self-contained sampling code

sample.py

## MM DiT model that was proposed by SD3 paper.
# I've tried to make this follow the work of MuP, so they scale in maximal feature-learning regime.

import math

import torch
import torch.nn as nn
import torch.nn.functional as F


class Fp32LayerNorm(nn.LayerNorm):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

    def forward(self, input):
        output = F.layer_norm(
            input.float(),
            self.normalized_shape,
            self.weight.float() if self.weight is not None else None,
            self.bias.float() if self.bias is not None else None,
            self.eps,
        )
        return output.type_as(input)


def modulate(x, shift, scale):
    return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)


def find_multiple(n: int, k: int) -> int:
    if n % k == 0:
        return n
    return n + k - (n % k)


class MLP(nn.Module):
    def __init__(self, dim, hidden_dim=None) -> None:
        super().__init__()
        if hidden_dim is None:
            hidden_dim = 4 * dim

        n_hidden = int(2 * hidden_dim / 3)
        n_hidden = find_multiple(n_hidden, 256)

        self.c_fc1 = nn.Linear(dim, n_hidden, bias=False)
        self.c_fc2 = nn.Linear(dim, n_hidden, bias=False)
        self.c_proj = nn.Linear(n_hidden, dim, bias=False)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = F.silu(self.c_fc1(x)) * self.c_fc2(x)
        x = self.c_proj(x)
        return x


class MultiHeadLayerNorm(nn.Module):
    def __init__(self, hidden_size=None, eps=1e-5):
        # Copy pasta from https://github.com/huggingface/transformers/blob/e5f71ecaae50ea476d1e12351003790273c4b2ed/src/transformers/models/cohere/modeling_cohere.py#L78

        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)
        mean = hidden_states.mean(-1, keepdim=True)
        variance = (hidden_states - mean).pow(2).mean(-1, keepdim=True)
        hidden_states = (hidden_states - mean) * torch.rsqrt(
            variance + self.variance_epsilon
        )
        hidden_states = self.weight.to(torch.float32) * hidden_states
        return hidden_states.to(input_dtype)

class SingleAttention(nn.Module):
    def __init__(self, dim, n_heads, mh_qknorm=False):
        super().__init__()

        self.n_heads = n_heads
        self.head_dim = dim // n_heads

        # this is for cond
        self.w1q = nn.Linear(dim, dim, bias=False)
        self.w1k = nn.Linear(dim, dim, bias=False)
        self.w1v = nn.Linear(dim, dim, bias=False)
        self.w1o = nn.Linear(dim, dim, bias=False)

        self.q_norm1 = (
            MultiHeadLayerNorm((self.n_heads, self.head_dim))
            if mh_qknorm
            else Fp32LayerNorm(self.head_dim, bias=False, elementwise_affine=False)
        )
        self.k_norm1 = (
            MultiHeadLayerNorm((self.n_heads, self.head_dim))
            if mh_qknorm
            else Fp32LayerNorm(self.head_dim, bias=False, elementwise_affine=False)
        )
    
    #@torch.compile()
    def forward(self, c):

        bsz, seqlen1, _ = c.shape

        q, k, v = self.w1q(c), self.w1k(c), self.w1v(c)
        q = q.view(bsz, seqlen1, self.n_heads, self.head_dim)
        k = k.view(bsz, seqlen1, self.n_heads, self.head_dim)
        v = v.view(bsz, seqlen1, self.n_heads, self.head_dim)
        q, k = self.q_norm1(q), self.k_norm1(k)

        output = F.scaled_dot_product_attention(
            q.permute(0, 2, 1, 3),
            k.permute(0, 2, 1, 3),
            v.permute(0, 2, 1, 3),
            dropout_p=0.0,
            is_causal=False,
            scale=1 / self.head_dim**0.5,
        ).permute(0, 2, 1, 3)
        output = output.flatten(-2)
        c = self.w1o(output)
        return c



class DoubleAttention(nn.Module):
    def __init__(self, dim, n_heads, mh_qknorm=False):
        super().__init__()

        self.n_heads = n_heads
        self.head_dim = dim // n_heads

        # this is for cond
        self.w1q = nn.Linear(dim, dim, bias=False)
        self.w1k = nn.Linear(dim, dim, bias=False)
        self.w1v = nn.Linear(dim, dim, bias=False)
        self.w1o = nn.Linear(dim, dim, bias=False)

