Ctrl + C ; Ctrl + V

extensions-builtin/PiD/nn/model.py

@@ -0,0 +1,239 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+import comfy.ldm.common_dit
+import comfy.patcher_extension
+from comfy.ldm.flux.math import apply_rope, rope
+from comfy.ldm.hidream.model import FeedForwardSwiGLU
+from comfy.ldm.modules.attention import optimized_attention
+from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder
+
+from .modules import (
+    FinalLayer,
+    PatchTokenEmbedder,
+    PiTBlock,
+    PixelTokenEmbedder,
+    apply_adaln_,
+    precompute_freqs_cis_2d,
+)
+
+
+class MMDiTJointAttention(nn.Module):
+    """Joint MMDiT attention with separate Q/K/V/proj for image and text streams.
+
+    RoPE is applied to each stream before concatenation so each stream uses its own
+    2D/1D positional encoding. Concat order is [text, image] (text first).
+    """
+    def __init__(self, dim, num_heads=8, qkv_bias=False, dtype=None, device=None, operations=None):
+        super().__init__()
+        assert dim % num_heads == 0
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+
+        self.qkv_x = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
+        self.qkv_y = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
+
+        self.q_norm_x = operations.RMSNorm(self.head_dim, eps=1e-6, dtype=dtype, device=device)
+        self.k_norm_x = operations.RMSNorm(self.head_dim, eps=1e-6, dtype=dtype, device=device)
+        self.q_norm_y = operations.RMSNorm(self.head_dim, eps=1e-6, dtype=dtype, device=device)
+        self.k_norm_y = operations.RMSNorm(self.head_dim, eps=1e-6, dtype=dtype, device=device)
+
+        self.proj_x = operations.Linear(dim, dim, dtype=dtype, device=device)
+        self.proj_y = operations.Linear(dim, dim, dtype=dtype, device=device)
+
+    def forward(self, x, y, pos_img, pos_txt=None, attn_mask=None, transformer_options={}):
+        B, Nx, _ = x.shape
+        _, Ny, _ = y.shape
+        H = self.num_heads
+        D = self.head_dim
+
+        qkv_x = self.qkv_x(x).reshape(B, Nx, 3, H, D).permute(2, 0, 3, 1, 4)
+        qx, kx, vx = qkv_x.unbind(0)
+        qx = self.q_norm_x(qx)
+        kx = self.k_norm_x(kx)
+
+        qkv_y = self.qkv_y(y).reshape(B, Ny, 3, H, D).permute(2, 0, 3, 1, 4)
+        qy, ky, vy = qkv_y.unbind(0)
+        qy = self.q_norm_y(qy)
+        ky = self.k_norm_y(ky)
+
+        qx, kx = apply_rope(qx, kx, pos_img[None, None])
+        if pos_txt is not None:
+            qy, ky = apply_rope(qy, ky, pos_txt[None, None])
+
+        q_joint = torch.cat([qy, qx], dim=2)
+        k_joint = torch.cat([ky, kx], dim=2)
+        v_joint = torch.cat([vy, vx], dim=2)
+
+        out_joint = optimized_attention(
+            q_joint, k_joint, v_joint, H,
+            mask=attn_mask, skip_reshape=True, skip_output_reshape=True,
+            transformer_options=transformer_options,
+        )
+
+        out_y = out_joint[:, :, :Ny, :].transpose(1, 2).reshape(B, Ny, H * D)
+        out_x = out_joint[:, :, Ny:, :].transpose(1, 2).reshape(B, Nx, H * D)
+
+        return self.proj_x(out_x), self.proj_y(out_y)
+
+
+class MMDiTBlockT2I(nn.Module):
+    def __init__(self, hidden_size, groups, mlp_ratio=4.0, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm_x1 = operations.RMSNorm(hidden_size, eps=1e-6, dtype=dtype, device=device)
+        self.norm_y1 = operations.RMSNorm(hidden_size, eps=1e-6, dtype=dtype, device=device)
+        self.attn = MMDiTJointAttention(hidden_size, num_heads=groups, qkv_bias=False, dtype=dtype, device=device, operations=operations)
+        self.norm_x2 = operations.RMSNorm(hidden_size, eps=1e-6, dtype=dtype, device=device)
+        self.norm_y2 = operations.RMSNorm(hidden_size, eps=1e-6, dtype=dtype, device=device)
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        self.mlp_x = FeedForwardSwiGLU(hidden_size, mlp_hidden_dim, multiple_of=1, dtype=dtype, device=device, operations=operations)
+        self.mlp_y = FeedForwardSwiGLU(hidden_size, mlp_hidden_dim, multiple_of=1, dtype=dtype, device=device, operations=operations)
+        self.adaLN_modulation_img = nn.Sequential(operations.Linear(hidden_size, 6 * hidden_size, bias=True, dtype=dtype, device=device))
