Merge pull request #340 from mit-han-lab/dev

feat: support PuLID, Double FBCache and TeaCache; better linter

nunchaku/models/pulid/eva_clip/model.py

+"""CLIP Model
+
+Adapted from https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.
+"""
+
+from dataclasses import dataclass
+from functools import partial
+from typing import Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+try:
+    from .hf_model import HFTextEncoder
+except ImportError:
+    HFTextEncoder = None
+from .eva_vit_model import EVAVisionTransformer
+from .modified_resnet import ModifiedResNet
+from .transformer import LayerNorm, QuickGELU, TextTransformer, VisionTransformer
+
+try:
+    from apex.normalization import FusedLayerNorm
+except ImportError:
+    FusedLayerNorm = LayerNorm
+    print("Please 'pip install apex'")
+
+try:
+    import xformers.ops as xops
+except ImportError:
+    xops = None
+    print("Please 'pip install xformers'")
+
+
+@dataclass
+class CLIPVisionCfg:
+    layers: Union[Tuple[int, int, int, int], int] = 12
+    width: int = 768
+    head_width: int = 64
+    mlp_ratio: float = 4.0
+    patch_size: int = 16
+    image_size: Union[Tuple[int, int], int] = 224
+    ls_init_value: Optional[float] = None  # layer scale initial value
+    patch_dropout: float = 0.0
+    global_average_pool: bool = (
+        False  # whether to global average pool the last embedding layer, instead of using CLS token (https://arxiv.org/abs/2205.01580)
+    )
+    drop_path_rate: Optional[float] = None  # drop path rate
+    timm_model_name: str = None  # a valid model name overrides layers, width, patch_size
+    timm_model_pretrained: bool = False  # use (imagenet) pretrained weights for named model
+    timm_pool: str = "avg"  # feature pooling for timm model ('abs_attn', 'rot_attn', 'avg', '')
+    timm_proj: str = "linear"  # linear projection for timm model output ('linear', 'mlp', '')
+    timm_proj_bias: bool = False  # enable bias final projection
+    eva_model_name: str = None  # a valid eva model name overrides layers, width, patch_size
+    qkv_bias: bool = True
+    fusedLN: bool = False
+    xattn: bool = False
+    postnorm: bool = False
+    rope: bool = False
+    pt_hw_seq_len: int = 16  # 224/14
+    intp_freq: bool = False
+    naiveswiglu: bool = False
+    subln: bool = False
+
+
+@dataclass
+class CLIPTextCfg:
+    context_length: int = 77
+    vocab_size: int = 49408
+    width: int = 512
+    heads: int = 8
+    layers: int = 12
+    ls_init_value: Optional[float] = None  # layer scale initial value
+    hf_model_name: str = None
+    hf_tokenizer_name: str = None
+    hf_model_pretrained: bool = True
+    proj: str = "mlp"
+    pooler_type: str = "mean_pooler"
+    masked_language_modeling: bool = False
+    fusedLN: bool = False
+    xattn: bool = False
+    attn_mask: bool = True
+
+
+def get_cast_dtype(precision: str):
+    cast_dtype = None
+    if precision == "bf16":
+        cast_dtype = torch.bfloat16
+    elif precision == "fp16":
+        cast_dtype = torch.float16
+    return cast_dtype
+
+
+def _build_vision_tower(
+    embed_dim: int, vision_cfg: CLIPVisionCfg, quick_gelu: bool = False, cast_dtype: Optional[torch.dtype] = None
+):
+    if isinstance(vision_cfg, dict):
+        vision_cfg = CLIPVisionCfg(**vision_cfg)
+
+    # OpenAI models are pretrained w/ QuickGELU but native nn.GELU is both faster and more
+    # memory efficient in recent PyTorch releases (>= 1.10).
+    # NOTE: timm models always use native GELU regardless of quick_gelu flag.
+    act_layer = QuickGELU if quick_gelu else nn.GELU
+
+    if vision_cfg.eva_model_name:
+        vision_heads = vision_cfg.width // vision_cfg.head_width
+        norm_layer = LayerNorm
+
+        visual = EVAVisionTransformer(
+            img_size=vision_cfg.image_size,
+            patch_size=vision_cfg.patch_size,
+            num_classes=embed_dim,
+            use_mean_pooling=vision_cfg.global_average_pool,  # False
+            init_values=vision_cfg.ls_init_value,
+            patch_dropout=vision_cfg.patch_dropout,
+            embed_dim=vision_cfg.width,
+            depth=vision_cfg.layers,
+            num_heads=vision_heads,
+            mlp_ratio=vision_cfg.mlp_ratio,
+            qkv_bias=vision_cfg.qkv_bias,
+            drop_path_rate=vision_cfg.drop_path_rate,
+            norm_layer=partial(FusedLayerNorm, eps=1e-6) if vision_cfg.fusedLN else partial(norm_layer, eps=1e-6),
+            xattn=vision_cfg.xattn,
+            rope=vision_cfg.rope,
+            postnorm=vision_cfg.postnorm,
+            pt_hw_seq_len=vision_cfg.pt_hw_seq_len,  # 224/14
+            intp_freq=vision_cfg.intp_freq,
+            naiveswiglu=vision_cfg.naiveswiglu,
+            subln=vision_cfg.subln,
+        )
+        act_layer = nn.GELU  # so that text transformer doesn't use QuickGELU w/ timm models
+    elif isinstance(vision_cfg.layers, (tuple, list)):
+        vision_heads = vision_cfg.width * 32 // vision_cfg.head_width
+        visual = ModifiedResNet(
+            layers=vision_cfg.layers,
+            output_dim=embed_dim,
+            heads=vision_heads,
+            image_size=vision_cfg.image_size,
+            width=vision_cfg.width,
+        )
+    else:
+        vision_heads = vision_cfg.width // vision_cfg.head_width
+        visual = VisionTransformer(
+            image_size=vision_cfg.image_size,
+            patch_size=vision_cfg.patch_size,
+            width=vision_cfg.width,
+            layers=vision_cfg.layers,
+            heads=vision_heads,
+            mlp_ratio=vision_cfg.mlp_ratio,
+            ls_init_value=vision_cfg.ls_init_value,
