Modify nn.MHA to attn with q/k/v

gen_img.py

@@ -20,7 +20,7 @@ import diffusers
 import numpy as np
 import torch
 
-from library.ipex_interop import init_ipex
+from library.device_utils import init_ipex, clean_memory, get_preferred_device
 
 init_ipex()
 
@@ -338,7 +338,7 @@ class PipelineLike:
         self.clip_vision_model: CLIPVisionModelWithProjection = None
         self.clip_vision_processor: CLIPImageProcessor = None
         self.clip_vision_strength = 0.0
-        
+
         # Textual Inversion
         self.token_replacements_list = []
         for _ in range(len(self.text_encoders)):

library/stable_cascade.py

@@ -5,6 +5,7 @@
 import math
 from types import SimpleNamespace
 from typing import List, Optional
+from einops import rearrange
 import numpy as np
 import torch
 import torch.nn as nn
@@ -148,7 +149,7 @@ class EfficientNetEncoder(nn.Module):
         The method to make it usable like VAE. It should be separated properly, but it is a temporary response.
         """
         # latents = vae.encode(img_tensors).latent_dist.sample().to("cpu")
-        
+
         # x is -1 to 1, so we need to convert it to 0 to 1, and then preprocess it with EfficientNet's preprocessing.
         x = (x + 1) / 2
         x = EFFNET_PREPROCESS(x)
@@ -172,6 +173,7 @@ from torch.nn import Conv2d
 from torch.nn import Linear
 
 
+r"""
 class Attention2D(nn.Module):
     def __init__(self, c, nhead, dropout=0.0):
         super().__init__()
@@ -185,6 +187,119 @@ class Attention2D(nn.Module):
         x = self.attn(x, kv, kv, need_weights=False)[0]
         x = x.permute(0, 2, 1).view(*orig_shape)
         return x
+"""
+
+
+class Attention(nn.Module):
+    def __init__(self, c, nhead, dropout=0.0):
+        # dropout is for attn_output_weights, so we may not need it. however, if we use sdpa, we enable it.
+        # xformers and normal attn are not affected by dropout
+        super().__init__()
+
+        self.to_q = Linear(c, c, bias=True)
+        self.to_k = Linear(c, c, bias=True)
+        self.to_v = Linear(c, c, bias=True)
+        self.to_out = Linear(c, c, bias=True)
+        self.nhead = nhead
+        self.dropout = dropout
+        self.scale = (c // nhead) ** -0.5
+
+        # default is to use sdpa
+        self.use_memory_efficient_attention_xformers = False
+        self.use_sdpa = True
+
+    def set_use_xformers_or_sdpa(self, xformers, sdpa):
+        # print(f"Attention: set_use_xformers_or_sdpa: xformers={xformers}, sdpa={sdpa}")
+        self.use_memory_efficient_attention_xformers = xformers
+        self.use_sdpa = sdpa
+
+    def forward(self, q_in, k_in, v_in):
+        q_in = self.to_q(q_in)
+        k_in = self.to_k(k_in)
+        v_in = self.to_v(v_in)
+
+        if self.use_memory_efficient_attention_xformers:
+            q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b n h d", h=self.nhead), (q_in, k_in, v_in))
+            del q_in, k_in, v_in
+            out = self.forward_memory_efficient_xformers(q, k, v)
+            del q, k, v
+            out = rearrange(out, "b n h d -> b n (h d)", h=self.nhead)
+        elif self.use_sdpa:
+            q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=self.nhead), (q_in, k_in, v_in))
+            del q_in, k_in, v_in
+            out = self.forward_sdpa(q, k, v)
+            del q, k, v
+            out = rearrange(out, "b h n d -> b n (h d)", h=self.nhead)
+        else:
+            q, k, v = map(lambda t: rearrange(t, "b n (h d) -> (b h) n d", h=self.nhead), (q_in, k_in, v_in))
+            del q_in, k_in, v_in
+            out = self._attention(q, k, v)
+            del q, k, v
+            out = rearrange(out, "(b h) n d -> b n (h d)", h=self.nhead)
+
+        return self.to_out(out)
+
+    def _attention(self, query, key, value):
+        # if self.upcast_attention:
+        #     query = query.float()
+        #     key = key.float()
+
+        attention_scores = torch.baddbmm(
+            torch.empty(query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device),
+            query,
+            key.transpose(-1, -2),
+            beta=0,
+            alpha=self.scale,
+        )
+        attention_probs = attention_scores.softmax(dim=-1)
+
+        # cast back to the original dtype
+        attention_probs = attention_probs.to(value.dtype)
+
+        # compute attention output
+        hidden_states = torch.bmm(attention_probs, value)
+
+        return hidden_states
+
+    def forward_memory_efficient_xformers(self, q, k, v):
+        import xformers.ops
+
+        q = q.contiguous()
+        k = k.contiguous()
+        v = v.contiguous()
+        out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None)  # 最適なのを選んでくれる
+        del q, k, v
+
+        return out
+
+    def forward_sdpa(self, q, k, v):
+        out = torch.nn.functional.scaled_dot_product_attention(q, k, v, dropout_p=self.dropout, is_causal=False)
+        return out
+
+
+class Attention2D(nn.Module):
+    r"""
+    to_q/k/v を個別に重みをもつように変更
+    modified to have separate weights for to_q/k/v
+    """
+
+    def __init__(self, c, nhead, dropout=0.0):
+        super().__init__()
+        # self.attn = nn.MultiheadAttention(c, nhead, dropout=dropout, bias=True, batch_first=True)
+        self.attn = Attention(c, nhead, dropout=dropout)  # , bias=True, batch_first=True)
+
+    def forward(self, x, kv, self_attn=False):
+        orig_shape = x.shape
+        x = x.view(x.size(0), x.size(1), -1).permute(0, 2, 1)  # Bx4xHxW -> Bx(HxW)x4
+        if self_attn:
+            kv = torch.cat([x, kv], dim=1)
+        # x = self.attn(x, kv, kv, need_weights=False)[0]
+        x = self.attn(x, kv, kv)
+        x = x.permute(0, 2, 1).view(*orig_shape)
+        return x
+
+    def set_use_xformers_or_sdpa(self, xformers, sdpa):
+        self.attn.set_use_xformers_or_sdpa(xformers, sdpa)
 