+        self.adaLN_modulation_txt = nn.Sequential(operations.Linear(hidden_size, 6 * hidden_size, bias=True, dtype=dtype, device=device))
+
+    def forward(self, x, y, c, pos_img, pos_txt=None, attn_mask=None, transformer_options={}):
+        shift_msa_x, scale_msa_x, gate_msa_x, shift_mlp_x, scale_mlp_x, gate_mlp_x = self.adaLN_modulation_img(c).chunk(6, dim=-1)
+        shift_msa_y, scale_msa_y, gate_msa_y, shift_mlp_y, scale_mlp_y, gate_mlp_y = self.adaLN_modulation_txt(c).chunk(6, dim=-1)
+
+        x_norm = apply_adaln_(self.norm_x1(x), shift_msa_x, scale_msa_x)
+        y_norm = apply_adaln_(self.norm_y1(y), shift_msa_y, scale_msa_y)
+        attn_x, attn_y = self.attn(x_norm, y_norm, pos_img, pos_txt, attn_mask, transformer_options=transformer_options)
+        x = torch.addcmul(x, gate_msa_x, attn_x)
+        y = torch.addcmul(y, gate_msa_y, attn_y)
+
+        x = torch.addcmul(x, gate_mlp_x, self.mlp_x(apply_adaln_(self.norm_x2(x), shift_mlp_x, scale_mlp_x)))
+        y = torch.addcmul(y, gate_mlp_y, self.mlp_y(apply_adaln_(self.norm_y2(y), shift_mlp_y, scale_mlp_y)))
+        return x, y
+
+
+class PixDiT_T2I(nn.Module):
+    """PixelDiT T2I model. Hardcoded for the released 1024px Stage-3 checkpoint
+    (also runs at 512px when fed the appropriate latent size and flow_shift).
+
+    Forward:
+      x:        [B, 3, H, W] pixel-space input (no VAE)
+      timesteps:[B] in [0, 1000] (ComfyUI flow sampling convention)
+      context:  [B, Ltxt, 2304] Gemma-2-2b-it hidden states (chi_prompt prepended)
+    Returns flow-matching velocity [B, 3, H, W].
+    """
+    def __init__(
+        self,
+        in_channels=3,
+        num_groups=24,
+        hidden_size=1536,
+        pixel_hidden_size=16,
+        pixel_attn_hidden_size=1152,
+        pixel_num_groups=16,
+        patch_depth=14,
+        pixel_depth=2,
+        patch_size=16,
+        txt_embed_dim=2304,
+        txt_max_length=300,
+        use_text_rope=True,
+        text_rope_theta=10000.0,
+        image_model=None,
+        dtype=None,
+        device=None,
+        operations=None,
+        pixel_mlp_chunks=2,
+    ):
+        super().__init__()
+        self.dtype = dtype
+        self.in_channels = in_channels
+        self.out_channels = in_channels
+        self.hidden_size = hidden_size
+        self.num_groups = num_groups
+        self.patch_depth = patch_depth
+        self.pixel_depth = pixel_depth
+        self.patch_size = patch_size
+        self.pixel_hidden_size = pixel_hidden_size
+        self.pixel_attn_hidden_size = pixel_attn_hidden_size
+        self.pixel_num_groups = pixel_num_groups
+        self.txt_embed_dim = txt_embed_dim
+        self.txt_max_length = txt_max_length
+        self.use_text_rope = use_text_rope
+        self.text_rope_theta = text_rope_theta
+
+        self.pixel_embedder = PixelTokenEmbedder(self.in_channels, self.pixel_hidden_size, dtype=dtype, device=device, operations=operations)
+        self.s_embedder = PatchTokenEmbedder(self.in_channels * self.patch_size ** 2, self.hidden_size, bias=True, dtype=dtype, device=device, operations=operations)
+        self.t_embedder = TimestepEmbedder(self.hidden_size, dtype=dtype, device=device, operations=operations, max_period=10)
+        self.y_embedder = PatchTokenEmbedder(self.txt_embed_dim, self.hidden_size, bias=True, use_norm=True, dtype=dtype, device=device, operations=operations)
+        self.y_pos_embedding = nn.Parameter(torch.empty(1, self.txt_max_length, self.hidden_size, dtype=dtype, device=device))
+
+        self.patch_blocks = nn.ModuleList([
+            MMDiTBlockT2I(self.hidden_size, self.num_groups,
+                          dtype=dtype, device=device, operations=operations)
+            for _ in range(self.patch_depth)
+        ])
+        self.pixel_blocks = nn.ModuleList([
+            PiTBlock(
+                self.pixel_hidden_size,
+                self.hidden_size,
+                patch_size=self.patch_size,
+                num_heads=self.num_groups,
+                attn_hidden_size=self.pixel_attn_hidden_size,
+                attn_num_heads=self.pixel_num_groups,
