From 762265eab58cdb8f2d6398769bab43d8b8db0075 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Wed, 10 May 2023 07:52:45 +0300 Subject: autofixes from ruff --- modules/sd_hijack_optimizations.py | 1 - 1 file changed, 1 deletion(-) (limited to 'modules/sd_hijack_optimizations.py') diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py index f10865cd..b623d53d 100644 --- a/modules/sd_hijack_optimizations.py +++ b/modules/sd_hijack_optimizations.py @@ -296,7 +296,6 @@ def sub_quad_attention(q, k, v, q_chunk_size=1024, kv_chunk_size=None, kv_chunk_ 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): -- cgit v1.2.1 From 028d3f6425d85f122027c127fba8bcbf4f66ee75 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Wed, 10 May 2023 11:05:02 +0300 Subject: ruff auto fixes --- modules/sd_hijack_optimizations.py | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) (limited to 'modules/sd_hijack_optimizations.py') diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py index b623d53d..a174bbe1 100644 --- a/modules/sd_hijack_optimizations.py +++ b/modules/sd_hijack_optimizations.py @@ -49,7 +49,7 @@ def split_cross_attention_forward_v1(self, x, context=None, mask=None): 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)) + q, k, v = (rearrange(t, 'b n (h d) -> (b h) n d', h=h) for t in (q_in, k_in, v_in)) del q_in, k_in, v_in dtype = q.dtype @@ -98,7 +98,7 @@ def split_cross_attention_forward(self, x, context=None, mask=None): 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)) + q, k, v = (rearrange(t, 'b n (h d) -> (b h) n d', h=h) for t in (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) @@ -229,7 +229,7 @@ def split_cross_attention_forward_invokeAI(self, x, context=None, mask=None): 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)) + q, k, v = (rearrange(t, 'b n (h d) -> (b h) n d', h=h) for t in (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)) @@ -334,7 +334,7 @@ def xformers_attention_forward(self, x, context=None, mask=None): 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)) + q, k, v = (rearrange(t, 'b n (h d) -> b n h d', h=h) for t in (q_in, k_in, v_in)) del q_in, k_in, v_in dtype = q.dtype @@ -460,7 +460,7 @@ def xformers_attnblock_forward(self, x): 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, k, v = (rearrange(t, 'b c h w -> b (h w) c') for t in (q, k, v)) dtype = q.dtype if shared.opts.upcast_attn: q, k = q.float(), k.float() @@ -482,7 +482,7 @@ def sdp_attnblock_forward(self, x): 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, k, v = (rearrange(t, 'b c h w -> b (h w) c') for t in (q, k, v)) dtype = q.dtype if shared.opts.upcast_attn: q, k = q.float(), k.float() @@ -506,7 +506,7 @@ def sub_quad_attnblock_forward(self, x): 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, k, v = (rearrange(t, 'b c h w -> b (h w) c') for t in (q, k, v)) q = q.contiguous() k = k.contiguous() v = v.contiguous() -- cgit v1.2.1 From 49a55b410b66b7dd9be9335d8a2e3a71e4f8b15c Mon Sep 17 00:00:00 2001 From: Aarni Koskela Date: Thu, 11 May 2023 18:28:15 +0300 Subject: Autofix Ruff W (not W605) (mostly whitespace) --- modules/sd_hijack_optimizations.py | 32 ++++++++++++++++---------------- 1 file changed, 16 insertions(+), 16 deletions(-) (limited to 'modules/sd_hijack_optimizations.py') diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py index a174bbe1..f00fe55c 100644 --- a/modules/sd_hijack_optimizations.py +++ b/modules/sd_hijack_optimizations.py @@ -62,10 +62,10 @@ def split_cross_attention_forward_v1(self, x, context=None, mask=None): 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 @@ -95,43 +95,43 @@ def split_cross_attention_forward(self, x, context=None, mask=None): with devices.without_autocast(disable=not shared.opts.upcast_attn): k_in = k_in * self.scale - + del context, x - + q, k, v = (rearrange(t, 'b n (h d) -> (b h) n d', h=h) for t in (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) @@ -228,7 +228,7 @@ def split_cross_attention_forward_invokeAI(self, x, context=None, mask=None): with devices.without_autocast(disable=not shared.opts.upcast_attn): k = k * self.scale - + q, k, v = (rearrange(t, 'b n (h d) -> (b h) n d', h=h) for t in (q, k, v)) r = einsum_op(q, k, v) r = r.to(dtype) @@ -369,7 +369,7 @@ def scaled_dot_product_attention_forward(self, x, context=None, mask=None): q = q_in.view(batch_size, -1, h, head_dim).transpose(1, 2) k = k_in.view(batch_size, -1, h, head_dim).transpose(1, 2) v = v_in.view(batch_size, -1, h, head_dim).transpose(1, 2) - + del q_in, k_in, v_in dtype = q.dtype @@ -451,7 +451,7 @@ def cross_attention_attnblock_forward(self, x): h3 += x return h3 - + def xformers_attnblock_forward(self, x): try: h_ = x -- cgit v1.2.1 From 2582a0fd3b3e91c5fba9e5e561cbdf5fee835063 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Thu, 18 May 2023 22:48:28 +0300 Subject: make it possible for scripts to add cross attention optimizations add UI selection for cross attention optimization --- modules/sd_hijack_optimizations.py | 135 ++++++++++++++++++++++++++++++++++++- 1 file changed, 132 insertions(+), 3 deletions(-) (limited to 'modules/sd_hijack_optimizations.py') diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py index f00fe55c..1c5b709b 100644 --- a/modules/sd_hijack_optimizations.py +++ b/modules/sd_hijack_optimizations.py @@ -9,10 +9,139 @@ from torch import einsum from ldm.util import default from einops import rearrange -from modules import shared, errors, devices +from modules import shared, errors, devices, sub_quadratic_attention, script_callbacks from modules.hypernetworks import hypernetwork -from .sub_quadratic_attention import efficient_dot_product_attention +import ldm.modules.attention +import ldm.modules.diffusionmodules.model + +diffusionmodules_model_AttnBlock_forward = ldm.modules.diffusionmodules.model.AttnBlock.forward + + +class SdOptimization: + def __init__(self, name, label=None, cmd_opt=None): + self.name = name + self.label = label + self.cmd_opt = cmd_opt + + def title(self): + if self.label is None: + return self.name + + return f"{self.name} - {self.label}" + + def is_available(self): + return True + + def priority(self): + return 0 + + def apply(self): + pass + + def undo(self): + ldm.modules.attention.CrossAttention.forward = hypernetwork.attention_CrossAttention_forward + ldm.modules.diffusionmodules.model.AttnBlock.forward = diffusionmodules_model_AttnBlock_forward + + +class SdOptimizationXformers(SdOptimization): + def __init__(self): + super().__init__("xformers", cmd_opt="xformers") + + def is_available(self): + return shared.cmd_opts.force_enable_xformers or (shared.xformers_available and torch.version.cuda and (6, 0) <= torch.cuda.get_device_capability(shared.device) <= (9, 0)) + + def priority(self): + return 100 + + def apply(self): + ldm.modules.attention.CrossAttention.forward = xformers_attention_forward + ldm.modules.diffusionmodules.model.AttnBlock.forward = xformers_attnblock_forward + + +class SdOptimizationSdpNoMem(SdOptimization): + def __init__(self, name="sdp-no-mem", label="scaled dot product without memory efficient attention", cmd_opt="opt_sdp_no_mem_attention"): + super().__init__(name, label, cmd_opt) + + def is_available(self): + return hasattr(torch.nn.functional, "scaled_dot_product_attention") and callable(torch.nn.functional.scaled_dot_product_attention) + + def priority(self): + return 90 + + def apply(self): + ldm.modules.attention.CrossAttention.forward = scaled_dot_product_no_mem_attention_forward + ldm.modules.diffusionmodules.model.AttnBlock.forward = sdp_no_mem_attnblock_forward + + +class SdOptimizationSdp(SdOptimizationSdpNoMem): + def __init__(self): + super().