diff --git a/README.md b/README.md index f51807ad5df1..cd8273b0d98c 100644 --- a/README.md +++ b/README.md @@ -173,7 +173,7 @@ There is a portable standalone build for Windows that should work for running on ### [Direct link to download](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia.7z) -Simply download, extract with [7-Zip](https://7-zip.org) and run. Make sure you put your Stable Diffusion checkpoints/models (the huge ckpt/safetensors files) in: ComfyUI\models\checkpoints +Simply download, extract with [7-Zip](https://7-zip.org) or with the windows explorer on recent windows versions and run. For smaller models you normally only need to put the checkpoints (the huge ckpt/safetensors files) in: ComfyUI\models\checkpoints but many of the larger models have multiple files. Make sure to follow the instructions to know which subfolder to put them in ComfyUI\models\ If you have trouble extracting it, right click the file -> properties -> unblock diff --git a/comfy/ldm/flux/layers.py b/comfy/ldm/flux/layers.py index a3eab0470d95..f4bf56e01b4d 100644 --- a/comfy/ldm/flux/layers.py +++ b/comfy/ldm/flux/layers.py @@ -167,39 +167,55 @@ def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=N img_modulated = self.img_norm1(img) img_modulated = apply_mod(img_modulated, (1 + img_mod1.scale), img_mod1.shift, modulation_dims_img) img_qkv = self.img_attn.qkv(img_modulated) + del img_modulated img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) + del img_qkv img_q, img_k = self.img_attn.norm(img_q, img_k, img_v) # prepare txt for attention txt_modulated = self.txt_norm1(txt) txt_modulated = apply_mod(txt_modulated, (1 + txt_mod1.scale), txt_mod1.shift, modulation_dims_txt) txt_qkv = self.txt_attn.qkv(txt_modulated) + del txt_modulated txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) + del txt_qkv txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v) if self.flipped_img_txt: + q = torch.cat((img_q, txt_q), dim=2) + del img_q, txt_q + k = torch.cat((img_k, txt_k), dim=2) + del img_k, txt_k + v = torch.cat((img_v, txt_v), dim=2) + del img_v, txt_v # run actual attention - attn = attention(torch.cat((img_q, txt_q), dim=2), - torch.cat((img_k, txt_k), dim=2), - torch.cat((img_v, txt_v), dim=2), + attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options) + del q, k, v img_attn, txt_attn = attn[:, : img.shape[1]], attn[:, img.shape[1]:] else: + q = torch.cat((txt_q, img_q), dim=2) + del txt_q, img_q + k = torch.cat((txt_k, img_k), dim=2) + del txt_k, img_k + v = torch.cat((txt_v, img_v), dim=2) + del txt_v, img_v # run actual attention - attn = attention(torch.cat((txt_q, img_q), dim=2), - torch.cat((txt_k, img_k), dim=2), - torch.cat((txt_v, img_v), dim=2), + attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options) + del q, k, v txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:] # calculate the img bloks img += apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img) + del img_attn img += apply_mod(self.img_mlp(apply_mod(self.img_norm2(img), (1 + img_mod2.scale), img_mod2.shift, modulation_dims_img)), img_mod2.gate, None, modulation_dims_img) # calculate the txt bloks txt += apply_mod(self.txt_attn.proj(txt_attn), txt_mod1.gate, None, modulation_dims_txt) + del txt_attn txt += apply_mod(self.txt_mlp(apply_mod(self.txt_norm2(txt), (1 + txt_mod2.scale), txt_mod2.shift, modulation_dims_txt)), txt_mod2.gate, None, modulation_dims_txt) if txt.dtype == torch.float16: @@ -249,12 +265,15 @@ def forward(self, x: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1) q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) + del qkv q, k = self.norm(q, k, v) # compute attention attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options) + del q, k, v # compute activation in mlp stream, cat again and run second linear layer - output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2)) + mlp = self.mlp_act(mlp) + output = self.linear2(torch.cat((attn, mlp), 2)) x += apply_mod(output, mod.gate, None, modulation_dims) if x.dtype == torch.float16: x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504) diff --git a/server.py b/server.py index 5d773b10a57f..d059d3dc9085 100644 --- a/server.py +++ b/server.py @@ -2,6 +2,7 @@ import sys import asyncio import traceback +import time import nodes import folder_paths @@ -733,6 +734,7 @@ async def post_prompt(request): for sensitive_val in execution.SENSITIVE_EXTRA_DATA_KEYS: if sensitive_val in extra_data: sensitive[sensitive_val] = extra_data.pop(sensitive_val) + extra_data["create_time"] = int(time.time() * 1000) # timestamp in milliseconds self.prompt_queue.put((number, prompt_id, prompt, extra_data, outputs_to_execute, sensitive)) response = {"prompt_id": prompt_id, "number": number, "node_errors": valid[3]} return web.json_response(response)