        # this is for x
        self.w2q = nn.Linear(dim, dim, bias=False)
        self.w2k = nn.Linear(dim, dim, bias=False)
        self.w2v = nn.Linear(dim, dim, bias=False)
        self.w2o = nn.Linear(dim, dim, bias=False)

        self.q_norm1 = (
            MultiHeadLayerNorm((self.n_heads, self.head_dim))
            if mh_qknorm
            else Fp32LayerNorm(self.head_dim, bias=False, elementwise_affine=False)
        )
        self.k_norm1 = (
            MultiHeadLayerNorm((self.n_heads, self.head_dim))
            if mh_qknorm
            else Fp32LayerNorm(self.head_dim, bias=False, elementwise_affine=False)
        )

        self.q_norm2 = (
            MultiHeadLayerNorm((self.n_heads, self.head_dim))
            if mh_qknorm
            else Fp32LayerNorm(self.head_dim, bias=False, elementwise_affine=False)
        )
        self.k_norm2 = (
            MultiHeadLayerNorm((self.n_heads, self.head_dim))
            if mh_qknorm
            else Fp32LayerNorm(self.head_dim, bias=False, elementwise_affine=False)
        )


    #@torch.compile()
    def forward(self, c, x):

        bsz, seqlen1, _ = c.shape
        bsz, seqlen2, _ = x.shape
        seqlen = seqlen1 + seqlen2

        cq, ck, cv = self.w1q(c), self.w1k(c), self.w1v(c)
        cq = cq.view(bsz, seqlen1, self.n_heads, self.head_dim)
        ck = ck.view(bsz, seqlen1, self.n_heads, self.head_dim)
        cv = cv.view(bsz, seqlen1, self.n_heads, self.head_dim)
        cq, ck = self.q_norm1(cq), self.k_norm1(ck)

        xq, xk, xv = self.w2q(x), self.w2k(x), self.w2v(x)
        xq = xq.view(bsz, seqlen2, self.n_heads, self.head_dim)
        xk = xk.view(bsz, seqlen2, self.n_heads, self.head_dim)
        xv = xv.view(bsz, seqlen2, self.n_heads, self.head_dim)
        xq, xk = self.q_norm2(xq), self.k_norm2(xk)

        # concat all
        q, k, v = (
            torch.cat([cq, xq], dim=1),
            torch.cat([ck, xk], dim=1),
            torch.cat([cv, xv], dim=1),
        )

        output = F.scaled_dot_product_attention(
            q.permute(0, 2, 1, 3),
            k.permute(0, 2, 1, 3),
            v.permute(0, 2, 1, 3),
            dropout_p=0.0,
            is_causal=False,
            scale=1 / self.head_dim**0.5,
        ).permute(0, 2, 1, 3)
        output = output.flatten(-2)
        c, x = output.split([seqlen1, seqlen2], dim=1)
        c = self.w1o(c)
        x = self.w2o(x)

        return c, x


class MMDiTBlock(nn.Module):
    def __init__(self, dim, heads=8, global_conddim=1024, is_last=False):
        super().__init__()

        self.normC1 = Fp32LayerNorm(dim, elementwise_affine=False, bias=False)
        self.normC2 = Fp32LayerNorm(dim, elementwise_affine=False, bias=False)
        if not is_last:
            self.mlpC = MLP(dim, hidden_dim=dim * 4)
            self.modC = nn.Sequential(
                nn.SiLU(),
                nn.Linear(global_conddim, 6 * dim, bias=False),
            )
        else:
            self.modC = nn.Sequential(
                nn.SiLU(),
                nn.Linear(global_conddim, 2 * dim, bias=False),
            )

