Add Birch-san's sub-quadratic attention implementation

4 files changed, 310 insertions, 34 deletions

diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py
index 690a9ec2..019a6f3f 100644
--- a/modules/sd_hijack.py
+++ b/modules/sd_hijack.py
@@ -7,8 +7,6 @@ from modules.hypernetworks import hypernetwork
 from modules.shared import cmd_opts

 from modules import sd_hijack_clip, sd_hijack_open_clip, sd_hijack_unet

 

-from modules.sd_hijack_optimizations import invokeAI_mps_available

-

 import ldm.modules.attention

 import ldm.modules.diffusionmodules.model

 import ldm.modules.diffusionmodules.openaimodel

@@ -40,17 +38,16 @@ def apply_optimizations():
         print("Applying xformers cross attention optimization.")

         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.xformers_attention_forward

         ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.xformers_attnblock_forward

+    elif cmd_opts.opt_sub_quad_attention:

+        print("Applying sub-quadratic cross attention optimization.")

+        ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.sub_quad_attention_forward

+        ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.sub_quad_attnblock_forward

     elif cmd_opts.opt_split_attention_v1:

         print("Applying v1 cross attention optimization.")

         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_v1

     elif not cmd_opts.disable_opt_split_attention and (cmd_opts.opt_split_attention_invokeai or not torch.cuda.is_available()):

-        if not invokeAI_mps_available and shared.device.type == 'mps':

-            print("The InvokeAI cross attention optimization for MPS requires the psutil package which is not installed.")

-            print("Applying v1 cross attention optimization.")

-            ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_v1

-        else:

-            print("Applying cross attention optimization (InvokeAI).")

-            ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_invokeAI

+        print("Applying cross attention optimization (InvokeAI).")

+        ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_invokeAI

     elif not cmd_opts.disable_opt_split_attention and (cmd_opts.opt_split_attention or torch.cuda.is_available()):

         print("Applying cross attention optimization (Doggettx).")

         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward

diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py
index 02c87f40..f5c153e8 100644
--- a/modules/sd_hijack_optimizations.py
+++ b/modules/sd_hijack_optimizations.py
@@ -1,7 +1,7 @@
 import math

 import sys

 import traceback

-import importlib

+import psutil

 

 import torch

 from torch import einsum

@@ -12,6 +12,8 @@ from einops import rearrange
 from modules import shared

 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:

@@ -22,6 +24,19 @@ if shared.cmd_opts.xformers or shared.cmd_opts.force_enable_xformers:
         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

@@ -76,12 +91,7 @@ def split_cross_attention_forward(self, x, context=None, mask=None):
 

     r1 = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)

 

-    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(torch.cuda.current_device())

-    mem_free_torch = mem_reserved - mem_active

-    mem_free_total = mem_free_cuda + mem_free_torch

+    mem_free_total = get_available_vram()

 

     gb = 1024 ** 3

     tensor_size = q.shape[0] * q.shape[1] * k.shape[1] * q.element_size()

@@ -118,19 +128,8 @@ def split_cross_attention_forward(self, x, context=None, mask=None):
     return self.to_out(r2)

 

 

-def check_for_psutil():

-    try:

-        spec = importlib.util.find_spec('psutil')

-        return spec is not None

-    except ModuleNotFoundError:

-        return False

-

-invokeAI_mps_available = check_for_psutil()

-

 # -- Taken from https://github.com/invoke-ai/InvokeAI and modified --

-if invokeAI_mps_available:

-    import psutil

-    mem_total_gb = psutil.virtual_memory().total // (1 << 30)

+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)

@@ -215,6 +214,70 @@ def split_cross_attention_forward_invokeAI(self, x, context=None, mask=None):
 

 # -- 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

+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_hypernetwork(shared.loaded_hypernetwork, 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)

+

+    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_bytes=shared.cmd_opts.sub_quad_chunk_threshold, use_checkpoint=self.training)

+

+    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_bytes=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

+

+    available_vram = int(get_available_vram() * 0.9) if q.device.type == 'mps' else int(get_available_vram() * 0.7)

+

+    if chunk_threshold_bytes is None:

+        chunk_threshold_bytes = available_vram

+    elif chunk_threshold_bytes == 0:

+        chunk_threshold_bytes = None

+

+    if kv_chunk_size_min is 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

+

+    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 xformers_attention_forward(self, x, context=None, mask=None):

     h = self.heads

     q_in = self.to_q(x)

@@ -252,12 +315,7 @@ def cross_attention_attnblock_forward(self, x):
 

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

 

-        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(torch.cuda.current_device())

-        mem_free_torch = mem_reserved - mem_active

-        mem_free_total = mem_free_cuda + mem_free_torch

+        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

@@ -312,3 +370,19 @@ def xformers_attnblock_forward(self, x):
         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_bytes=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

diff --git a/modules/shared.py b/modules/shared.py
index d4ddeea0..487a7792 100644
--- a/modules/shared.py
+++ b/modules/shared.py
@@ -56,6 +56,10 @@ parser.add_argument("--xformers", action='store_true', help="enable xformers for
 parser.add_argument("--force-enable-xformers", action='store_true', help="enable xformers for cross attention layers regardless of whether the checking code thinks you can run it; do not make bug reports if this fails to work")

 parser.add_argument("--deepdanbooru", action='store_true', help="does not do anything")

 parser.add_argument("--opt-split-attention", action='store_true', help="force-enables Doggettx's cross-attention layer optimization. By default, it's on for torch cuda.")