+                dtype=dtype, device=device, operations=operations,
+                mlp_chunks=pixel_mlp_chunks,
+            )
+            for _ in range(self.pixel_depth)
+        ])
+
+        self.final_layer = FinalLayer(self.pixel_hidden_size, self.out_channels, dtype=dtype, device=device, operations=operations)
+
+    def _fetch_patch_pos(self, height, width, device, dtype, **rope_opts):
+        return precompute_freqs_cis_2d(self.hidden_size // self.num_groups, height, width, device=device, dtype=dtype, **rope_opts)
+
+    def _fetch_text_pos(self, length, device, dtype):
+        return rope(torch.arange(length, dtype=torch.float32, device=device).reshape(1, -1), self.hidden_size // self.num_groups, self.text_rope_theta).squeeze(0).to(dtype=dtype)
+
+    def forward(self, x, timesteps, context=None, attention_mask=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward, self, comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options),
+        ).execute(x, timesteps, context, attention_mask, transformer_options, **kwargs)
+
+    def _pre_patch_block(self, s, i, **kwargs):
+        """Hook for subclasses to inject per-block state into the patch stream (e.g. PiD's LQ gate)."""
+        return s
+
+    def _forward(self, x, timesteps, context=None, attention_mask=None, transformer_options={}, **kwargs):
+        H_orig, W_orig = x.shape[2], x.shape[3]
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
+        B, _, H, W = x.shape
+        Hs = H // self.patch_size
+        Ws = W // self.patch_size
+        L = Hs * Ws
+
+        pos_img = self._fetch_patch_pos(Hs, Ws, x.device, x.dtype, **(transformer_options.get("rope_options") or {}))
+        x_patches = F.unfold(x, kernel_size=self.patch_size, stride=self.patch_size).transpose(1, 2)
+
+        t_emb = self.t_embedder(timesteps.view(-1), x.dtype).view(B, -1, self.hidden_size)
+
+        if context is None or context.dim() != 3:
+            raise ValueError("PixDiT_T2I requires context (text embeddings) of shape [B, L, D]")
+        Ltxt = min(context.shape[1], self.txt_max_length)
+        y = context[:, :Ltxt, :]
+        y_emb = self.y_embedder(y).view(B, Ltxt, self.hidden_size)
+        y_emb = y_emb + self.y_pos_embedding[:, :Ltxt, :].to(y_emb) # y_pos_embedding is a raw nn.Parameter
+
+        condition = F.silu(t_emb)
+        pos_txt = self._fetch_text_pos(Ltxt, x.device, x.dtype) if self.use_text_rope else None
+
+        s = self.s_embedder(x_patches)
+        for i, blk in enumerate(self.patch_blocks):
+            s = self._pre_patch_block(s, i, **kwargs)
+            s, y_emb = blk(s, y_emb, condition, pos_img, pos_txt, None, transformer_options=transformer_options)
+        s = F.silu(t_emb + s)
+
+        s_cond = s.view(B * L, self.hidden_size)
+        x_pixels = self.pixel_embedder(x, patch_size=self.patch_size)
+        for blk in self.pixel_blocks:
+            x_pixels = blk(x_pixels, s_cond, H, W, self.patch_size, mask=None, transformer_options=transformer_options)
+
+        x_pixels = self.final_layer(x_pixels)
+        C_out = self.out_channels
+        P2 = self.patch_size * self.patch_size
+        x_pixels = x_pixels.view(B, L, P2, C_out).permute(0, 3, 2, 1).reshape(B, C_out * P2, L)
+        out = F.fold(x_pixels, (H, W), kernel_size=self.patch_size, stride=self.patch_size)
+        return out[:, :, :H_orig, :W_orig]

extensions-builtin/PiD/nn/modules.py

@@ -0,0 +1,187 @@
+import torch
+import torch.nn as nn
+
+from comfy.ldm.flux.math import apply_rope, rope
+from comfy.ldm.modules.attention import optimized_attention
+from comfy.ldm.modules.diffusionmodules.mmdit import Mlp, get_1d_sincos_pos_embed_from_grid_torch
+
+
+def apply_adaln_(x, shift, scale):
+    return x.addcmul_(x, scale).add_(shift)
+
+
+def precompute_freqs_cis_2d(dim, height, width, theta=10000.0, scale=16.0,
+                            ref_grid_h=None, ref_grid_w=None,
+                            scale_x=1.0, scale_y=1.0, shift_x=0.0, shift_y=0.0,
+                            device=None, dtype=torch.float32, **kwargs):
+    """2D RoPE with x/y axis frequencies interleaved at stride 2 across head dim.
