File size: 8,338 Bytes
fbaeba7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
"""Streaming one-step DiT for SwiftVR.

SwiftVR collapses iterative diffusion sampling to a single forward pass taken at the
fully-degraded endpoint of the flow (t = 1). No sampling scheduler is therefore
needed at inference time: the conditioning timestep is the constant
``INFERENCE_TIMESTEP`` below, equivalent to a flow-matching schedule evaluated by
``set_timesteps(1)`` (whose ``scale_model_input`` is the identity). Adjust it if
your training schedule uses a different number of train timesteps.
"""

from typing import Optional

import torch

from .chunk import ChunkSpec

INFERENCE_TIMESTEP = 1000.0
ROPE_EXTEND_MARGIN = 256


def _ensure_rope_cache_len(rope, required_len):
    cos, sin = rope.freqs_cos, rope.freqs_sin
    old_len = cos.shape[0]
    if required_len <= old_len:
        return
    if old_len < 2:
        raise RuntimeError(f"Cannot extend RoPE cache: length={old_len}")
    with torch.no_grad():
        dev, old_dtype = cos.device, cos.dtype
        cos64, sin64 = cos.to(torch.float64), sin.to(torch.float64)
        cos0, sin0, cos1, sin1 = cos64[0:1], sin64[0:1], cos64[1:2], sin64[1:2]
        cos_delta = cos1 * cos0 + sin1 * sin0
        sin_delta = sin1 * cos0 - cos1 * sin0
        angle0 = torch.atan2(sin0, cos0)
        step = torch.atan2(sin_delta, cos_delta)
        pos = torch.arange(old_len, required_len, device=dev, dtype=torch.float64).view(-1, 1)
        angle = angle0 + pos * step
        rope.freqs_cos = torch.cat([cos, torch.cos(angle).to(old_dtype)], 0).contiguous()
        rope.freqs_sin = torch.cat([sin, torch.sin(angle).to(old_dtype)], 0).contiguous()


def _rope_with_offset(rope, ppf, pph, ppw, t_off=0, h_off=0, w_off=0):
    required_len = max(t_off + ppf, h_off + pph, w_off + ppw)
    if required_len > rope.freqs_cos.shape[0]:
        required_len += max(0, int(ROPE_EXTEND_MARGIN))
    _ensure_rope_cache_len(rope, required_len)
    sp = [rope.t_dim, rope.h_dim, rope.w_dim]
    fc = rope.freqs_cos.split(sp, dim=1)
    fs = rope.freqs_sin.split(sp, dim=1)
    cf = fc[0][t_off:t_off + ppf].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
    ch = fc[1][h_off:h_off + pph].view(1, pph, 1, -1).expand(ppf, pph, ppw, -1)
    cw = fc[2][w_off:w_off + ppw].view(1, 1, ppw, -1).expand(ppf, pph, ppw, -1)
    sf = fs[0][t_off:t_off + ppf].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
    sh = fs[1][h_off:h_off + pph].view(1, pph, 1, -1).expand(ppf, pph, ppw, -1)
    sw = fs[2][w_off:w_off + ppw].view(1, 1, ppw, -1).expand(ppf, pph, ppw, -1)
    return (torch.cat([cf, ch, cw], -1).reshape(1, ppf * pph * ppw, 1, -1),
            torch.cat([sf, sh, sw], -1).reshape(1, ppf * pph * ppw, 1, -1))


def _precompute_cond(transformer, B, prompt_emb, timestep):
    pe = prompt_emb.clone()
    if pe.ndim == 2:
        pe = pe.unsqueeze(0).expand(B, -1, -1)
    elif pe.shape[0] != B:
        pe = pe.expand(B, -1, -1)
    ts = timestep.clone()
    ts_seq = None
    if ts.ndim == 2:
        ts_seq = ts.shape[1]
        ts = ts.flatten()
    temb, tp, enc_hs, enc_img = transformer.condition_embedder(ts, pe, None, timestep_seq_len=ts_seq)
    tp = tp.unflatten(2 if ts_seq else 1, (6, -1))
    if enc_img is not None:
        enc_hs = torch.cat([enc_img, enc_hs], dim=1)
    return temb, tp, enc_hs


@torch.inference_mode()
def _dit_forward_chunk(transformer, chunk, temb, tp, enc_hs, t_off=0):
    """One forward pass of the DiT, returning the predicted degradation velocity."""
    p_t, p_h, p_w = transformer.config.patch_size
    B, C, F, H, W = chunk.shape
    ppf, pph, ppw = F // p_t, H // p_h, W // p_w
    rope = _rope_with_offset(transformer.rope, ppf, pph, ppw, t_off)
    hs = transformer.patch_embedding(chunk).flatten(2).transpose(1, 2)

    thw_global = (ppf, pph, ppw)
    for i, blk in enumerate(transformer.blocks):
        underlying = getattr(blk, "_orig_mod", blk)
        if hasattr(underlying, "attn1"):
            underlying.attn1._thw = thw_global
            underlying.attn1._layer_id = i
    for blk in transformer.blocks:
        hs = blk(hs, enc_hs, tp, rope)

