Mochi VAE encoder.

comfy/ldm/genmo/vae/model.py

@@ -2,12 +2,16 @@
 #adapted to ComfyUI
 
 from typing import Callable, List, Optional, Tuple, Union
+from functools import partial
+import math
 
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
 from einops import rearrange
 
+from comfy.ldm.modules.attention import optimized_attention
+
 import comfy.ops
 ops = comfy.ops.disable_weight_init
 
@@ -158,8 +162,10 @@ class ResBlock(nn.Module):
         *,
         affine: bool = True,
         attn_block: Optional[nn.Module] = None,
-        padding_mode: str = "replicate",
         causal: bool = True,
+        prune_bottleneck: bool = False,
+        padding_mode: str,
+        bias: bool = True,
     ):
         super().__init__()
         self.channels = channels
@@ -170,23 +176,23 @@ class ResBlock(nn.Module):
             nn.SiLU(inplace=True),
             PConv3d(
                 in_channels=channels,
-                out_channels=channels,
+                out_channels=channels // 2 if prune_bottleneck else channels,
                 kernel_size=(3, 3, 3),
                 stride=(1, 1, 1),
                 padding_mode=padding_mode,
-                bias=True,
+                bias=bias,
-                # causal=causal,
+                causal=causal,
             ),
             norm_fn(channels, affine=affine),
             nn.SiLU(inplace=True),
             PConv3d(
-                in_channels=channels,
+                in_channels=channels // 2 if prune_bottleneck else channels,
                 out_channels=channels,
                 kernel_size=(3, 3, 3),
                 stride=(1, 1, 1),
                 padding_mode=padding_mode,
-                bias=True,
+                bias=bias,
-                # causal=causal,
+                causal=causal,
             ),
         )
 
@@ -206,6 +212,81 @@ class ResBlock(nn.Module):
         return self.attn_block(x)
 
 
+class Attention(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        head_dim: int = 32,
+        qkv_bias: bool = False,
+        out_bias: bool = True,
+        qk_norm: bool = True,
+    ) -> None:
+        super().__init__()
+        self.head_dim = head_dim
+        self.num_heads = dim // head_dim
+        self.qk_norm = qk_norm
+
+        self.qkv = nn.Linear(dim, 3 * dim, bias=qkv_bias)
+        self.out = nn.Linear(dim, dim, bias=out_bias)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+    ) -> torch.Tensor:
+        """Compute temporal self-attention.
+
+        Args:
+            x: Input tensor. Shape: [B, C, T, H, W].
+            chunk_size: Chunk size for large tensors.
+
+        Returns:
+            x: Output tensor. Shape: [B, C, T, H, W].
+        """
+        B, _, T, H, W = x.shape
+
+        if T == 1:
+            # No attention for single frame.
+            x = x.movedim(1, -1)  # [B, C, T, H, W] -> [B, T, H, W, C]
+            qkv = self.qkv(x)
+            _, _, x = qkv.chunk(3, dim=-1)  # Throw away queries and keys.
+            x = self.out(x)
+            return x.movedim(-1, 1)  # [B, T, H, W, C] -> [B, C, T, H, W]
+
+        # 1D temporal attention.
+        x = rearrange(x, "B C t h w -> (B h w) t C")
+        qkv = self.qkv(x)
+
+        # Input: qkv with shape [B, t, 3 * num_heads * head_dim]
+        # Output: x with shape [B, num_heads, t, head_dim]
+        q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, self.head_dim).transpose(1, 3).unbind(2)
+
+        if self.qk_norm:
+            q = F.normalize(q, p=2, dim=-1)
+            k = F.normalize(k, p=2, dim=-1)
+
+        x = optimized_attention(q, k, v, self.num_heads, skip_reshape=True)
+
+        assert x.size(0) == q.size(0)
+
+        x = self.out(x)
+        x = rearrange(x, "(B h w) t C -> B C t h w", B=B, h=H, w=W)
+        return x
+
+
+class AttentionBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        **attn_kwargs,
+    ) -> None:
+        super().__init__()
+        self.norm = norm_fn(dim)
+        self.attn = Attention(dim, **attn_kwargs)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x + self.attn(self.norm(x))
+
+
 class CausalUpsampleBlock(nn.Module):
     def __init__(
         self,
@@ -244,14 +325,9 @@ class CausalUpsampleBlock(nn.Module):
         return x
 