__init__("sdp", "scaled dot product", cmd_opt="opt_sdp_attention") + + def priority(self): + return 80 + + def apply(self): + ldm.modules.attention.CrossAttention.forward = scaled_dot_product_attention_forward + ldm.modules.diffusionmodules.model.AttnBlock.forward = sdp_attnblock_forward + + +class SdOptimizationSubQuad(SdOptimization): + def __init__(self): + super().__init__("sub-quadratic", cmd_opt="opt_sub_quad_attention") + + def priority(self): + return 10 + + def apply(self): + ldm.modules.attention.CrossAttention.forward = sub_quad_attention_forward + ldm.modules.diffusionmodules.model.AttnBlock.forward = sub_quad_attnblock_forward + + +class SdOptimizationV1(SdOptimization): + def __init__(self): + super().__init__("V1", "original v1", cmd_opt="opt_split_attention_v1") + + def priority(self): + return 10 + + def apply(self): + ldm.modules.attention.CrossAttention.forward = split_cross_attention_forward_v1 + + +class SdOptimizationInvokeAI(SdOptimization): + def __init__(self): + super().__init__("InvokeAI", cmd_opt="opt_split_attention_invokeai") + + def priority(self): + return 1000 if not torch.cuda.is_available() else 10 + + def apply(self): + ldm.modules.attention.CrossAttention.forward = split_cross_attention_forward_invokeAI + + +class SdOptimizationDoggettx(SdOptimization): + def __init__(self): + super().__init__("Doggettx", cmd_opt="opt_split_attention") + + def priority(self): + return 20 + + def apply(self): + ldm.modules.attention.CrossAttention.forward = split_cross_attention_forward + ldm.modules.diffusionmodules.model.AttnBlock.forward = cross_attention_attnblock_forward + + +def list_optimizers(res): + res.extend([ + SdOptimizationXformers(), + SdOptimizationSdpNoMem(), + SdOptimizationSdp(), + SdOptimizationSubQuad(), + SdOptimizationV1(), + SdOptimizationInvokeAI(), + SdOptimizationDoggettx(), + ]) if shared.cmd_opts.xformers or shared.cmd_opts.force_enable_xformers: @@ -299,7 +428,7 @@ def sub_quad_attention(q, k, v, q_chunk_size=1024, kv_chunk_size=None, kv_chunk_ kv_chunk_size = k_tokens with devices.without_autocast(disable=q.dtype == v.dtype): - return efficient_dot_product_attention( + return sub_quadratic_attention.efficient_dot_product_attention( q, k, v, -- cgit v1.2.1 From 8a3d232839930376898634f65bd6c16f3a41e5b4 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 19 May 2023 00:03:27 +0300 Subject: fix linter issues --- modules/sd_hijack_optimizations.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/sd_hijack_optimizations.py') diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py index 1c5b709b..db1e4367 100644 --- a/modules/sd_hijack_optimizations.py +++ b/modules/sd_hijack_optimizations.py @@ -9,7 +9,7 @@ from torch import einsum from ldm.util import default from einops import rearrange -from modules import shared, errors, devices, sub_quadratic_attention, script_callbacks +from modules import shared, errors, devices, sub_quadratic_attention from modules.hypernetworks import hypernetwork import ldm.modules.attention -- cgit v1.2.1 From 1e5afd4fa9774314d649bddb3d18a9a75871902b Mon Sep 17 00:00:00 2001 From: AUTOMATIC1111 <16777216c@gmail.com> Date: Fri, 19 May 2023 09:17:36 +0300 Subject: Apply suggestions from code review Co-authored-by: Aarni Koskela --- modules/sd_hijack_optimizations.py | 66 ++++++++++++++++---------------------- 1 file changed, 28 insertions(+), 38 deletions(-) (limited to 'modules/sd_hijack_optimizations.py') diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py index db1e4367..0eb4c525 100644 --- a/modules/sd_hijack_optimizations.py +++ b/modules/sd_hijack_optimizations.py @@ -19,10 +19,10 @@ diffusionmodules_model_AttnBlock_forward = ldm.modules.diffusionmodules.model.At class SdOptimization: - def __init__(self, name, label=None, cmd_opt=None): - self.name = name - self.label = label - self.cmd_opt = cmd_opt + name: str = None + label: str | None = None + cmd_opt: str | None = None + priority: int = 0 def title(self): if self.label is None: @@ -33,9 +33,6 @@ class SdOptimization: def is_available(self): return True - def priority(self): - return 0 - def apply(self): pass @@ -45,41 +42,37 @@ class SdOptimization: class SdOptimizationXformers(SdOptimization): - def __init__(self): - super().