+
+    rope_options:
+      scale_x / scale_y multiply the position range (RoPE extrapolation).
+      shift_x / shift_y offset the position origin (tiled / regional inference).
+    With ref_grid_h/w set, also applies NTK-aware per-axis theta scaling
+    (rope_mode='ntk_aware'): theta_axis = theta * (current/ref)^(dim_axis/(dim_axis-2)).
+    Returns Flux-format rotation matrices of shape [H*W, dim/2, 2, 2].
+    Layout of head-dim pairs: [x_0, y_0, x_1, y_1, ..., x_{dim/4-1}, y_{dim/4-1}].
+    """
+    dim_axis = dim // 2
+    if ref_grid_h is not None and dim_axis > 2:
+        h_ntk = (height / ref_grid_h) ** (dim_axis / (dim_axis - 2))
+        w_ntk = (width / ref_grid_w) ** (dim_axis / (dim_axis - 2))
+    else:
+        h_ntk = w_ntk = 1.0
+
+    x_lin = torch.linspace(shift_x, scale * scale_x + shift_x, width, device=device)
+    y_lin = torch.linspace(shift_y, scale * scale_y + shift_y, height, device=device)
+    y_grid, x_grid = torch.meshgrid(y_lin, x_lin, indexing="ij")
+    x_rope = rope(x_grid.reshape(1, -1), dim_axis, theta * w_ntk).squeeze(0)
+    y_rope = rope(y_grid.reshape(1, -1), dim_axis, theta * h_ntk).squeeze(0)
+    out = torch.stack([x_rope, y_rope], dim=2).reshape(height * width, dim // 2, 2, 2)
+    return out.to(dtype=dtype)
+
+
+def get_2d_sincos_pos_embed(embed_dim, height, width, device=None, dtype=torch.float32):
+    """Standard 2D sin/cos absolute positional embedding (ViT-style).
+
+    first half encodes W-coordinates, second half H.
+    """
+    assert embed_dim % 4 == 0
+    grid_h = torch.arange(height, dtype=torch.float32, device=device)
+    grid_w = torch.arange(width, dtype=torch.float32, device=device)
+    grid_y, grid_x = torch.meshgrid(grid_h, grid_w, indexing="ij")
+    emb_w = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_x.reshape(-1), device=device)
+    emb_h = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_y.reshape(-1), device=device)
+    return torch.cat([emb_w, emb_h], dim=1).to(dtype=dtype)
+
+
+class RotaryAttention(nn.Module):
+    """Single-stream self-attention with rotary positional encoding (used inside PiTBlock)."""
+    def __init__(self, dim, num_heads=8, qkv_bias=False, dtype=None, device=None, operations=None):
+        super().__init__()
+        assert dim % num_heads == 0
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
+        self.q_norm = operations.RMSNorm(self.head_dim, eps=1e-6, dtype=dtype, device=device)
+        self.k_norm = operations.RMSNorm(self.head_dim, eps=1e-6, dtype=dtype, device=device)
+        self.proj = operations.Linear(dim, dim, dtype=dtype, device=device)
+
+    def forward(self, x, pos, mask=None, transformer_options={}):
+        B, N, C = x.shape
+        H = self.num_heads
+        D = self.head_dim
+        qkv = self.qkv(x).reshape(B, N, 3, H, D).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv.unbind(0)
+        q, k = apply_rope(self.q_norm(q), self.k_norm(k), pos[None, None])
+        x = optimized_attention(q, k, v, H, mask=mask, skip_reshape=True, transformer_options=transformer_options)
+        return self.proj(x)
+
+
+class FinalLayer(nn.Module):
+    def __init__(self, hidden_size, out_channels, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm = operations.RMSNorm(hidden_size, eps=1e-6, dtype=dtype, device=device)
+        self.linear = operations.Linear(hidden_size, out_channels, bias=True, dtype=dtype, device=device)
+
+    def forward(self, x):
+        return self.linear(self.norm(x))
+
+
+class PatchTokenEmbedder(nn.Module):
+    """Linear projection used both for patchified-image tokens and text-feature tokens."""
