sd_hijack.py

import math
import os
import sys
import traceback
import torch
import numpy as np
from torch import einsum

from modules.shared import opts, device, cmd_opts

from ldm.util import default
from einops import rearrange
import ldm.modules.attention



# see https://github.com/basujindal/stable-diffusion/pull/117 for discussion
def split_cross_attention_forward_v1(self, x, context=None, mask=None):
    h = self.heads

    q = self.to_q(x)
    context = default(context, x)
    k = self.to_k(context)
    v = self.to_v(context)
    del context, x

    q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))

    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

        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

    r2 = rearrange(r1, '(b h) n d -> b n (h d)', h=h)
    del r1

    return self.to_out(r2)


# taken from https://github.com/Doggettx/stable-diffusion
def split_cross_attention_forward(self, x, context=None, mask=None):
    h = self.heads

    q_in = self.to_q(x)
    context = default(context, x)
    k_in = self.to_k(context)
    v_in = self.to_v(context)
    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)

    stats = torch.cuda.memory_stats(q.device)
    mem_active = stats['active_bytes.all.current']
    mem_reserved = stats['reserved_bytes.all.current']
    mem_free_cuda, _ = torch.cuda.mem_get_info(torch.cuda.current_device())
    mem_free_torch = mem_reserved - mem_active
    mem_free_total = mem_free_cuda + mem_free_torch

    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) * self.scale

        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

    r2 = rearrange(r1, '(b h) n d -> b n (h d)', h=h)
    del r1

    return self.to_out(r2)


class StableDiffusionModelHijack:
    ids_lookup = {}
    word_embeddings = {}
    word_embeddings_checksums = {}
    fixes = None
    comments = []
    dir_mtime = None
    layers = None
    circular_enabled = False

    def load_textual_inversion_embeddings(self, dirname, model):
        mt = os.path.getmtime(dirname)
        if self.dir_mtime is not None and mt <= self.dir_mtime:
            return

        self.dir_mtime = mt
        self.ids_lookup.clear()
        self.word_embeddings.clear()

        tokenizer = model.cond_stage_model.tokenizer

        def const_hash(a):
            r = 0
            for v in a:
                r = (r * 281 ^ int(v) * 997) & 0xFFFFFFFF
            return r

        def process_file(path, filename):
            name = os.path.splitext(filename)[0]

            data = torch.load(path)

            # textual inversion embeddings
            if 'string_to_param' in data:
                param_dict = data['string_to_param']
                if hasattr(param_dict, '_parameters'):
                    param_dict = getattr(param_dict, '_parameters')  # fix for torch 1.12.1 loading saved file from torch 1.11
                assert len(param_dict) == 1, 'embedding file has multiple terms in it'
                emb = next(iter(param_dict.items()))[1]
            elif type(data) == dict and type(next(iter(data.values()))) == torch.Tensor:
                assert len(data.keys()) == 1, 'embedding file has multiple terms in it'

                emb = next(iter(data.values()))
                if len(emb.shape) == 1:
                    emb = emb.unsqueeze(0)

            self.word_embeddings[name] = emb.detach()
            self.word_embeddings_checksums[name] = f'{const_hash(emb.reshape(-1))&0xffff:04x}'

            ids = tokenizer([name], add_special_tokens=False)['input_ids'][0]

            first_id = ids[0]
            if first_id not in self.ids_lookup:
                self.ids_lookup[first_id] = []
            self.ids_lookup[first_id].append((ids, name))

        for fn in os.listdir(dirname):
            try:
                process_file(os.path.join(dirname, fn), fn)
            except Exception:
                print(f"Error loading emedding {fn}:", file=sys.stderr)
                print(traceback.format_exc(), file=sys.stderr)
                continue

        print(f"Loaded a total of {len(self.word_embeddings)} text inversion embeddings.")

    def hijack(self, m):
        model_embeddings = m.cond_stage_model.transformer.text_model.embeddings

        model_embeddings.token_embedding = EmbeddingsWithFixes(model_embeddings.token_embedding, self)
        m.cond_stage_model = FrozenCLIPEmbedderWithCustomWords(m.cond_stage_model, self)

        if cmd_opts.opt_split_attention:
            ldm.modules.attention.CrossAttention.forward = split_cross_attention_forward
        elif cmd_opts.opt_split_attention_v1:
            ldm.modules.attention.CrossAttention.forward = split_cross_attention_forward_v1

        def flatten(el):
            flattened = [flatten(children) for children in el.children()]
            res = [el]
            for c in flattened:
                res += c
            return res

        self.layers = flatten(m)

    def apply_circular(self, enable):
        if self.circular_enabled == enable:
            return

        self.circular_enabled = enable

        for layer in [layer for layer in self.layers if type(layer) == torch.nn.Conv2d]:
            layer.padding_mode = 'circular' if enable else 'zeros'


class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
    def __init__(self, wrapped, hijack):
        super().__init__()
        self.wrapped = wrapped
        self.hijack = hijack
        self.tokenizer = wrapped.tokenizer
        self.max_length = wrapped.max_length
        self.token_mults = {}