+parser.add_argument("--opt-sub-quad-attention", action='store_true', help="enable memory efficient sub-quadratic cross-attention layer optimization")

+parser.add_argument("--sub-quad-q-chunk-size", type=int, help="query chunk size for the sub-quadratic cross-attention layer optimization to use", default=1024)

+parser.add_argument("--sub-quad-kv-chunk-size", type=int, help="kv chunk size for the sub-quadratic cross-attention layer optimization to use", default=None)

+parser.add_argument("--sub-quad-chunk-threshold", type=int, help="the size threshold in bytes for the sub-quadratic cross-attention layer optimization to use chunking", default=None)

 parser.add_argument("--opt-split-attention-invokeai", action='store_true', help="force-enables InvokeAI's cross-attention layer optimization. By default, it's on when cuda is unavailable.")

 parser.add_argument("--opt-split-attention-v1", action='store_true', help="enable older version of split attention optimization that does not consume all the VRAM it can find")

 parser.add_argument("--disable-opt-split-attention", action='store_true', help="force-disables cross-attention layer optimization")

diff --git a/modules/sub_quadratic_attention.py b/modules/sub_quadratic_attention.py
new file mode 100644
index 00000000..b11dc1c7
--- /dev/null
+++ b/modules/sub_quadratic_attention.py
@@ -0,0 +1,201 @@
+# original source:
+#   https://github.com/AminRezaei0x443/memory-efficient-attention/blob/1bc0d9e6ac5f82ea43a375135c4e1d3896ee1694/memory_efficient_attention/attention_torch.py
+# license:
+#   unspecified
+# credit:
+#   Amin Rezaei (original author)
+#   Alex Birch (optimized algorithm for 3D tensors, at the expense of removing bias, masking and callbacks)
+# implementation of:
+#   Self-attention Does Not Need O(n2) Memory":
+#   https://arxiv.org/abs/2112.05682v2
+
+from functools import partial
+import torch
+from torch import Tensor
+from torch.utils.checkpoint import checkpoint
+import math
+from typing import Optional, NamedTuple, Protocol, List
+
+def dynamic_slice(
+    x: Tensor,
+    starts: List[int],
+    sizes: List[int],
+) -> Tensor:
+    slicing = [slice(start, start + size) for start, size in zip(starts, sizes)]
+    return x[slicing]
+
+class AttnChunk(NamedTuple):
+    exp_values: Tensor
+    exp_weights_sum: Tensor
+    max_score: Tensor
+
+class SummarizeChunk(Protocol):
+    @staticmethod
+    def __call__(
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+    ) -> AttnChunk: ...
+
+class ComputeQueryChunkAttn(Protocol):
+    @staticmethod
+    def __call__(
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+    ) -> Tensor: ...
+
+def _summarize_chunk(
+    query: Tensor,
+    key: Tensor,
+    value: Tensor,
+    scale: float,
+) -> AttnChunk:
+    attn_weights = torch.baddbmm(
+        torch.empty(1, 1, 1, device=query.device, dtype=query.dtype),
+        query,
+        key.transpose(1,2),
+        alpha=scale,
+        beta=0,
+    )
+    max_score, _ = torch.max(attn_weights, -1, keepdim=True)
+    max_score = max_score.detach()
+    exp_weights = torch.exp(attn_weights - max_score)
+    exp_values = torch.bmm(exp_weights, value)
+    max_score = max_score.squeeze(-1)
+    return AttnChunk(exp_values, exp_weights.sum(dim=-1), max_score)
+
+def _query_chunk_attention(
+    query: Tensor,
+    key: Tensor,
+    value: Tensor,
+    summarize_chunk: SummarizeChunk,
+    kv_chunk_size: int,
+) -> Tensor:
+    batch_x_heads, k_tokens, k_channels_per_head = key.shape
+    _, _, v_channels_per_head = value.shape
+
+    def chunk_scanner(chunk_idx: int) -> AttnChunk:
+        key_chunk = dynamic_slice(
+            key,
+            (0, chunk_idx, 0),
+            (batch_x_heads, kv_chunk_size, k_channels_per_head)
+        )
+        value_chunk = dynamic_slice(
+            value,
+            (0, chunk_idx, 0),
+            (batch_x_heads, kv_chunk_size, v_channels_per_head)
+        )
+        return summarize_chunk(query, key_chunk, value_chunk)
+
+    chunks: List[AttnChunk] = [
+        chunk_scanner(chunk) for chunk in torch.arange(0, k_tokens, kv_chunk_size)
+    ]
+    acc_chunk = AttnChunk(*map(torch.stack, zip(*chunks)))
+    chunk_values, chunk_weights, chunk_max = acc_chunk
+
+    global_max, _ = torch.max(chunk_max, 0, keepdim=True)