        self.normX1 = Fp32LayerNorm(dim, elementwise_affine=False, bias=False)
        self.normX2 = Fp32LayerNorm(dim, elementwise_affine=False, bias=False)
        self.mlpX = MLP(dim, hidden_dim=dim * 4)
        self.modX = nn.Sequential(
            nn.SiLU(),
            nn.Linear(global_conddim, 6 * dim, bias=False),
        )

        self.attn = DoubleAttention(dim, heads)
        self.is_last = is_last

    #@torch.compile()
    def forward(self, c, x, global_cond, **kwargs):

        cres, xres = c, x
        
        cshift_msa, cscale_msa, cgate_msa, cshift_mlp, cscale_mlp, cgate_mlp = (
            self.modC(global_cond).chunk(6, dim=1)
        )
      
        c = modulate(self.normC1(c), cshift_msa, cscale_msa)

        # xpath
        xshift_msa, xscale_msa, xgate_msa, xshift_mlp, xscale_mlp, xgate_mlp = (
            self.modX(global_cond).chunk(6, dim=1)
        )

        x = modulate(self.normX1(x), xshift_msa, xscale_msa)

        # attention
        c, x = self.attn(c, x)


        c = self.normC2(cres + cgate_msa.unsqueeze(1) * c)
        c = cgate_mlp.unsqueeze(1) * self.mlpC(modulate(c, cshift_mlp, cscale_mlp))
        c = cres + c

        x = self.normX2(xres + xgate_msa.unsqueeze(1) * x)
        x = xgate_mlp.unsqueeze(1) * self.mlpX(modulate(x, xshift_mlp, xscale_mlp))
        x = xres + x

        return c, x

class DiTBlock(nn.Module):
    # like MMDiTBlock, but it only has X
    def __init__(self, dim, heads=8, global_conddim=1024):
        super().__init__()

        self.norm1 = Fp32LayerNorm(dim, elementwise_affine=False, bias=False)
        self.norm2 = Fp32LayerNorm(dim, elementwise_affine=False, bias=False)

        self.modCX = nn.Sequential(
            nn.SiLU(),
            nn.Linear(global_conddim, 6 * dim, bias=False),
        )

        self.attn = SingleAttention(dim, heads)
        self.mlp = MLP(dim, hidden_dim=dim * 4)

    #@torch.compile()
    def forward(self, cx, global_cond, **kwargs):
        cxres = cx
        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.modCX(
            global_cond
        ).chunk(6, dim=1)
        cx = modulate(self.norm1(cx), shift_msa, scale_msa)
        cx = self.attn(cx)
        cx = self.norm2(cxres + gate_msa.unsqueeze(1) * cx)
        mlpout = self.mlp(modulate(cx, shift_mlp, scale_mlp))
        cx = gate_mlp.unsqueeze(1) * mlpout

        cx = cxres + cx

        return cx



class TimestepEmbedder(nn.Module):
    def __init__(self, hidden_size, frequency_embedding_size=256):
        super().__init__()
        self.mlp = nn.Sequential(
            nn.Linear(frequency_embedding_size, hidden_size),
            nn.SiLU(),
            nn.Linear(hidden_size, hidden_size),
        )
        self.frequency_embedding_size = frequency_embedding_size

    @staticmethod
    def timestep_embedding(t, dim, max_period=10000):
        half = dim // 2
        freqs = 1000 * torch.exp(
            -math.log(max_period) * torch.arange(start=0, end=half) / half
        ).to(t.device)
        args = t[:, None] * freqs[None]
        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
        if dim % 2:
            embedding = torch.cat(
                [embedding, torch.zeros_like(embedding[:, :1])], dim=-1
            )
        return embedding