        tokens_with_parens = [(k, v) for k, v in self.tokenizer.get_vocab().items() if '(' in k or ')' in k or '[' in k or ']' in k]
        for text, ident in tokens_with_parens:
            mult = 1.0
            for c in text:
                if c == '[':
                    mult /= 1.1
                if c == ']':
                    mult *= 1.1
                if c == '(':
                    mult *= 1.1
                if c == ')':
                    mult /= 1.1

            if mult != 1.0:
                self.token_mults[ident] = mult

    def forward(self, text):
        self.hijack.fixes = []
        self.hijack.comments = []
        remade_batch_tokens = []
        id_start = self.wrapped.tokenizer.bos_token_id
        id_end = self.wrapped.tokenizer.eos_token_id
        maxlen = self.wrapped.max_length - 2
        used_custom_terms = []

        cache = {}
        batch_tokens = self.wrapped.tokenizer(text, truncation=False, add_special_tokens=False)["input_ids"]
        batch_multipliers = []
        for tokens in batch_tokens:
            tuple_tokens = tuple(tokens)

            if tuple_tokens in cache:
                remade_tokens, fixes, multipliers = cache[tuple_tokens]
            else:
                fixes = []
                remade_tokens = []
                multipliers = []
                mult = 1.0

                i = 0
                while i < len(tokens):
                    token = tokens[i]

                    possible_matches = self.hijack.ids_lookup.get(token, None)

                    mult_change = self.token_mults.get(token) if opts.enable_emphasis else None
                    if mult_change is not None:
                        mult *= mult_change
                    elif possible_matches is None:
                        remade_tokens.append(token)
                        multipliers.append(mult)
                    else:
                        found = False
                        for ids, word in possible_matches:
                            if tokens[i:i+len(ids)] == ids:
                                emb_len = int(self.hijack.word_embeddings[word].shape[0])
                                fixes.append((len(remade_tokens), word))
                                remade_tokens += [0] * emb_len
                                multipliers += [mult] * emb_len
                                i += len(ids) - 1
                                found = True
                                used_custom_terms.append((word, self.hijack.word_embeddings_checksums[word]))
                                break

                        if not found:
                            remade_tokens.append(token)
                            multipliers.append(mult)

                    i += 1

                if len(remade_tokens) > maxlen - 2:
                    vocab = {v: k for k, v in self.wrapped.tokenizer.get_vocab().items()}
                    ovf = remade_tokens[maxlen - 2:]
                    overflowing_words = [vocab.get(int(x), "") for x in ovf]
                    overflowing_text = self.wrapped.tokenizer.convert_tokens_to_string(''.join(overflowing_words))

                    self.hijack.comments.append(f"Warning: too many input tokens; some ({len(overflowing_words)}) have been truncated:\n{overflowing_text}\n")

                remade_tokens = remade_tokens + [id_end] * (maxlen - 2 - len(remade_tokens))
                remade_tokens = [id_start] + remade_tokens[0:maxlen-2] + [id_end]
                cache[tuple_tokens] = (remade_tokens, fixes, multipliers)

            multipliers = multipliers + [1.0] * (maxlen - 2 - len(multipliers))
            multipliers = [1.0] + multipliers[0:maxlen - 2] + [1.0]

            remade_batch_tokens.append(remade_tokens)
            self.hijack.fixes.append(fixes)
            batch_multipliers.append(multipliers)

        if len(used_custom_terms) > 0:
            self.hijack.comments.append("Used custom terms: " + ", ".join([f'{word} [{checksum}]' for word, checksum in used_custom_terms]))

        tokens = torch.asarray(remade_batch_tokens).to(device)
        outputs = self.wrapped.transformer(input_ids=tokens)
        z = outputs.last_hidden_state

        # restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise
        batch_multipliers = torch.asarray(batch_multipliers).to(device)
        original_mean = z.mean()
        z *= batch_multipliers.reshape(batch_multipliers.shape + (1,)).expand(z.shape)
        new_mean = z.mean()
        z *= original_mean / new_mean

        return z


class EmbeddingsWithFixes(torch.nn.Module):
    def __init__(self, wrapped, embeddings):
        super().__init__()
        self.wrapped = wrapped
        self.embeddings = embeddings

    def forward(self, input_ids):
        batch_fixes = self.embeddings.fixes
        self.embeddings.fixes = None

        inputs_embeds = self.wrapped(input_ids)

        if batch_fixes is not None:
            for fixes, tensor in zip(batch_fixes, inputs_embeds):
                for offset, word in fixes:
                    emb = self.embeddings.word_embeddings[word]
                    emb_len = min(tensor.shape[0]-offset, emb.shape[0])
                    tensor[offset:offset+emb_len] = self.embeddings.word_embeddings[word][0:emb_len]

        return inputs_embeds


def add_circular_option_to_conv_2d():
    conv2d_constructor = torch.nn.Conv2d.__init__

    def conv2d_constructor_circular(self, *args, **kwargs):
        return conv2d_constructor(self, *args, padding_mode='circular', **kwargs)

    torch.nn.Conv2d.__init__ = conv2d_constructor_circular


model_hijack = StableDiffusionModelHijack()