+    def __init__(self, in_chans, embed_dim, use_norm=False, bias=True, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.proj = operations.Linear(in_chans, embed_dim, bias=bias, dtype=dtype, device=device)
+        self.norm = operations.RMSNorm(embed_dim, eps=1e-6, dtype=dtype, device=device) if use_norm else nn.Identity()
+
+    def forward(self, x):
+        return self.norm(self.proj(x))
+
+
+class PixelTokenEmbedder(nn.Module):
+    """Pixel-level embedder: lifts each RGB pixel to hidden_size and packs into per-patch sequences."""
+    def __init__(self, in_channels, hidden_size_output, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.in_channels = in_channels
+        self.hidden_size_output = hidden_size_output
+        self.proj = operations.Linear(self.in_channels, self.hidden_size_output, bias=True, dtype=dtype, device=device)
+
+    def forward(self, inputs, patch_size):
+        B, _, H, W = inputs.shape
+        Hs, Ws = H // patch_size, W // patch_size
+        P2 = patch_size * patch_size
+        x = inputs.permute(0, 2, 3, 1).contiguous()
+        x = self.proj(x)
+        pos_full = get_2d_sincos_pos_embed(self.hidden_size_output, H, W, device=x.device, dtype=x.dtype).view(H, W, self.hidden_size_output)
+        x = x + pos_full.unsqueeze(0)
+        x = x.view(B, Hs, patch_size, Ws, patch_size, self.hidden_size_output)
+        return x.permute(0, 1, 3, 2, 4, 5).reshape(B * Hs * Ws, P2, self.hidden_size_output)
+
+
+class PiTBlock(nn.Module):
+    """Pixel-level transformer block.
+
+    Compresses each patch's P^2 pixel tokens → 1 attention token via a linear,
+    runs global self-attention across patches with 2D RoPE, then expands back to P^2 tokens.
+    Conditioning is per-pixel adaLN from the patch-level features.
+    """
+    def __init__(self, pixel_hidden_size, patch_hidden_size, patch_size, num_heads, mlp_ratio=4.0,
+                 attn_hidden_size=None, attn_num_heads=None, dtype=None, device=None, operations=None, mlp_chunks=1):
+        super().__init__()
+        self.pixel_dim = pixel_hidden_size
+        self.context_dim = patch_hidden_size
+        self.attn_dim = attn_hidden_size if attn_hidden_size is not None else patch_hidden_size
+        self.num_heads = attn_num_heads if attn_num_heads is not None else num_heads
+        assert self.attn_dim % self.num_heads == 0
+
+        p2 = patch_size * patch_size
+        self.compress_to_attn = operations.Linear(p2 * self.pixel_dim, self.attn_dim, bias=True, dtype=dtype, device=device)
+        self.expand_from_attn = operations.Linear(self.attn_dim, p2 * self.pixel_dim, bias=True, dtype=dtype, device=device)
+
+        self.norm1 = operations.RMSNorm(self.pixel_dim, eps=1e-6, dtype=dtype, device=device)
+        self.attn = RotaryAttention(self.attn_dim, num_heads=self.num_heads, qkv_bias=False, dtype=dtype, device=device, operations=operations)
+        self.norm2 = operations.RMSNorm(self.pixel_dim, eps=1e-6, dtype=dtype, device=device)
+        self.mlp = Mlp(self.pixel_dim, hidden_features=int(self.pixel_dim * mlp_ratio), dtype=dtype, device=device, operations=operations)
+
+        self.adaLN_modulation_msa = operations.Linear(self.context_dim, 3 * self.pixel_dim * p2, bias=True, dtype=dtype, device=device)
+        self.adaLN_modulation_mlp = operations.Linear(self.context_dim, 3 * self.pixel_dim * p2, bias=True, dtype=dtype, device=device)
+
+        self._rope_fn = precompute_freqs_cis_2d
+        self.mlp_chunks = max(1, int(mlp_chunks))
+
+    def _fetch_pos(self, height, width, device, dtype, **rope_opts):
+        return self._rope_fn(self.attn_dim // self.num_heads, height, width, device=device, dtype=dtype, **rope_opts)
+
+    def forward(self, x, s_cond, image_height, image_width, patch_size, mask=None, transformer_options={}):
+        BL, P2, _ = x.shape
+        Hs, Ws = image_height // patch_size, image_width // patch_size
+        L = Hs * Ws
+        B = BL // L
+
+        # Attention path uses only msa params; compute, use, free before mlp params allocate.