    #@torch.compile()
    def forward(self, t):
        t_freq = self.timestep_embedding(t, self.frequency_embedding_size).to(
            dtype=next(self.parameters()).dtype
        )
        t_emb = self.mlp(t_freq)
        return t_emb


class MMDiT(nn.Module):
    def __init__(
        self,
        in_channels=4,
        out_channels=4,
        patch_size=2,
        dim=2048,
        n_layers=8,
        n_double_layers=4,
        n_heads=4,
        global_conddim=1024,
        cond_seq_dim=2048,
        max_seq=16 * 16,
    ):
        super().__init__()

        self.t_embedder = TimestepEmbedder(global_conddim)
      
        self.cond_seq_linear = nn.Linear(
            cond_seq_dim, dim, bias=False
        )  # linear for something like text sequence.
        self.init_x_linear = nn.Linear(
            patch_size * patch_size * in_channels, dim
        )  # init linear for patchified image.

        self.positional_encoding = nn.Parameter(torch.randn(1, max_seq, dim) * 0.1)
        self.register_tokens = nn.Parameter(torch.randn(1, 8, dim) * 0.02)

        self.double_layers = nn.ModuleList([])
        self.single_layers = nn.ModuleList([])


        for idx in range(n_double_layers):
            self.double_layers.append(
                MMDiTBlock(dim, n_heads, global_conddim, is_last=(idx == n_layers - 1))
            )
        
        for idx in range(n_double_layers, n_layers):
            self.single_layers.append(
                DiTBlock(dim, n_heads, global_conddim)
            )


        self.final_linear = nn.Linear(
            dim, patch_size * patch_size * out_channels, bias=False
        )

        self.modF = nn.Sequential(
            nn.SiLU(),
            nn.Linear(global_conddim, 2 * dim, bias=False),
        )
        nn.init.constant_(self.final_linear.weight, 0)
       
        self.out_channels = out_channels
        self.patch_size = patch_size
        self.n_double_layers = n_double_layers
        self.n_layers = n_layers

        for pn, p in self.named_parameters():
            if ".mod" in pn:
                nn.init.constant_(p, 0)
                print("zeroed", pn)

        # if cond_seq_linear
        nn.init.constant_(self.cond_seq_linear.weight, 0)
        self.h_max = int(max_seq**0.5)
        self.w_max = int(max_seq**0.5)

    @torch.no_grad()
    def extend_pe(self, init_dim=(16, 16), target_dim=(64, 64)):
        # extend pe
        pe_data = self.positional_encoding.data.squeeze(0)[: init_dim[0] * init_dim[1]]

        pe_as_2d = pe_data.view(init_dim[0], init_dim[1], -1).permute(2, 0, 1)

        # now we need to extend this to target_dim. for this we will use interpolation.
        # we will use torch.nn.functional.interpolate
        pe_as_2d = F.interpolate(
            pe_as_2d.unsqueeze(0), size=target_dim, mode="bilinear"
        )
        pe_new = pe_as_2d.squeeze(0).permute(1, 2, 0).flatten(0, 1)
        self.positional_encoding.data = pe_new.unsqueeze(0).contiguous()
        self.h_max, self.w_max = target_dim
        print("PE extended to", target_dim)

    def pe_selection_index_based_on_dim(self, h, w):
        h_p, w_p = h // self.patch_size, w // self.patch_size
        original_pe_indexes = torch.arange(self.positional_encoding.shape[1])
        original_pe_indexes = original_pe_indexes.view(self.h_max, self.w_max)
        original_pe_indexes = original_pe_indexes[
            self.h_max // 2 - h_p // 2 : self.h_max // 2 + h_p // 2,
            self.w_max // 2 - w_p // 2 : self.w_max // 2 + w_p // 2,
        ]
        return original_pe_indexes.flatten()

    def unpatchify(self, x, h, w):
        c = self.out_channels
        p = self.patch_size