    if temb.ndim == 3:
        shift, scale = (transformer.scale_shift_table.unsqueeze(0).to(temb.device)
                        + temb.unsqueeze(2)).chunk(2, dim=2)
        shift, scale = shift.squeeze(2), scale.squeeze(2)
    else:
        shift, scale = (transformer.scale_shift_table.to(temb.device)
                        + temb.unsqueeze(1)).chunk(2, dim=1)
    hs = (transformer.norm_out(hs.float()) * (1 + scale.to(hs.device)) + shift.to(hs.device)).type_as(hs)
    hs = transformer.proj_out(hs)
    hs = hs.reshape(B, ppf, pph, ppw, p_t, p_h, p_w, -1).permute(0, 7, 1, 4, 2, 5, 3, 6)
    return hs.flatten(6, 7).flatten(4, 5).flatten(2, 3)


class StreamingDiT:
    """One-step DiT with temporal overlap blending across chunks."""

    def __init__(self, transformer, overlap=0):
        self.transformer = transformer
        self.overlap = overlap
        self._prev_lq = self._prev_out = None
        self._g_off = 0
        self._cond_cache_key = self._cond_cache = None

    def reset(self):
        self._prev_lq = self._prev_out = None
        self._g_off = 0

    def _get_cached_condition(self, B, prompt_emb, dev, dt):
        cache_key = (int(B), dev.type, dev.index, str(dt), tuple(prompt_emb.shape))
        if self._cond_cache_key != cache_key or self._cond_cache is None:
            ts = torch.full((B,), INFERENCE_TIMESTEP, device=dev, dtype=torch.float32)
            self._cond_cache = _precompute_cond(self.transformer, B, prompt_emb.to(dev, dt), ts)
            self._cond_cache_key = cache_key
        return self._cond_cache

    @torch.inference_mode()
    def denoise(self, lq, prompt_emb):
        """Restore one chunk of latents ``lq`` (shape [B, C, F, H, W])."""
        dev, dt = lq.device, lq.dtype
        B, C, F_cur, H, W = lq.shape

        ol = 0
        if self._prev_lq is not None and self.overlap > 0:
            ol = self._prev_lq.shape[2]
            lq_ext = torch.cat([self._prev_lq.to(dev), lq], dim=2)
            t_rope = self._g_off - ol
        else:
            lq_ext = lq
            t_rope = self._g_off

        temb, tp, enc_hs = self._get_cached_condition(B, prompt_emb, dev, dt)
        pred = _dit_forward_chunk(self.transformer, lq_ext, temb, tp, enc_hs, t_off=t_rope)
        den_ext = lq_ext - pred

        if ol > 0 and self._prev_out is not None:
            ramp = torch.linspace(0, 1, ol, device=dev, dtype=dt).view(1, 1, ol, 1, 1)
            den_ext[:, :, :ol] = self._prev_out.to(dev) * (1 - ramp) + den_ext[:, :, :ol] * ramp
            den_out = den_ext[:, :, ol:]
        else:
            den_out = den_ext

        k = min(self.overlap, F_cur)
        if k > 0:
            self._prev_lq = lq[:, :, -k:].detach().cpu().clone()
            self._prev_out = den_out[:, :, -k:].detach().cpu().clone()
        else:
            self._prev_lq = self._prev_out = None

        self._g_off += F_cur
        return den_out

    @torch.inference_mode()
    def denoise_last_chunk(self, z_new_ntchw, spec: ChunkSpec, prompt_emb,
                           prev_dit_out_cpu: Optional[torch.Tensor], n_lat: int, device, dtype):
        """LAST chunk: pad the (b+1) new latents up to (n_lat+1) with the previous
        chunk's latents (or zeros) for a correct RoPE offset, run one pass, and
        keep only the new ``b+1`` denoised latents.
        """
        lat_count = spec.b + 1
        pad_count = (n_lat + 1) - lat_count

        z_bcfhw = z_new_ntchw.permute(0, 2, 1, 3, 4).contiguous()
        if pad_count > 0:
            if prev_dit_out_cpu is not None:
                pad_z = prev_dit_out_cpu[:, :, -pad_count:].to(device=device, dtype=dtype)
            else:
                pad_z = torch.zeros(z_bcfhw.shape[0], z_bcfhw.shape[1], pad_count,
                                    z_bcfhw.shape[3], z_bcfhw.shape[4], device=device, dtype=dtype)
            z_bcfhw = torch.cat([pad_z, z_bcfhw], dim=2)

        t_off = max(0, self._g_off - pad_count)
        temb, tp, enc_hs = self._get_cached_condition(z_bcfhw.shape[0], prompt_emb, device, dtype)
        pred = _dit_forward_chunk(self.transformer, z_bcfhw, temb, tp, enc_hs, t_off=t_off)
        z_den = (z_bcfhw - pred)[:, :, -lat_count:].contiguous()

        self._g_off += lat_count
        return z_den.permute(0, 2, 1, 3, 4).contiguous()