+            patch_dropout=vision_cfg.patch_dropout,
+            global_average_pool=vision_cfg.global_average_pool,
+            output_dim=embed_dim,
+            act_layer=act_layer,
+            norm_layer=norm_layer,
+        )
+
+    return visual
+
+
+def _build_text_tower(
+    embed_dim: int,
+    text_cfg: CLIPTextCfg,
+    quick_gelu: bool = False,
+    cast_dtype: Optional[torch.dtype] = None,
+):
+    if isinstance(text_cfg, dict):
+        text_cfg = CLIPTextCfg(**text_cfg)
+
+    if text_cfg.hf_model_name:
+        text = HFTextEncoder(
+            text_cfg.hf_model_name,
+            output_dim=embed_dim,
+            tokenizer_name=text_cfg.hf_tokenizer_name,
+            proj=text_cfg.proj,
+            pooler_type=text_cfg.pooler_type,
+            masked_language_modeling=text_cfg.masked_language_modeling,
+        )
+    else:
+        act_layer = QuickGELU if quick_gelu else nn.GELU
+        norm_layer = LayerNorm
+
+        text = TextTransformer(
+            context_length=text_cfg.context_length,
+            vocab_size=text_cfg.vocab_size,
+            width=text_cfg.width,
+            heads=text_cfg.heads,
+            layers=text_cfg.layers,
+            ls_init_value=text_cfg.ls_init_value,
+            output_dim=embed_dim,
+            act_layer=act_layer,
+            norm_layer=FusedLayerNorm if text_cfg.fusedLN else norm_layer,
+            xattn=text_cfg.xattn,
+            attn_mask=text_cfg.attn_mask,
+        )
+    return text
+
+
+class CLIP(nn.Module):
+    def __init__(
+        self,
+        embed_dim: int,
+        vision_cfg: CLIPVisionCfg,
+        text_cfg: CLIPTextCfg,
+        quick_gelu: bool = False,
+        cast_dtype: Optional[torch.dtype] = None,
+    ):
+        super().__init__()
+        self.visual = _build_vision_tower(embed_dim, vision_cfg, quick_gelu, cast_dtype)
+
+        text = _build_text_tower(embed_dim, text_cfg, quick_gelu, cast_dtype)
+        self.transformer = text.transformer
+        self.vocab_size = text.vocab_size
+        self.token_embedding = text.token_embedding
+        self.positional_embedding = text.positional_embedding
+        self.ln_final = text.ln_final
+        self.text_projection = text.text_projection
+        self.register_buffer("attn_mask", text.attn_mask, persistent=False)
+
+        self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        self.visual.set_grad_checkpointing(enable)
+        self.transformer.grad_checkpointing = enable
+
+    def encode_image(self, image, normalize: bool = False):
+        features = self.visual(image)
+        return F.normalize(features, dim=-1) if normalize else features
+
+    def encode_text(self, text, normalize: bool = False):
+        cast_dtype = self.transformer.get_cast_dtype()
+
+        x = self.token_embedding(text).to(cast_dtype)  # [batch_size, n_ctx, d_model]
+
+        x = x + self.positional_embedding.to(cast_dtype)
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.transformer(x, attn_mask=self.attn_mask)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+        x = self.ln_final(x)  # [batch_size, n_ctx, transformer.width]
+        # take features from the eot embedding (eot_token is the highest number in each sequence)
+        x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection
+        return F.normalize(x, dim=-1) if normalize else x
+
+    def forward(self, image, text):
+        image_features = self.encode_image(image, normalize=True)
+        text_features = self.encode_text(text, normalize=True)
+        return image_features, text_features, self.logit_scale.exp()
+
+
+class CustomCLIP(nn.Module):
+    def __init__(
+        self,
+        embed_dim: int,
+        vision_cfg: CLIPVisionCfg,
+        text_cfg: CLIPTextCfg,
+        quick_gelu: bool = False,
+        cast_dtype: Optional[torch.dtype] = None,
+        itm_task: bool = False,
+    ):
+        super().__init__()
+        self.visual = _build_vision_tower(embed_dim, vision_cfg, quick_gelu, cast_dtype)
+        self.text = _build_text_tower(embed_dim, text_cfg, quick_gelu, cast_dtype)
+        self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        self.visual.set_grad_checkpointing(enable)
+        self.text.set_grad_checkpointing(enable)
+
+    def encode_image(self, image, normalize: bool = False):
+        features = self.visual(image)
+        return F.normalize(features, dim=-1) if normalize else features
+
+    def encode_text(self, text, normalize: bool = False):
+        features = self.text(text)
+        return F.normalize(features, dim=-1) if normalize else features
+
+    def forward(self, image, text):
+        image_features = self.encode_image(image, normalize=True)
+        text_features = self.encode_text(text, normalize=True)
+        return image_features, text_features, self.logit_scale.exp()
+
+
+# used to maintain checkpoint compatibility
+def convert_to_custom_text_state_dict(state_dict: dict):
+    if "text_projection" in state_dict:
+        # old format state_dict, move text tower -> .text
+        new_state_dict = {}
+        for k, v in state_dict.items():
+            if any(
+                k.startswith(p)
+                for p in (
+                    "text_projection",
+                    "positional_embedding",
+                    "token_embedding",
+                    "transformer",
+                    "ln_final",
+                    "logit_scale",
+                )
+            ):
+                k = "text." + k
+            new_state_dict[k] = v
+        return new_state_dict
+    return state_dict