__init__("xformers", cmd_opt="xformers") + name = "xformers" + cmd_opt = "xformers" + priority = 100 def is_available(self): return shared.cmd_opts.force_enable_xformers or (shared.xformers_available and torch.version.cuda and (6, 0) <= torch.cuda.get_device_capability(shared.device) <= (9, 0)) - def priority(self): - return 100 - def apply(self): ldm.modules.attention.CrossAttention.forward = xformers_attention_forward ldm.modules.diffusionmodules.model.AttnBlock.forward = xformers_attnblock_forward class SdOptimizationSdpNoMem(SdOptimization): - def __init__(self, name="sdp-no-mem", label="scaled dot product without memory efficient attention", cmd_opt="opt_sdp_no_mem_attention"): - super().__init__(name, label, cmd_opt) + name = "sdp-no-mem" + label = "scaled dot product without memory efficient attention" + cmd_opt = "opt_sdp_no_mem_attention" + priority = 90 def is_available(self): return hasattr(torch.nn.functional, "scaled_dot_product_attention") and callable(torch.nn.functional.scaled_dot_product_attention) - def priority(self): - return 90 - def apply(self): ldm.modules.attention.CrossAttention.forward = scaled_dot_product_no_mem_attention_forward ldm.modules.diffusionmodules.model.AttnBlock.forward = sdp_no_mem_attnblock_forward class SdOptimizationSdp(SdOptimizationSdpNoMem): - def __init__(self): - super().__init__("sdp", "scaled dot product", cmd_opt="opt_sdp_attention") - - def priority(self): - return 80 + name = "sdp" + label = "scaled dot product" + cmd_opt = "opt_sdp_attention" + priority = 80 def apply(self): ldm.modules.attention.CrossAttention.forward = scaled_dot_product_attention_forward @@ -87,11 +80,9 @@ class SdOptimizationSdp(SdOptimizationSdpNoMem): class SdOptimizationSubQuad(SdOptimization): - def __init__(self): - super().__init__("sub-quadratic", cmd_opt="opt_sub_quad_attention") - - def priority(self): - return 10 + name = "sub-quadratic" + cmd_opt = "opt_sub_quad_attention" + priority = 10 def apply(self): ldm.modules.attention.CrossAttention.forward = sub_quad_attention_forward @@ -99,20 +90,21 @@ class SdOptimizationSubQuad(SdOptimization): class SdOptimizationV1(SdOptimization): - def __init__(self): - super().__init__("V1", "original v1", cmd_opt="opt_split_attention_v1") + name = "V1" + label = "original v1" + cmd_opt = "opt_split_attention_v1" + priority = 10 - def priority(self): - return 10 def apply(self): ldm.modules.attention.CrossAttention.forward = split_cross_attention_forward_v1 class SdOptimizationInvokeAI(SdOptimization): - def __init__(self): - super().__init__("InvokeAI", cmd_opt="opt_split_attention_invokeai") + name = "InvokeAI" + cmd_opt = "opt_split_attention_invokeai" + @property def priority(self): return 1000 if not torch.cuda.is_available() else 10 @@ -121,11 +113,9 @@ class SdOptimizationInvokeAI(SdOptimization): class SdOptimizationDoggettx(SdOptimization): - def __init__(self): - super().__init__("Doggettx", cmd_opt="opt_split_attention") - - def priority(self): - return 20 + name = "Doggettx" + cmd_opt = "opt_split_attention" + priority = 20 def apply(self): ldm.modules.attention.CrossAttention.forward = split_cross_attention_forward -- cgit v1.2.1 From df004be2fc4b2c68adfb75565d97551a1a5e7ed6 Mon Sep 17 00:00:00 2001 From: Aarni Koskela Date: Sun, 21 May 2023 00:26:16 +0300 Subject: Add a couple `from __future__ import annotations`es for Py3.9 compat --- modules/sd_hijack_optimizations.py | 1 + 1 file changed, 1 insertion(+) (limited to 'modules/sd_hijack_optimizations.py') diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py index 0eb4c525..2ec0b049 100644 --- a/modules/sd_hijack_optimizations.py +++ b/modules/sd_hijack_optimizations.py @@ -1,3 +1,4 @@ +from __future__ import annotations import math import sys import traceback -- cgit v1.2.1