From 84d9ce30cb427759547bc7876ed80ab91787d175 Mon Sep 17 00:00:00 2001 From: brkirch Date: Tue, 24 Jan 2023 23:51:45 -0500 Subject: Add option for float32 sampling with float16 UNet This also handles type casting so that ROCm and MPS torch devices work correctly without --no-half. One cast is required for deepbooru in deepbooru_model.py, some explicit casting is required for img2img and inpainting. depth_model can't be converted to float16 or it won't work correctly on some systems (it's known to have issues on MPS) so in sd_models.py model.depth_model is removed for model.half(). --- modules/processing.py | 15 ++++++++------- 1 file changed, 8 insertions(+), 7 deletions(-) (limited to 'modules/processing.py') diff --git a/modules/processing.py b/modules/processing.py index bc541e2f..2d186ba0 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -172,7 +172,8 @@ class StableDiffusionProcessing: midas_in = torch.from_numpy(transformed["midas_in"][None, ...]).to(device=shared.device) midas_in = repeat(midas_in, "1 ... -> n ...", n=self.batch_size) - conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(source_image)) + conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(source_image.to(devices.dtype_unet) if devices.unet_needs_upcast else source_image)) + conditioning_image = conditioning_image.float() if devices.unet_needs_upcast else conditioning_image conditioning = torch.nn.functional.interpolate( self.sd_model.depth_model(midas_in), size=conditioning_image.shape[2:], @@ -203,7 +204,7 @@ class StableDiffusionProcessing: # Create another latent image, this time with a masked version of the original input. # Smoothly interpolate between the masked and unmasked latent conditioning image using a parameter. - conditioning_mask = conditioning_mask.to(source_image.device).to(source_image.dtype) + conditioning_mask = conditioning_mask.to(device=source_image.device, dtype=source_image.dtype) conditioning_image = torch.lerp( source_image, source_image * (1.0 - conditioning_mask), @@ -211,7 +212,7 @@ class StableDiffusionProcessing: ) # Encode the new masked image using first stage of network. - conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(conditioning_image)) + conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(conditioning_image.to(devices.dtype_unet) if devices.unet_needs_upcast else conditioning_image)) # Create the concatenated conditioning tensor to be fed to `c_concat` conditioning_mask = torch.nn.functional.interpolate(conditioning_mask, size=latent_image.shape[-2:]) @@ -225,10 +226,10 @@ class StableDiffusionProcessing: # HACK: Using introspection as the Depth2Image model doesn't appear to uniquely # identify itself with a field common to all models. The conditioning_key is also hybrid. if isinstance(self.sd_model, LatentDepth2ImageDiffusion): - return self.depth2img_image_conditioning(source_image) + return self.depth2img_image_conditioning(source_image.float() if devices.unet_needs_upcast else source_image) if self.sampler.conditioning_key in {'hybrid', 'concat'}: - return self.inpainting_image_conditioning(source_image, latent_image, image_mask=image_mask) + return self.inpainting_image_conditioning(source_image.float() if devices.unet_needs_upcast else source_image, latent_image, image_mask=image_mask) # Dummy zero conditioning if we're not using inpainting or depth model. return latent_image.new_zeros(latent_image.shape[0], 5, 1, 1) @@ -610,7 +611,7 @@ def process_images_inner(p: StableDiffusionProcessing) -> Processed: if p.n_iter > 1: shared.state.job = f"Batch {n+1} out of {p.n_iter}" - with devices.autocast(): + with devices.autocast(disable=devices.unet_needs_upcast): samples_ddim = p.sample(conditioning=c, unconditional_conditioning=uc, seeds=seeds, subseeds=subseeds, subseed_strength=p.subseed_strength, prompts=prompts) x_samples_ddim = [decode_first_stage(p.sd_model, samples_ddim[i:i+1].to(dtype=devices.dtype_vae))[0].cpu() for i in range(samples_ddim.size(0))] @@ -988,7 +989,7 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing): image = torch.from_numpy(batch_images) image = 2. * image - 1. - image = image.to(shared.device) + image = image.to(device=shared.device, dtype=devices.dtype_unet if devices.unet_needs_upcast else None) self.init_latent = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(image)) -- cgit v1.2.3 From e3b53fd295aca784253dfc8668ec87b537a72f43 Mon Sep 17 00:00:00 2001 From: brkirch Date: Wed, 25 Jan 2023 00:23:10 -0500 Subject: Add UI setting for upcasting attention to float32 Adds "Upcast cross attention layer to float32" option in Stable Diffusion settings. This allows for generating images using SD 2.1 models without --no-half or xFormers. In order to make upcasting cross attention layer optimizations possible it is necessary to indent several sections of code in sd_hijack_optimizations.py so that a context manager can be used to disable autocast. Also, even though Stable Diffusion (and Diffusers) only upcast q and k, unfortunately my findings were that most of the cross attention layer optimizations could not function unless v is upcast also. --- modules/devices.py | 6 +- modules/processing.py | 2 +- modules/sd_hijack_optimizations.py | 159 +++++++++++++++++++++++-------------- modules/shared.py | 1 + modules/sub_quadratic_attention.py | 4 +- 5 files changed, 108 insertions(+), 64 deletions(-) (limited to 'modules/processing.py') diff --git a/modules/devices.py b/modules/devices.py index 0981ef80..6b36622c 100644 --- a/modules/devices.py +++ b/modules/devices.py @@ -108,6 +108,10 @@ def autocast(disable=False): return torch.autocast("cuda") +def without_autocast(disable=False): + return torch.autocast("cuda", enabled=False) if torch.is_autocast_enabled() and not disable else contextlib.nullcontext() + + class NansException(Exception): pass @@ -125,7 +129,7 @@ def test_for_nans(x, where): message = "A tensor with all NaNs was produced in Unet." if not shared.cmd_opts.no_half: - message += " This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try using --no-half commandline argument to fix this." + message += " This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the \"Upcast cross attention layer to float32\" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this." elif where == "vae": message = "A tensor with all NaNs was produced in VAE." diff --git a/modules/processing.py b/modules/processing.py index 2d186ba0..a850082d 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -611,7 +611,7 @@ def process_images_inner(p: StableDiffusionProcessing) -> Processed: if p.n_iter > 1: shared.state.job = f"Batch {n+1} out of {p.n_iter}" - with devices.autocast(disable=devices.unet_needs_upcast): + with devices.without_autocast() if devices.unet_needs_upcast else devices.autocast(): samples_ddim = p.sample(conditioning=c, unconditional_conditioning=uc, seeds=seeds, subseeds=subseeds, subseed_strength=p.subseed_strength, prompts=prompts) x_samples_ddim = [decode_first_stage(p.sd_model, samples_ddim[i:i+1].to(dtype=devices.dtype_vae))[0].cpu() for i in range(samples_ddim.size(0))] diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py index 74452709..c02d954c 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 +from modules import shared, errors, devices from modules.hypernetworks import hypernetwork from .sub_quadratic_attention import efficient_dot_product_attention @@ -52,18 +52,25 @@ def split_cross_attention_forward_v1(self, x, context=None, mask=None): 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) - 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 + dtype = q.dtype + if shared.opts.upcast_attn: + q, k, v = q.float(), k.float(), v.float() - s2 = s1.softmax(dim=-1) - del s1 + 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[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 @@ -82,45 +89,52 @@ def split_cross_attention_forward(self, x, context=None, mask=None): k_in = self.to_k(context_k) v_in = self.to_v(context_v) - 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}") + 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() - 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 + 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 - del q, k, v + r1 = r1.to(dtype) r2 = rearrange(r1, '(b h) n d -> b n (h d)', h=h) del r1 @@ -204,12 +218,20 @@ def split_cross_attention_forward_invokeAI(self, x, context=None, mask=None): context = default(context, x) context_k, context_v = hypernetwork.apply_hypernetworks(shared.loaded_hypernetworks, context) - k = self.to_k(context_k) * self.scale + k = self.to_k(context_k) v = 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, k, v)) - r = einsum_op(q, k, v) + 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 -- @@ -234,8 +256,14 @@ def sub_quad_attention_forward(self, x, context=None, mask=None): 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 @@ -268,15 +296,16 @@ def sub_quad_attention(q, k, v, q_chunk_size=1024, kv_chunk_size=None, kv_chunk_ 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, - ) + 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): @@ -306,8 +335,14 @@ def xformers_attention_forward(self, x, context=None, mask=None): 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) @@ -378,10 +413,14 @@ def xformers_attnblock_forward(self, x): 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 diff --git a/modules/shared.py b/modules/shared.py index 4ce1209b..6a0b96cb 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -410,6 +410,7 @@ options_templates.update(options_section(('sd', "Stable Diffusion"), { "comma_padding_backtrack": OptionInfo(20, "Increase coherency by padding from the last comma within n tokens when using more than 75 tokens", gr.Slider, {"minimum": 0, "maximum": 74, "step": 1 }), "CLIP_stop_at_last_layers": OptionInfo(1, "Clip skip", gr.Slider, {"minimum": 1, "maximum": 12, "step": 1}), "extra_networks_default_multiplier": OptionInfo(1.0, "Multiplier for extra networks", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.01}), + "upcast_attn": OptionInfo(False, "Upcast cross attention layer to float32"), })) options_templates.update(options_section(('compatibility', "Compatibility"), { diff --git a/modules/sub_quadratic_attention.py b/modules/sub_quadratic_attention.py index 55052815..05595323 100644 --- a/modules/sub_quadratic_attention.py +++ b/modules/sub_quadratic_attention.py @@ -67,7 +67,7 @@ def _summarize_chunk( 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) + exp_values = torch.bmm(exp_weights, value) if query.device.type == 'mps' else torch.bmm(exp_weights, value.to(exp_weights.dtype)).to(value.dtype) max_score = max_score.squeeze(-1) return AttnChunk(exp_values, exp_weights.sum(dim=-1), max_score) @@ -129,7 +129,7 @@ def _get_attention_scores_no_kv_chunking( ) attn_probs = attn_scores.softmax(dim=-1) del attn_scores - hidden_states_slice = torch.bmm(attn_probs, value) + hidden_states_slice = torch.bmm(attn_probs, value) if query.device.type == 'mps' else torch.bmm(attn_probs, value.to(attn_probs.dtype)).to(value.dtype) return hidden_states_slice -- cgit v1.2.3