+        msa_params = self.adaLN_modulation_msa(s_cond).view(BL, P2, 3 * self.pixel_dim)
+        shift_msa, scale_msa, gate_msa = msa_params.chunk(3, dim=-1)
+
+        x_norm = apply_adaln_(self.norm1(x), shift_msa, scale_msa)
+        x_flat = x_norm.view(BL, P2 * self.pixel_dim)
+
+        x_comp = self.compress_to_attn(x_flat).view(B, L, self.attn_dim)
+        pos_comp = self._fetch_pos(Hs, Ws, x.device, x.dtype, **(transformer_options.get("rope_options") or {}))
+        attn_out = self.attn(x_comp, pos_comp, mask=mask, transformer_options=transformer_options)
+        attn_flat = self.expand_from_attn(attn_out.view(B * L, self.attn_dim))
+        attn_exp = attn_flat.view(BL, P2, self.pixel_dim)
+        x = torch.addcmul(x, gate_msa, attn_exp)
+        del msa_params, shift_msa, scale_msa, gate_msa
+
+        mlp_params = self.adaLN_modulation_mlp(s_cond).view(BL, P2, 3 * self.pixel_dim)
+        shift_mlp, scale_mlp, gate_mlp = mlp_params.chunk(3, dim=-1)
+        gate_mlp = gate_mlp.contiguous()  # detach from mlp_params so the del below frees shift+scale storage before the MLP
+        mlp_input = apply_adaln_(self.norm2(x), shift_mlp, scale_mlp)
+        del mlp_params, shift_mlp, scale_mlp
+
+        # MLP in chunks since the peak memory usage is huge here
+        chunk_size = (BL + self.mlp_chunks - 1) // self.mlp_chunks
+        for s in range(0, BL, chunk_size):
+            e = min(s + chunk_size, BL)
+            x[s:e].addcmul_(gate_mlp[s:e], self.mlp(mlp_input[s:e]))
+        return x

extensions-builtin/PiD/nn/pid.py

@@ -0,0 +1,227 @@
+"""PiD — Pixel Diffusion Decoder. Decodes a Flux/SD3/Flux2/Z-Image latent
+directly to a 4x-upscaled image in 4 distilled flow-matching steps. PixDiT_T2I
+body + LQ projection branch injected before each MMDiT patch block.
+"""
+
+from typing import List
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .model import PixDiT_T2I
+from .modules import precompute_freqs_cis_2d
+
+
+class SigmaAwareGatePerTokenPerDim(nn.Module):
+    """gate = sigmoid(content_proj(cat[x, lq]) - exp(log_alpha) * sigma); out = x + gate * lq.
+
+    Trained init gives ~0.88 gate at sigma=0, ~0.05 at sigma=1.
+    """
+
+    def __init__(self, dim: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.content_proj = operations.Linear(dim * 2, dim, dtype=dtype, device=device)
+        self.log_alpha = nn.Parameter(torch.empty((), dtype=dtype, device=device))
+
+    def forward(self, x: torch.Tensor, lq: torch.Tensor, sigma: torch.Tensor) -> torch.Tensor:
+        content_logit = self.content_proj(torch.cat([x, lq], dim=-1))
+        # log_alpha is a raw nn.Parameter -> doesn't auto-cast under dynamic VRAM.
+        log_alpha = self.log_alpha.to(device=x.device, dtype=torch.float32)
+        sigma_offset = -log_alpha.exp() * sigma.float().view(-1, 1, 1)
+        gate = torch.sigmoid(content_logit + sigma_offset)
+        return x + (gate * lq).to(x.dtype)
+
+
+class ResBlock(nn.Module):
+    """Pre-activation ResNet block: GN -> SiLU -> Conv -> GN -> SiLU -> Conv + skip."""