        x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
        x = torch.einsum("nhwpqc->nchpwq", x)
        imgs = x.reshape(shape=(x.shape[0], c, h * p, w * p))
        return imgs

    def patchify(self, x):
        B, C, H, W = x.size()
        x = x.view(
            B,
            C,
            H // self.patch_size,
            self.patch_size,
            W // self.patch_size,
            self.patch_size,
        )
        x = x.permute(0, 2, 4, 1, 3, 5).flatten(-3).flatten(1, 2)
        return x

    def forward(self, x, t, conds, **kwargs):
        # patchify x, add PE
        b, c, h, w = x.shape

        # pe_indexes = self.pe_selection_index_based_on_dim(h, w)
        # print(pe_indexes.shape)

        x = self.init_x_linear(self.patchify(x))  # B, T_x, D
        x = x + self.positional_encoding[:, : x.size(1)]

        # process conditions for MMDiT Blocks
        c_seq = conds["c_seq"][0:b]  # B, T_c, D_c
        t = t[0:b]
   
        c = self.cond_seq_linear(c_seq)  # B, T_c, D
        c = torch.cat([self.register_tokens.repeat(c.size(0), 1, 1), c], dim=1)
        
        global_cond = self.t_embedder(t)  # B, D
     
        if len(self.double_layers) > 0:
            for layer in self.double_layers:
              
                c, x = layer(c, x, global_cond, **kwargs)
                
        if len(self.single_layers) > 0:
            c_len = c.size(1)
            cx = torch.cat([c, x], dim=1)
            for layer in self.single_layers:
      
                cx = layer(cx, global_cond, **kwargs)

            x = cx[:, c_len:]

        fshift, fscale = self.modF(global_cond).chunk(2, dim=1)

        x = modulate(x, fshift, fscale)
        x = self.final_linear(x)
        x = self.unpatchify(x, h // self.patch_size, w // self.patch_size)
        return x



class RF(torch.nn.Module):
    def __init__(self, model, ln=True):
        super().__init__()
        self.model = model
        self.ln = ln
        self.stratified = False

    def forward(self, x, cond):

        b = x.size(0)
        if self.ln:
            if self.stratified:
                # stratified sampling of normals
                # first stratified sample from uniform
                quantiles = torch.linspace(0, 1, b + 1).to(x.device)
                z = quantiles[:-1] + torch.rand((b,)).to(x.device) / b
                # now transform to normal
                z = torch.erfinv(2 * z - 1) * math.sqrt(2)
                t = torch.sigmoid(z)
            else:
                nt = torch.randn((b,)).to(x.device)
                t = torch.sigmoid(nt)
        else:
            t = torch.rand((b,)).to(x.device)
        texp = t.view([b, *([1] * len(x.shape[1:]))])
        z1 = torch.randn_like(x)
        zt = (1 - texp) * x + texp * z1

        # make t, zt into same dtype as x
        zt, t = zt.to(x.dtype), t.to(x.dtype)
        vtheta = self.model(zt, t, cond)
        batchwise_mse = ((z1 - x - vtheta) ** 2).mean(dim=list(range(1, len(x.shape))))
        tlist = batchwise_mse.detach().cpu().reshape(-1).tolist()
        ttloss = [(tv, tloss) for tv, tloss in zip(t, tlist)]
        return batchwise_mse.mean(), {"batchwise_loss": ttloss}