+    max_diffs = torch.exp(chunk_max - global_max)
+    chunk_values *= torch.unsqueeze(max_diffs, -1)
+    chunk_weights *= max_diffs
+
+    all_values = chunk_values.sum(dim=0)
+    all_weights = torch.unsqueeze(chunk_weights, -1).sum(dim=0)
+    return all_values / all_weights
+
+# TODO: refactor CrossAttention#get_attention_scores to share code with this
+def _get_attention_scores_no_kv_chunking(
+    query: Tensor,
+    key: Tensor,
+    value: Tensor,
+    scale: float,
+) -> Tensor:
+    attn_scores = torch.baddbmm(
+        torch.empty(1, 1, 1, device=query.device, dtype=query.dtype),
+        query,
+        key.transpose(1,2),
+        alpha=scale,
+        beta=0,
+    )
+    attn_probs = attn_scores.softmax(dim=-1)
+    del attn_scores
+    hidden_states_slice = torch.bmm(attn_probs, value)
+    return hidden_states_slice
+
+class ScannedChunk(NamedTuple):
+    chunk_idx: int
+    attn_chunk: AttnChunk
+
+def efficient_dot_product_attention(
+    query: Tensor,
+    key: Tensor,
+    value: Tensor,
+    query_chunk_size=1024,
+    kv_chunk_size: Optional[int] = None,
+    kv_chunk_size_min: Optional[int] = None,
+    use_checkpoint=True,
+):
+    """Computes efficient dot-product attention given query, key, and value.
+      This is efficient version of attention presented in
+      https://arxiv.org/abs/2112.05682v2 which comes with O(sqrt(n)) memory requirements.
+      Args:
+        query: queries for calculating attention with shape of
+          `[batch * num_heads, tokens, channels_per_head]`.
+        key: keys for calculating attention with shape of
+          `[batch * num_heads, tokens, channels_per_head]`.
+        value: values to be used in attention with shape of
+          `[batch * num_heads, tokens, channels_per_head]`.
+        query_chunk_size: int: query chunks size
+        kv_chunk_size: Optional[int]: key/value chunks size. if None: defaults to sqrt(key_tokens)
+        kv_chunk_size_min: Optional[int]: key/value minimum chunk size. only considered when kv_chunk_size is None. changes `sqrt(key_tokens)` into `max(sqrt(key_tokens), kv_chunk_size_min)`, to ensure our chunk sizes don't get too small (smaller chunks = more chunks = less concurrent work done).
+        use_checkpoint: bool: whether to use checkpointing (recommended True for training, False for inference)
+      Returns:
+        Output of shape `[batch * num_heads, query_tokens, channels_per_head]`.
+      """
+    batch_x_heads, q_tokens, q_channels_per_head = query.shape
+    _, k_tokens, _ = key.shape
+    scale = q_channels_per_head ** -0.5
+
+    kv_chunk_size = min(kv_chunk_size or int(math.sqrt(k_tokens)), k_tokens)
+    if kv_chunk_size_min is not None:
+        kv_chunk_size = max(kv_chunk_size, kv_chunk_size_min)
+
+    def get_query_chunk(chunk_idx: int) -> Tensor:
+        return dynamic_slice(
+            query,
+            (0, chunk_idx, 0),
+            (batch_x_heads, min(query_chunk_size, q_tokens), q_channels_per_head)
+        )
+    
+    summarize_chunk: SummarizeChunk = partial(_summarize_chunk, scale=scale)
+    summarize_chunk: SummarizeChunk = partial(checkpoint, summarize_chunk) if use_checkpoint else summarize_chunk
+    compute_query_chunk_attn: ComputeQueryChunkAttn = partial(
+        _get_attention_scores_no_kv_chunking,
+        scale=scale
+    ) if k_tokens <= kv_chunk_size else (
+        # fast-path for when there's just 1 key-value chunk per query chunk (this is just sliced attention btw)
+        partial(
+            _query_chunk_attention,
+            kv_chunk_size=kv_chunk_size,
+            summarize_chunk=summarize_chunk,
+        )
+    )
+
+    if q_tokens <= query_chunk_size:
+        # fast-path for when there's just 1 query chunk
+        return compute_query_chunk_attn(
+            query=query,
+            key=key,
+            value=value,
+        )
+    
+    # TODO: maybe we should use torch.empty_like(query) to allocate storage in-advance,
+    # and pass slices to be mutated, instead of torch.cat()ing the returned slices
+    res = torch.cat([
+        compute_query_chunk_attn(
+            query=get_query_chunk(i * query_chunk_size),
+            key=key,
+            value=value,
+        ) for i in range(math.ceil(q_tokens / query_chunk_size))
+    ], dim=1)
+    return res