 
 class LayerNorm2d(nn.LayerNorm):
@@ -262,6 +377,9 @@ class AttnBlock(nn.Module):
     def set_gradient_checkpointing(self, value):
         self.gradient_checkpointing = value
 
+    def set_use_xformers_or_sdpa(self, xformers, sdpa):
+        self.attention.set_use_xformers_or_sdpa(xformers, sdpa)
+
     def forward_body(self, x, kv):
         kv = self.kv_mapper(kv)
         x = x + self.attention(self.norm(x), kv, self_attn=self.self_attn)
@@ -657,6 +775,12 @@ class StageB(nn.Module):
             if m.bias is not None:
                 nn.init.constant_(m.bias, 0)
 
+    def set_use_xformers_or_sdpa(self, xformers, sdpa):
+        for block in self.down_blocks + self.up_blocks:
+            for layer in block:
+                if hasattr(layer, "set_use_xformers_or_sdpa"):
+                    layer.set_use_xformers_or_sdpa(xformers, sdpa)
+
     def gen_r_embedding(self, r, max_positions=10000):
         r = r * max_positions
         half_dim = self.c_r // 2
@@ -920,6 +1044,12 @@ class StageC(nn.Module):
                 if hasattr(layer, "set_gradient_checkpointing"):
                     layer.set_gradient_checkpointing(value)
 
