Mirrored_Repos/InvokeAI
1
0
Fork 0
mirror of https://github.com/invoke-ai/InvokeAI synced 2024-08-30 20:32:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update attention.py

Browse Source 
Performance improvements to generate larger images in M1 #431

Update attention.py

Added dtype=r1.dtype to softmax
This commit is contained in:
Any-Winter-4079 2022-09-12 02:52:36 +02:00 committed by Lincoln Stein
parent 49a96b90d8
commit 9cdf3aca7d
1 changed files with 91 additions and 38 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

129
ldm/modules/attention.py
View File

@ -168,6 +168,96 @@ class CrossAttention(nn.Module):
nn.Dropout(dropout)
)
if not torch.cuda.is_available():
mem_av = psutil.virtual_memory().available / (1024**3)
if mem_av > 32:
self.einsum_op = self.einsum_op_v1
elif mem_av > 12:
self.einsum_op = self.einsum_op_v2
else:
self.einsum_op = self.einsum_op_v3
del mem_av
else:
self.einsum_op = self.einsum_op_v4
# mps 64-128 GB
def einsum_op_v1(self, q, k, v, r1):
if q.shape[1] <= 4096: # for 512x512: the max q.shape[1] is 4096
s1 = einsum('b i d, b j d -> b i j', q, k) * self.scale # aggressive/faster: operation in one go
s2 = s1.softmax(dim=-1, dtype=q.dtype)
del s1
r1 = einsum('b i j, b j d -> b i d', s2, v)
del s2
else:
# q.shape[0] * q.shape[1] * slice_size >= 2**31 throws err
# needs around half of that slice_size to not generate noise
slice_size = math.floor(2**30 / (q.shape[0] * 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) * self.scale
s2 = s1.softmax(dim=-1, dtype=r1.dtype)
del s1
r1[:, i:end] = einsum('b i j, b j d -> b i d', s2, v)
del s2
return r1
# mps 16-32 GB (can be optimized)
def einsum_op_v2(self, q, k, v, r1):
slice_size = math.floor(2**30 / (q.shape[0] * q.shape[1]))
for i in range(0, q.shape[1], slice_size): # conservative/less mem: operation in steps
end = i + slice_size
s1 = einsum('b i d, b j d -> b i j', q[:, i:end], k) * self.scale
s2 = s1.softmax(dim=-1, dtype=r1.dtype)
del s1
r1[:, i:end] = einsum('b i j, b j d -> b i d', s2, v)
del s2
return r1
# mps 8 GB
def einsum_op_v3(self, q, k, v, r1):
slice_size = 1
for i in range(0, q.shape[0], slice_size): # iterate over q.shape[0]
end = min(q.shape[0], i + slice_size)
s1 = einsum('b i d, b j d -> b i j', q[i:end], k[i:end]) # adapted einsum for mem
s1 *= self.scale
s2 = s1.softmax(dim=-1, dtype=r1.dtype)
del s1
r1[i:end] = einsum('b i j, b j d -> b i d', s2, v[i:end]) # adapted einsum for mem
del s2
return r1
# cuda
def einsum_op_v4(self, q, k, v, r1):
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
gb = 1024 ** 3
tensor_size = q.shape[0] * q.shape[1] * k.shape[1] * 4
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)))
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 = min(q.shape[1], i + slice_size)
s1 = einsum('b i d, b j d -> b i j', q[:, i:end], k) * self.scale
s2 = s1.softmax(dim=-1, dtype=r1.dtype)
del s1
r1[:, i:end] = einsum('b i j, b j d -> b i d', s2, v)
del s2
return r1
def forward(self, x, context=None, mask=None):
h = self.heads
@ -180,45 +270,8 @@ class CrossAttention(nn.Module):
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)
if device_type == 'mps':
mem_free_total = psutil.virtual_memory().available
else:
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
gb = 1024 ** 3
tensor_size = q.shape[0] * q.shape[1] * k.shape[1] * 4
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)))
# 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) * self.scale
s2 = s1.softmax(dim=-1, dtype=r1.dtype)
del s1
r1[:, i:end] = einsum('b i j, b j d -> b i d', s2, v)
del s2
r1 = self.einsum_op(q, k, v, r1)
del q, k, v
r2 = rearrange(r1, '(b h) n d -> b n (h d)', h=h)