nunchaku/models/pulid/eva_clip/model_configs/EVA02-CLIP-L-14-336.json

+{
+    "embed_dim": 768,
+    "vision_cfg": {
+        "image_size": 336,
+        "layers": 24,
+        "width": 1024,
+        "drop_path_rate": 0,
+        "head_width": 64,
+        "mlp_ratio": 2.6667,
+        "patch_size": 14,
+        "eva_model_name": "eva-clip-l-14-336",
+        "xattn": true,
+        "fusedLN": true,
+        "rope": true,
+        "pt_hw_seq_len": 16,
+        "intp_freq": true,
+        "naiveswiglu": true,
+        "subln": true
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 768,
+        "heads": 12,
+        "layers": 12,
+        "xattn": false,
+        "fusedLN": true
+    }
+}

nunchaku/models/pulid/eva_clip/modified_resnet.py

+from collections import OrderedDict
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(self, inplanes, planes, stride=1):
+        super().__init__()
+
+        # all conv layers have stride 1. an avgpool is performed after the second convolution when stride > 1
+        self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(planes)
+        self.act1 = nn.ReLU(inplace=True)
+
+        self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.act2 = nn.ReLU(inplace=True)
+
+        self.avgpool = nn.AvgPool2d(stride) if stride > 1 else nn.Identity()
+
+        self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False)
+        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
+        self.act3 = nn.ReLU(inplace=True)
+
+        self.downsample = None
+        self.stride = stride
+
+        if stride > 1 or inplanes != planes * Bottleneck.expansion:
+            # downsampling layer is prepended with an avgpool, and the subsequent convolution has stride 1
+            self.downsample = nn.Sequential(
+                OrderedDict(
+                    [
+                        ("-1", nn.AvgPool2d(stride)),
+                        ("0", nn.Conv2d(inplanes, planes * self.expansion, 1, stride=1, bias=False)),
+                        ("1", nn.BatchNorm2d(planes * self.expansion)),
+                    ]
+                )
+            )
+
+    def forward(self, x: torch.Tensor):
+        identity = x
+
+        out = self.act1(self.bn1(self.conv1(x)))
+        out = self.act2(self.bn2(self.conv2(out)))
+        out = self.avgpool(out)
+        out = self.bn3(self.conv3(out))
+
+        if self.downsample is not None:
+            identity = self.downsample(x)
+
+        out += identity
+        out = self.act3(out)
+        return out
+
+
+class AttentionPool2d(nn.Module):
+    def __init__(self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(torch.randn(spacial_dim**2 + 1, embed_dim) / embed_dim**0.5)
+        self.k_proj = nn.Linear(embed_dim, embed_dim)
+        self.q_proj = nn.Linear(embed_dim, embed_dim)
+        self.v_proj = nn.Linear(embed_dim, embed_dim)
+        self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)
+        self.num_heads = num_heads
+
+    def forward(self, x):
+        x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3]).permute(2, 0, 1)  # NCHW -> (HW)NC
+        x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0)  # (HW+1)NC
+        x = x + self.positional_embedding[:, None, :].to(x.dtype)  # (HW+1)NC
+        x, _ = F.multi_head_attention_forward(
+            query=x,
+            key=x,
+            value=x,
+            embed_dim_to_check=x.shape[-1],
+            num_heads=self.num_heads,
+            q_proj_weight=self.q_proj.weight,
+            k_proj_weight=self.k_proj.weight,
+            v_proj_weight=self.v_proj.weight,
+            in_proj_weight=None,
+            in_proj_bias=torch.cat([self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]),
+            bias_k=None,
+            bias_v=None,
+            add_zero_attn=False,
+            dropout_p=0.0,