+    def set_use_xformers_or_sdpa(self, xformers, sdpa):
+        for block in self.down_blocks + self.up_blocks:
+            for layer in block:
+                if hasattr(layer, "set_use_xformers_or_sdpa"):
+                    layer.set_use_xformers_or_sdpa(xformers, sdpa)
+
     def gen_r_embedding(self, r, max_positions=10000):
         r = r * max_positions
         half_dim = self.c_r // 2

library/stable_cascade_utils.py

@@ -103,6 +103,9 @@ def load_stage_c_model(stage_c_checkpoint_path, dtype=None, device="cpu") -> sc.
         generator_c = sc.StageC()
     logger.info(f"Loading Stage C generator from {stage_c_checkpoint_path}")
     stage_c_checkpoint = load_file(stage_c_checkpoint_path)
+
+    stage_c_checkpoint = convert_state_dict_mha_to_normal_attn(stage_c_checkpoint)
+
     logger.info(f"Loading state dict")
     info = _load_state_dict_on_device(generator_c, stage_c_checkpoint, device, dtype=dtype)
     logger.info(info)
@@ -115,6 +118,9 @@ def load_stage_b_model(stage_b_checkpoint_path, dtype=None, device="cpu") -> sc.
         generator_b = sc.StageB()
     logger.info(f"Loading Stage B generator from {stage_b_checkpoint_path}")
     stage_b_checkpoint = load_file(stage_b_checkpoint_path)
+
+    stage_b_checkpoint = convert_state_dict_mha_to_normal_attn(stage_b_checkpoint)
+
     logger.info(f"Loading state dict")
     info = _load_state_dict_on_device(generator_b, stage_b_checkpoint, device, dtype=dtype)
     logger.info(info)
@@ -189,6 +195,55 @@ def load_previewer_model(previewer_checkpoint_path, dtype=None, device="cpu") ->
     return previewer
 
 
+def convert_state_dict_mha_to_normal_attn(state_dict):
+    # convert nn.MultiheadAttention to to_q/k/v and to_out
+    print("convert_state_dict_mha_to_normal_attn")
+    for key in list(state_dict.keys()):
+        if "attention.attn." in key:
+            if "in_proj_bias" in key:
+                value = state_dict.pop(key)
+                qkv = torch.chunk(value, 3, dim=0)
+                state_dict[key.replace("in_proj_bias", "to_q.bias")] = qkv[0]
+                state_dict[key.replace("in_proj_bias", "to_k.bias")] = qkv[1]
+                state_dict[key.replace("in_proj_bias", "to_v.bias")] = qkv[2]
+            elif "in_proj_weight" in key:
+                value = state_dict.pop(key)
+                qkv = torch.chunk(value, 3, dim=0)
+                state_dict[key.replace("in_proj_weight", "to_q.weight")] = qkv[0]
+                state_dict[key.replace("in_proj_weight", "to_k.weight")] = qkv[1]
+                state_dict[key.replace("in_proj_weight", "to_v.weight")] = qkv[2]
+            elif "out_proj.bias" in key:
+                value = state_dict.pop(key)
+                state_dict[key.replace("out_proj.bias", "to_out.bias")] = value
+            elif "out_proj.weight" in key:
+                value = state_dict.pop(key)
+                state_dict[key.replace("out_proj.weight", "to_out.weight")] = value
+    return state_dict
+
+
+def convert_state_dict_normal_attn_to_mha(state_dict):
+    # convert to_q/k/v and to_out to nn.MultiheadAttention
+    for key in list(state_dict.keys()):
+        if "attention.attn." in key:
+            if "to_q.bias" in key:
+                q = state_dict.pop(key)
+                k = state_dict.pop(key.replace("to_q.bias", "to_k.bias"))
+                v = state_dict.pop(key.replace("to_q.bias", "to_v.bias"))
+                state_dict[key.replace("to_q.bias", "in_proj_bias")] = torch.cat([q, k, v])
+            elif "to_q.weight" in key:
+                q = state_dict.pop(key)
+                k = state_dict.pop(key.replace("to_q.weight", "to_k.weight"))
+                v = state_dict.pop(key.replace("to_q.weight", "to_v.weight"))
+                state_dict[key.replace("to_q.weight", "in_proj_weight")] = torch.cat([q, k, v])
+            elif "to_out.bias" in key:
+                v = state_dict.pop(key)
+                state_dict[key.replace("to_out.bias", "out_proj.bias")] = v
+            elif "to_out.weight" in key:
+                v = state_dict.pop(key)
+                state_dict[key.replace("to_out.weight", "out_proj.weight")] = v
+    return state_dict
+
+
 def get_sai_model_spec(args, lora=False):
     timestamp = time.time()
 