 
-def block_fn(channels, *, has_attention: bool = False, **block_kwargs):
+def block_fn(channels, *, affine: bool = True, has_attention: bool = False, **block_kwargs):
-    assert has_attention is False #NOTE: if this is ever true add back the attention code.
+    attn_block = AttentionBlock(channels) if has_attention else None
-
+    return ResBlock(channels, affine=affine, attn_block=attn_block, **block_kwargs)
-    attn_block = None #AttentionBlock(channels) if has_attention else None
-
-    return ResBlock(
-        channels, affine=True, attn_block=attn_block, **block_kwargs
-    )
 
 
 class DownsampleBlock(nn.Module):
@@ -288,8 +364,9 @@ class DownsampleBlock(nn.Module):
                 out_channels=out_channels,
                 kernel_size=(temporal_reduction, spatial_reduction, spatial_reduction),
                 stride=(temporal_reduction, spatial_reduction, spatial_reduction),
+                # First layer in each block always uses replicate padding
                 padding_mode="replicate",
-                bias=True,
+                bias=block_kwargs["bias"],
             )
         )
 
@@ -382,7 +459,7 @@ class Decoder(nn.Module):
         blocks = []
 
         first_block = [
-            nn.Conv3d(latent_dim, ch[-1], kernel_size=(1, 1, 1))
+            ops.Conv3d(latent_dim, ch[-1], kernel_size=(1, 1, 1))
         ]  # Input layer.
         # First set of blocks preserve channel count.
         for _ in range(num_res_blocks[-1]):
@@ -452,11 +529,165 @@ class Decoder(nn.Module):
 
         return self.output_proj(x).contiguous()
 
+class LatentDistribution:
+    def __init__(self, mean: torch.Tensor, logvar: torch.Tensor):
+        """Initialize latent distribution.
+
+        Args:
+            mean: Mean of the distribution. Shape: [B, C, T, H, W].
+            logvar: Logarithm of variance of the distribution. Shape: [B, C, T, H, W].
+        """
+        assert mean.shape == logvar.shape
+        self.mean = mean
+        self.logvar = logvar
+
+    def sample(self, temperature=1.0, generator: torch.Generator = None, noise=None):
+        if temperature == 0.0:
+            return self.mean
+
+        if noise is None:
+            noise = torch.randn(self.mean.shape, device=self.mean.device, dtype=self.mean.dtype, generator=generator)
+        else:
+            assert noise.device == self.mean.device
+            noise = noise.to(self.mean.dtype)
+
+        if temperature != 1.0:
+            raise NotImplementedError(f"Temperature {temperature} is not supported.")
+
+        # Just Gaussian sample with no scaling of variance.
+        return noise * torch.exp(self.logvar * 0.5) + self.mean
+
+    def mode(self):
+        return self.mean
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        in_channels: int,
+        base_channels: int,
+        channel_multipliers: List[int],
+        num_res_blocks: List[int],
+        latent_dim: int,
+        temporal_reductions: List[int],
+        spatial_reductions: List[int],
+        prune_bottlenecks: List[bool],
+        has_attentions: List[bool],
+        affine: bool = True,
+        bias: bool = True,
+        input_is_conv_1x1: bool = False,
+        padding_mode: str,
+    ):
+        super().__init__()
+        self.temporal_reductions = temporal_reductions
+        self.spatial_reductions = spatial_reductions
+        self.base_channels = base_channels
+        self.channel_multipliers = channel_multipliers
+        self.num_res_blocks = num_res_blocks
+        self.latent_dim = latent_dim
+
+        self.fourier_features = FourierFeatures()
+        ch = [mult * base_channels for mult in channel_multipliers]
+        num_down_blocks = len(ch) - 1
+        assert len(num_res_blocks) == num_down_blocks + 2
+
+        layers = (
+            [ops.Conv3d(in_channels, ch[0], kernel_size=(1, 1, 1), bias=True)]
+            if not input_is_conv_1x1
+            else [Conv1x1(in_channels, ch[0])]
+        )
+
+        assert len(prune_bottlenecks) == num_down_blocks + 2
+        assert len(has_attentions) == num_down_blocks + 2
+        block = partial(block_fn, padding_mode=padding_mode, affine=affine, bias=bias)
+
+        for _ in range(num_res_blocks[0]):
+            layers.append(block(ch[0], has_attention=has_attentions[0], prune_bottleneck=prune_bottlenecks[0]))
+        prune_bottlenecks = prune_bottlenecks[1:]
+        has_attentions = has_attentions[1:]
+
+        assert len(temporal_reductions) == len(spatial_reductions) == len(ch) - 1
+        for i in range(num_down_blocks):
+            layer = DownsampleBlock(
+                ch[i],
+                ch[i + 1],
+                num_res_blocks=num_res_blocks[i + 1],
+                temporal_reduction=temporal_reductions[i],
+                spatial_reduction=spatial_reductions[i],
+                prune_bottleneck=prune_bottlenecks[i],
+                has_attention=has_attentions[i],
+                affine=affine,
+                bias=bias,
+                padding_mode=padding_mode,
+            )
+
+            layers.append(layer)
+
+        # Additional blocks.
+        for _ in range(num_res_blocks[-1]):
+            layers.append(block(ch[-1], has_attention=has_attentions[-1], prune_bottleneck=prune_bottlenecks[-1]))
+
+        self.layers = nn.Sequential(*layers)
+
+        # Output layers.
+        self.output_norm = norm_fn(ch[-1])
+        self.output_proj = Conv1x1(ch[-1], 2 * latent_dim, bias=False)
+
+    @property
+    def temporal_downsample(self):
+        return math.prod(self.temporal_reductions)
+
+    @property
+    def spatial_downsample(self):
+        return math.prod(self.spatial_reductions)
+
+    def forward(self, x) -> LatentDistribution:
+        """Forward pass.
+
+        Args:
+            x: Input video tensor. Shape: [B, C, T, H, W]. Scaled to [-1, 1]
+
+        Returns:
+            means: Latent tensor. Shape: [B, latent_dim, t, h, w]. Scaled [-1, 1].
+                   h = H // 8, w = W // 8, t - 1 = (T - 1) // 6
+            logvar: Shape: [B, latent_dim, t, h, w].
+        """
+        assert x.ndim == 5, f"Expected 5D input, got {x.shape}"
+        x = self.fourier_features(x)
+
+        x = self.layers(x)
+
+        x = self.output_norm(x)
+        x = F.silu(x, inplace=True)
+        x = self.output_proj(x)
+
+        means, logvar = torch.chunk(x, 2, dim=1)
+
+        assert means.ndim == 5
+        assert logvar.shape == means.shape
+        assert means.size(1) == self.latent_dim
+
+        return LatentDistribution(means, logvar)
+
 
