=None):
    h = self.heads

    q = self.to_q(x)
    context = default(context, x)

    context_k, context_v = hypernetwork.apply_hypernetworks(shared.loaded_hypernetworks, context)
    k = self.to_k(context_k)
    v = self.to_v(context_v)
    del context, context_k, context_v, x

    dtype = q.dtype
    if shared.opts.upcast_attn:
        q, k, v = q.float(), k.float(), v if v.device.type == 'mps' else v.float()

    with devices.without_autocast(disable=not shared.opts.upcast_attn):
        k = k * self.scale
    
        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
        r = einsum_op(q, k, v)
    r = r.to(dtype)
    return self.to_out(rearrange(r, '(b h) n d -> b n (h d)', h=h))

# -- End of code from https://github.com/invoke-ai/InvokeAI --


# Based on Birch-san's modified implementation of sub-quadratic attention from https://github.com/Birch-san/diffusers/pull/1
# The sub_quad_attention_forward function is under the MIT License listed under Memory Efficient Attention in the Licenses section of the web UI interface
def sub_quad_attention_forward(self, x, context=None, mask=None):
    assert mask is None, "attention-mask not currently implemented for SubQuadraticCrossAttnProcessor."

    h = self.heads

    q = self.to_q(x)
    context = default(context, x)

    context_k, context_v = hypernetwork.apply_hypernetworks(shared.loaded_hypernetworks, context)
    k = self.to_k(context_k)
    v = self.to_v(context_v)
    del context, context_k, context_v, x

    q = q.unflatten(-1, (h, -1)).transpose(1,2).flatten(end_dim=1)
    k = k.unflatten(-1, (h, -1)).transpose(1,2).flatten(end_dim=1)
    v = v.unflatten(-1, (h, -1)).transpose(1,2).flatten(end_dim=1)

    dtype = q.dtype
    if shared.opts.upcast_attn:
        q, k = q.float(), k.float()

    x = sub_quad_attention(q, k, v, q_chunk_size=shared.cmd_opts.sub_quad_q_chunk_size, kv_chunk_size=shared.cmd_opts.sub_quad_kv_chunk_size, chunk_threshold=shared.cmd_opts.sub_quad_chunk_threshold, use_checkpoint=self.training)

    x = x.to(dtype)

    x = x.unflatten(0, (-1, h)).transpose(1,2).flatten(start_dim=2)

    out_proj, dropout = self.to_out
    x = out_proj(x)
    x = dropout(x)

    return x

def sub_quad_attention(q, k, v, q_chunk_size=1024, kv_chunk_size=None, kv_chunk_size_min=None, chunk_threshold=None, use_checkpoint=True):
    bytes_per_token = torch.finfo(q.dtype).bits//8
    batch_x_heads, q_tokens, _ = q.shape
    _, k_tokens, _ = k.shape
    qk_matmul_size_bytes = batch_x_heads * bytes_per_token * q_tokens * k_tokens

    if chunk_threshold is None:
        chunk_threshold_bytes = int(get_available_vram() * 0.9) if q.device.type == 'mps' else int(get_available_vram() * 0.7)
    elif chunk_threshold == 0:
        chunk_threshold_bytes = None
    else:
        chunk_threshold_bytes = int(0.01 * chunk_threshold * get_available_vram())

    if kv_chunk_size_min is None and chunk_threshold_bytes is not None:
        kv_chunk_size_min = chunk_threshold_bytes // (batch_x_heads * bytes_per_token * (k.shape[2] + v.shape[2]))
    elif kv_chunk_size_min == 0:
        kv_chunk_size_min = None

    if chunk_threshold_bytes is not None and qk_matmul_size_bytes <= chunk_threshold_bytes:
        # the big matmul fits into our memory limit; do everything in 1 chunk,
        # i.e. send it down the unchunked fast-path
        query_chunk_size = q_tokens
        kv_chunk_size = k_tokens

    with devices.without_autocast(disable=q.dtype == v.dtype):
        return efficient_dot_product_attention(
            q,
            k,
            v,
            query_chunk_size=q_chunk_size,
            kv_chunk_size=kv_chunk_size,
            kv_chunk_size_min = kv_chunk_size_min,
            use_checkpoint=use_checkpoint,
        )


def get_xformers_flash_attention_op(q, k, v):
    if not shared.cmd_opts.xformers_flash_attention:
        return None