+
+    def __init__(self, channels: int, num_groups: int = 4, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.block = nn.Sequential(
+            operations.GroupNorm(num_groups, channels, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(channels, channels, kernel_size=3, padding=1, dtype=dtype, device=device),
+            operations.GroupNorm(num_groups, channels, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(channels, channels, kernel_size=3, padding=1, dtype=dtype, device=device),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x + self.block(x)
+
+
+class LQProjection2D(nn.Module):
+    """LQ latent -> per-block patch-aligned features for controlnet-style injection."""
+
+    def __init__(
+        self,
+        latent_channels: int,
+        hidden_dim: int = 512,
+        out_dim: int = 1536,
+        patch_size: int = 16,
+        sr_scale: int = 4,
+        latent_spatial_down_factor: int = 8,
+        num_res_blocks: int = 4,
+        num_outputs: int = 7,
+        interval: int = 2,
+        dtype=None, device=None, operations=None,
+    ):
+        super().__init__()
+        self.latent_channels = latent_channels
+        self.hidden_dim = hidden_dim
+        self.out_dim = out_dim
+        self.patch_size = patch_size
+        self.sr_scale = sr_scale
+        self.latent_spatial_down_factor = latent_spatial_down_factor
+        self.num_outputs = num_outputs
+        self.interval = interval
+
+        z_to_patch_ratio = (sr_scale * latent_spatial_down_factor) / patch_size
+        self.z_to_patch_ratio = z_to_patch_ratio
+        if z_to_patch_ratio >= 1:
+            self.latent_fold_factor = 0
+            latent_proj_in_ch = latent_channels
+        else:
+            fold_factor = int(1 / z_to_patch_ratio)
+            assert fold_factor * z_to_patch_ratio == 1.0
+            self.latent_fold_factor = fold_factor
+            latent_proj_in_ch = latent_channels * fold_factor * fold_factor
+
+        layers = [
+            operations.Conv2d(latent_proj_in_ch, hidden_dim, kernel_size=3, padding=1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(hidden_dim, hidden_dim, kernel_size=3, padding=1, dtype=dtype, device=device),
+        ]
+        for _ in range(num_res_blocks):
+            layers.append(ResBlock(hidden_dim, dtype=dtype, device=device, operations=operations))
+        self.latent_proj = nn.Sequential(*layers)
+
+        self.output_heads = nn.ModuleList(
+            [operations.Linear(hidden_dim, out_dim, dtype=dtype, device=device) for _ in range(num_outputs)]
+        )
+        self.gate_modules = nn.ModuleList(
+            [SigmaAwareGatePerTokenPerDim(out_dim, dtype=dtype, device=device, operations=operations)
+             for _ in range(num_outputs)]
+        )
+
+    def is_gate_active(self, block_idx: int) -> bool:
+        return block_idx % self.interval == 0
+
+    def output_index(self, block_idx: int) -> int:
+        return block_idx // self.interval
+
+    def gate(self, x: torch.Tensor, lq_feature: torch.Tensor, sigma: torch.Tensor, out_idx: int) -> torch.Tensor:
+        return self.gate_modules[out_idx](x, lq_feature, sigma)
+
+    def _align_latent_to_patch_grid(self, lq_latent: torch.Tensor, pH: int, pW: int) -> torch.Tensor:
+        B, z_dim = lq_latent.shape[:2]
+        if self.z_to_patch_ratio >= 1:
+            if lq_latent.shape[2] != pH or lq_latent.shape[3] != pW:
+                z_aligned = F.interpolate(lq_latent, size=(pH, pW), mode="nearest")
+            else:
+                z_aligned = lq_latent
+        else:
+            f = self.latent_fold_factor
+            zH_expected, zW_expected = pH * f, pW * f
+            if lq_latent.shape[2] != zH_expected or lq_latent.shape[3] != zW_expected:
+                lq_latent = F.interpolate(lq_latent, size=(zH_expected, zW_expected), mode="nearest")
+            z_aligned = lq_latent.reshape(B, z_dim, pH, f, pW, f).permute(0, 1, 3, 5, 2, 4)
+            z_aligned = z_aligned.reshape(B, z_dim * f * f, pH, pW)
+        return self.latent_proj(z_aligned)
+
+    def forward(self, lq_latent: torch.Tensor, target_pH: int, target_pW: int) -> List[torch.Tensor]:
+        feat = self._align_latent_to_patch_grid(lq_latent, target_pH, target_pW)
+        B, C, H, W = feat.shape
+        tokens = feat.permute(0, 2, 3, 1).contiguous().view(B, H * W, C)
+        return [head(tokens) for head in self.output_heads]
+
+
+class PidNet(PixDiT_T2I):
+    """PixDiT_T2I + LQ injection (one sigma-gated feature inserted before each patch block)."""