@@ -230,6 +285,8 @@ def stage_c_saver_common(ckpt_file, stage_c, text_model, save_dtype, sai_metadat
     if save_dtype is not None:
         state_dict = {k: v.to(save_dtype) for k, v in state_dict.items()}
 
+    state_dict = convert_state_dict_normal_attn_to_mha(state_dict)
+
     save_file(state_dict, ckpt_file, metadata=sai_metadata)
 
     # save text model

networks/check_lora_weights.py

@@ -29,7 +29,7 @@ def main(file):
 
     for key, value in values:
         value = value.to(torch.float32)
-        logger.info(f"{key},{str(tuple(value.size())).replace(', ', '-')},{torch.mean(torch.abs(value))},{torch.min(torch.abs(value))}")
+        print(f"{key},{str(tuple(value.size())).replace(', ', '-')},{torch.mean(torch.abs(value))},{torch.min(torch.abs(value))}")
 
 
 def setup_parser() -> argparse.ArgumentParser:

stable_cascade_gen_img.py

@@ -40,9 +40,15 @@ def main(args):
 
     generator_c = sc_utils.load_stage_c_model(args.stage_c_checkpoint_path, dtype=dtype, device=loading_device)
     generator_c.eval().requires_grad_(False).to(loading_device)
+    # if args.xformers or args.sdpa:
+    print(f"Stage C: use_xformers_or_sdpa: {args.xformers} {args.sdpa}")
+    generator_c.set_use_xformers_or_sdpa(args.xformers, args.sdpa)
 
     generator_b = sc_utils.load_stage_b_model(args.stage_b_checkpoint_path, dtype=dtype, device=loading_device)
     generator_b.eval().requires_grad_(False).to(loading_device)
+    # if args.xformers or args.sdpa:
+    print(f"Stage B: use_xformers_or_sdpa: {args.xformers} {args.sdpa}")
+    generator_b.set_use_xformers_or_sdpa(args.xformers, args.sdpa)
 
     # CLIP encoders
     tokenizer = sc_utils.load_tokenizer(args)
@@ -332,6 +338,8 @@ if __name__ == "__main__":
     sc_utils.add_text_model_arguments(parser)
     parser.add_argument("--bf16", action="store_true")
     parser.add_argument("--fp16", action="store_true")
+    parser.add_argument("--xformers", action="store_true")
+    parser.add_argument("--sdpa", action="store_true")
     parser.add_argument("--outdir", type=str, default="../outputs", help="dir to write results to / 生成画像の出力先")
     parser.add_argument("--lowvram", action="store_true", help="if specified, use low VRAM mode")
     parser.add_argument(

stable_cascade_train_c_network.py

@@ -295,10 +295,9 @@ class NetworkTrainer:
         # text_encoder is List[CLIPTextModel] or CLIPTextModel
         text_encoders = text_encoder if isinstance(text_encoder, list) else [text_encoder]
 
-        # # モデルに xformers とか memory efficient attention を組み込む
+        # モデルに xformers とか memory efficient attention を組み込む
         # train_util.replace_unet_modules(unet, args.mem_eff_attn, args.xformers, args.sdpa)
-        # if torch.__version__ >= "2.0.0":  # PyTorch 2.0.0 以上対応のxformersなら以下が使える
+        stage_c.set_use_xformers_or_sdpa(args.xformers, args.sdpa)
-        #     vae.set_use_memory_efficient_attention_xformers(args.xformers)
 
         # 差分追加学習のためにモデルを読み込む
         sys.path.append(os.path.dirname(__file__))

stable_cascade_train_stage_c.py

@@ -147,6 +147,9 @@ def train(args):
     else:
         previewer = None
 
+    # モデルに xformers とか memory efficient attention を組み込む
+    stage_c.set_use_xformers_or_sdpa(args.xformers, args.sdpa)
+
     # 学習を準備する
     if cache_latents:
         effnet.to(accelerator.device, dtype=effnet_dtype)