    try:
        flash_attention_op = xformers.ops.MemoryEfficientAttentionFlashAttentionOp
        fw, bw = flash_attention_op
        if fw.supports(xformers.ops.fmha.Inputs(query=q, key=k, value=v, attn_bias=None)):
            return flash_attention_op
    except Exception as e:
        errors.display_once(e, "enabling flash attention")

    return None


def xformers_attention_forward(self, x, context=None, mask=None):
    h = self.heads
    q_in = self.to_q(x)
    context = default(context, x)

    context_k, context_v = hypernetwork.apply_hypernetworks(shared.loaded_hypernetworks, context)
    k_in = self.to_k(context_k)
    v_in = self.to_v(context_v)

    q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b n h d', h=h), (q_in, k_in, v_in))
    del q_in, k_in, v_in

    dtype = q.dtype
    if shared.opts.upcast_attn:
        q, k = q.float(), k.float()

    out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=get_xformers_flash_attention_op(q, k, v))

    out = out.to(dtype)

    out = rearrange(out, 'b n h d -> b n (h d)', h=h)
    return self.to_out(out)

def cross_attention_attnblock_forward(self, x):
        h_ = x
        h_ = self.norm(h_)
        q1 = self.q(h_)
        k1 = self.k(h_)
        v = self.v(h_)

        # compute attention
        b, c, h, w = q1.shape

        q2 = q1.reshape(b, c, h*w)
        del q1

        q = q2.permute(0, 2, 1)   # b,hw,c
        del q2

        k = k1.reshape(b, c, h*w) # b,c,hw
        del k1

        h_ = torch.zeros_like(k, device=q.device)

        mem_free_total = get_available_vram()

        tensor_size = q.shape[0] * q.shape[1] * k.shape[2] * q.element_size()
        mem_required = tensor_size * 2.5
        steps = 1

        if mem_required > mem_free_total:
            steps = 2**(math.ceil(math.log(mem_required / mem_free_total, 2)))

        slice_size = q.shape[1] // steps if (q.shape[1] % steps) == 0 else q.shape[1]
        for i in range(0, q.shape[1], slice_size):
            end = i + slice_size

            w1 = torch.bmm(q[:, i:end], k)     # b,hw,hw    w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
            w2 = w1 * (int(c)**(-0.5))
            del w1
            w3 = torch.nn.functional.softmax(w2, dim=2, dtype=q.dtype)
            del w2

            # attend to values
            v1 = v.reshape(b, c, h*w)
            w4 = w3.permute(0, 2, 1)   # b,hw,hw (first hw of k, second of q)
            del w3

            h_[:, :, i:end] = torch.bmm(v1, w4)     # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
            del v1, w4

        h2 = h_.reshape(b, c, h, w)
        del h_

        h3 = self.proj_out(h2)
        del h2

        h3 += x

        return h3
    
def xformers_attnblock_forward(self, x):
    try:
        h_ = x
        h_ = self.norm(h_)
        q = self.q(h_)
        k = self.k(h_)
        v = self.v(h_)
        b, c, h, w = q.shape
        q, k, v = map(lambda t: rearrange(t, 'b c h w -> b (h w) c'), (q, k, v))
        dtype = q.dtype
        if shared.opts.upcast_attn:
            q, k = q.float(), k.float()
        q = q.contiguous()
        k = k.contiguous()
        v = v.contiguous()
        out = xformers.ops.memory_efficient_attention(q, k, v, op=get_xformers_flash_attention_op(q, k, v))
        out = out.to(dtype)
        out = rearrange(out, 'b (h w) c -> b c h w', h=h)
        out = self.proj_out(out)
        return x + out
    except NotImplementedError:
        return cross_attention_attnblock_forward(self, x)

def sub_quad_attnblock_forward(self, x):
    h_ = x
    h_ = self.norm(h_)
    q = self.q(h_)
    k = self.k(h_)
    v = self.v(h_)
    b, c, h, w = q.shape
    q, k, v = map(lambda t: rearrange(t, 'b c h w -> b (h w) c'), (q, k, v))
    q = q.contiguous()
    k = k.contiguous()
    v = v.contiguous()
    out = sub_quad_attention(q, k, v, q_chunk_size=shared.cmd_opts.sub_quad_q_chunk_size, kv_chunk_size=shared.cmd_opts.sub_quad_kv_chunk_size, chunk_threshold=shared.cmd_opts.sub_quad_chunk_threshold, use_checkpoint=self.training)
    out = rearrange(out, 'b (h w) c -> b c h w', h=h)
    out = self.proj_out(out)
    return x + out