+
+    def __init__(
+        self,
+        lq_latent_channels: int = 16,
+        lq_hidden_dim: int = 512,
+        lq_num_res_blocks: int = 4,
+        lq_interval: int = 2,
+        sr_scale: int = 4,
+        latent_spatial_down_factor: int = 8,
+        rope_ref_h: int = 1024, # NTK ref resolution in PIXEL units: 1024px / patch=16 -> grid_ref=64.
+        rope_ref_w: int = 1024,
+        image_model=None,
+        dtype=None, device=None, operations=None,
+        **pixdit_kwargs,
+    ):
+        super().__init__(dtype=dtype, device=device, operations=operations, **pixdit_kwargs)
+
+        self.rope_ref_grid_h = rope_ref_h // self.patch_size
+        self.rope_ref_grid_w = rope_ref_w // self.patch_size
+
+        # Parent's PiTBlocks were built with plain RoPE — swap in NTK-aware.
+        def _pit_rope_fn(head_dim, h, w, device=None, dtype=torch.float32, **rope_opts):
+            return precompute_freqs_cis_2d(head_dim, h, w, ref_grid_h=self.rope_ref_grid_h, ref_grid_w=self.rope_ref_grid_w, device=device, dtype=dtype, **rope_opts)
+        for blk in self.pixel_blocks:
+            blk._rope_fn = _pit_rope_fn
+
+        num_lq_outputs = (self.patch_depth + lq_interval - 1) // lq_interval
+        self.lq_proj = LQProjection2D(
+            latent_channels=lq_latent_channels,
+            hidden_dim=lq_hidden_dim,
+            out_dim=self.hidden_size,
+            patch_size=self.patch_size,
+            sr_scale=sr_scale,
+            latent_spatial_down_factor=latent_spatial_down_factor,
+            num_res_blocks=lq_num_res_blocks,
+            num_outputs=num_lq_outputs,
+            interval=lq_interval,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+    def _fetch_patch_pos(self, height, width, device, dtype, **rope_opts):
+        return precompute_freqs_cis_2d(
+            self.hidden_size // self.num_groups,
+            height, width,
+            ref_grid_h=self.rope_ref_grid_h, ref_grid_w=self.rope_ref_grid_w,
+            device=device, dtype=dtype, **rope_opts,
+        )
+
+    def _pre_patch_block(self, s, i, pid_lq_features, pid_degrade_sigma, **kwargs):
+        if not self.lq_proj.is_gate_active(i):
+            return s
+        out_idx = self.lq_proj.output_index(i)
+        if out_idx >= len(pid_lq_features):
+            return s
+        return self.lq_proj.gate(s, pid_lq_features[out_idx], pid_degrade_sigma, out_idx)
+
+    def _forward(self, x, timesteps, context=None, attention_mask=None, transformer_options={}, lq_latent=None, degrade_sigma=None, **kwargs):
+        if lq_latent is None:
+            raise ValueError("PidNet requires lq_latent — attach via PiDConditioning")
+        expected_c = self.lq_proj.latent_channels
+        if lq_latent.shape[1] != expected_c:
+            raise ValueError(
+                f"Input latent has {lq_latent.shape[1]} channels, this model variant expects {expected_c}. "
+                f"Flux1/SD3 = 16 channels, Flux2 = 128 channels."
+            )
+        B = x.shape[0]
+        # Match the backbone's pad_to_patch_size (round up) so the LQ grid lines up with the patch stream.
+        Hs = -(-x.shape[2] // self.patch_size)
+        Ws = -(-x.shape[3] // self.patch_size)
+
+        degrade_sigma = degrade_sigma.to(device=x.device, dtype=torch.float32).reshape(-1)
+        if degrade_sigma.numel() == 1 and B > 1:
+            degrade_sigma = degrade_sigma.expand(B).contiguous()
+
+        lq_features = self.lq_proj(lq_latent=lq_latent.to(x), target_pH=Hs, target_pW=Ws)
+
+        return super()._forward(
+            x, timesteps,
+            context=context, attention_mask=attention_mask,
+            transformer_options=transformer_options,
+            pid_lq_features=lq_features,
+            pid_degrade_sigma=degrade_sigma,
+            **kwargs,
+        )