+            out_proj_weight=self.c_proj.weight,
+            out_proj_bias=self.c_proj.bias,
+            use_separate_proj_weight=True,
+            training=self.training,
+            need_weights=False,
+        )
+
+        return x[0]
+
+
+class ModifiedResNet(nn.Module):
+    """
+    A ResNet class that is similar to torchvision's but contains the following changes:
+    - There are now 3 "stem" convolutions as opposed to 1, with an average pool instead of a max pool.
+    - Performs anti-aliasing strided convolutions, where an avgpool is prepended to convolutions with stride > 1
+    - The final pooling layer is a QKV attention instead of an average pool
+    """
+
+    def __init__(self, layers, output_dim, heads, image_size=224, width=64):
+        super().__init__()
+        self.output_dim = output_dim
+        self.image_size = image_size
+
+        # the 3-layer stem
+        self.conv1 = nn.Conv2d(3, width // 2, kernel_size=3, stride=2, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(width // 2)
+        self.act1 = nn.ReLU(inplace=True)
+        self.conv2 = nn.Conv2d(width // 2, width // 2, kernel_size=3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(width // 2)
+        self.act2 = nn.ReLU(inplace=True)
+        self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False)
+        self.bn3 = nn.BatchNorm2d(width)
+        self.act3 = nn.ReLU(inplace=True)
+        self.avgpool = nn.AvgPool2d(2)
+
+        # residual layers
+        self._inplanes = width  # this is a *mutable* variable used during construction
+        self.layer1 = self._make_layer(width, layers[0])
+        self.layer2 = self._make_layer(width * 2, layers[1], stride=2)
+        self.layer3 = self._make_layer(width * 4, layers[2], stride=2)
+        self.layer4 = self._make_layer(width * 8, layers[3], stride=2)
+
+        embed_dim = width * 32  # the ResNet feature dimension
+        self.attnpool = AttentionPool2d(image_size // 32, embed_dim, heads, output_dim)
+
+        self.init_parameters()
+
+    def _make_layer(self, planes, blocks, stride=1):
+        layers = [Bottleneck(self._inplanes, planes, stride)]
+
+        self._inplanes = planes * Bottleneck.expansion
+        for _ in range(1, blocks):
+            layers.append(Bottleneck(self._inplanes, planes))
+
+        return nn.Sequential(*layers)
+
+    def init_parameters(self):
+        if self.attnpool is not None:
+            std = self.attnpool.c_proj.in_features**-0.5
+            nn.init.normal_(self.attnpool.q_proj.weight, std=std)
+            nn.init.normal_(self.attnpool.k_proj.weight, std=std)
+            nn.init.normal_(self.attnpool.v_proj.weight, std=std)
+            nn.init.normal_(self.attnpool.c_proj.weight, std=std)
+
+        for resnet_block in [self.layer1, self.layer2, self.layer3, self.layer4]:
+            for name, param in resnet_block.named_parameters():
+                if name.endswith("bn3.weight"):
+                    nn.init.zeros_(param)
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        # FIXME support for non-transformer
+        pass
+
+    def stem(self, x):
+        x = self.act1(self.bn1(self.conv1(x)))
+        x = self.act2(self.bn2(self.conv2(x)))
+        x = self.act3(self.bn3(self.conv3(x)))
+        x = self.avgpool(x)
+        return x
+
+    def forward(self, x):
+        x = self.stem(x)
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.attnpool(x)
+
+        return x

nunchaku/models/transformers/__init__.py

 from .transformer_flux import NunchakuFluxTransformer2dModel
 from .transformer_sana import NunchakuSanaTransformer2DModel
+
+__all__ = ["NunchakuFluxTransformer2dModel", "NunchakuSanaTransformer2DModel"]

nunchaku/pipeline/__init__.py

+from .pipeline_flux_pulid import PuLIDFluxPipeline
+
+__all__ = ["PuLIDFluxPipeline"]