    @torch.no_grad()
    def sample(self, z, cond, null_cond=None, sample_steps=50, cfg=2.0):
        b = z.size(0)
        dt = 1.0 / sample_steps
        dt = torch.tensor([dt] * b).to(z.device).view([b, *([1] * len(z.shape[1:]))])
        images = [z]
        for i in range(sample_steps, 0, -1):
            t = i / sample_steps
            t = torch.tensor([t] * b).to(z.device, dtype = z.dtype)

            vc = self.model(z, t, cond)
            if null_cond is not None:
                vu = self.model(z, t, null_cond)
                vc = vu + cfg * (vc - vu)

            z = (z - dt * vc).to(z.dtype)
            images.append(z)
        return images




#### Inference utils
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from diffusers.image_processor import VaeImageProcessor
from diffusers.models import AutoencoderKL
import torch

class RFPipeline:

    def __init__(self, ckpt_path = "/jfs/stage2_0623_stage1/model_24586/ema1.pt", device = "cuda:0"):
        self.device = device
        self.model = RF(
            MMDiT(
                in_channels=4,
                out_channels=4,
                dim=3072,
                global_conddim=3072,
                n_layers=36,
                n_heads=12,
                cond_seq_dim=2048,
                max_seq = 32 * 32,
            ),
            True,
        )
        self.model.load_state_dict(torch.load(ckpt_path, map_location='cpu'))
        self.model.eval().to(device).half()

        


        self.t5tokenizer = AutoTokenizer.from_pretrained("EleutherAI/pile-t5-xl", use_fast=True)
        self.t5tokenizer.pad_token = self.t5tokenizer.bos_token
        
        t5model = AutoModelForSeq2SeqLM.from_pretrained("EleutherAI/pile-t5-xl").half()
        
        self.t5model = t5model.to(device).eval()
        self.vae = AutoencoderKL.from_pretrained("stabilityai/sdxl-vae").to(device)



    @torch.no_grad()
    def _make_text_cond(self, cond_prompts, uncond_prompts = None):
        
        
        uncond_prompts = ["This is watermark, jpeg image white background, web image"] * len(cond_prompts)
        cond_prompts = [f"{prompt}" for prompt in cond_prompts]
            
        allprompts = cond_prompts + uncond_prompts
        #print(allprompts)
        t5_inputs = self.t5tokenizer(
            allprompts,
            truncation=True,
            max_length=256,
            padding="max_length",    
            return_tensors="pt",
        )
        t5_inputs = {k: v.to(self.device) for k, v in t5_inputs.items()}
        t5_outputs = self.t5model.encoder(**t5_inputs)[0] # B, T, D
        # mask that by 0 for padding tokens
        mask = t5_inputs["attention_mask"].unsqueeze(-1).expand(t5_outputs.shape)
        t5_outputs = t5_outputs * mask

        return {'c_seq': t5_outputs[:len(cond_prompts)]}, {'c_seq': t5_outputs[len(cond_prompts):]}

    @torch.no_grad()
    def __call__(self, prompt, num_inference_steps=50, height = 256, width = 256, guidance_scale = 8.5, num_images_per_prompt = 1, negative_prompt = None, seed = None):

        # make cond
        cond, null_cond = self._make_text_cond([prompt] * num_images_per_prompt, None)

        #print(cond, null_cond)
        L = num_images_per_prompt
        if seed is not None:
            gen = torch.Generator().manual_seed(1)
           
            init_noise = torch.randn(L, 4, 64, 64, generator = gen).to(self.device).half()
        else:
            init_noise = torch.randn(L, 4, 64, 64).to(self.device).half()

        images = self.model.sample(init_noise, cond, null_cond, num_inference_steps, guidance_scale)

        pil_images = []
        for i in range(L):
            x = self.vae.decode(images[-1][i : i + 1].to(self.device).float()/0.13025).sample
            img = VaeImageProcessor().postprocess(
                image=x.detach(), do_denormalize=[True, True]
            )[0]
            pil_images.append(img)

        return pil_images

    def to(self, device):
        self.device = device
        self.model.to(device)
        self.t5model.to(device)
        self.vae.to(device)
        return

if __name__ == "__main__":
    rf = RFPipeline("ema1.pt")
    images = rf("A painting of a beautiful sunset over the ocean.", num_images_per_prompt=1)
    images[0].show()
    images[0].save("output.png")