 class VideoVAE(nn.Module):
     def __init__(self):
         super().__init__()
-        self.encoder = None #TODO once the model releases
+        self.encoder = Encoder(
+            in_channels=15,
+            base_channels=64,
+            channel_multipliers=[1, 2, 4, 6],
+            num_res_blocks=[3, 3, 4, 6, 3],
+            latent_dim=12,
+            temporal_reductions=[1, 2, 3],
+            spatial_reductions=[2, 2, 2],
+            prune_bottlenecks=[False, False, False, False, False],
+            has_attentions=[False, True, True, True, True],
+            affine=True,
+            bias=True,
+            input_is_conv_1x1=True,
+            padding_mode="replicate"
+        )
         self.decoder = Decoder(
             out_channels=3,
             base_channels=128,
@@ -474,7 +705,7 @@ class VideoVAE(nn.Module):
         )
 
     def encode(self, x):
-        return self.encoder(x)
+        return self.encoder(x).mode()
 
     def decode(self, x):
         return self.decoder(x)

comfy/ldm/modules/attention.py

@@ -393,6 +393,13 @@ def attention_xformers(q, k, v, heads, mask=None, attn_precision=None, skip_resh
 
     return out
 
+if model_management.is_nvidia(): #pytorch 2.3 and up seem to have this issue.
+    SDP_BATCH_LIMIT = 2**15
+else:
+    #TODO: other GPUs ?
+    SDP_BATCH_LIMIT = 2**31
+
+
 def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False):
     if skip_reshape:
         b, _, _, dim_head = q.shape
@@ -404,10 +411,15 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
             (q, k, v),
         )
 
+    if SDP_BATCH_LIMIT >= q.shape[0]:
         out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
         out = (
             out.transpose(1, 2).reshape(b, -1, heads * dim_head)
         )
+    else:
+        out = torch.empty((q.shape[0], q.shape[2], heads * dim_head), dtype=q.dtype, layout=q.layout, device=q.device)
+        for i in range(0, q.shape[0], SDP_BATCH_LIMIT):
+            out[i : i + SDP_BATCH_LIMIT] = torch.nn.functional.scaled_dot_product_attention(q[i : i + SDP_BATCH_LIMIT], k[i : i + SDP_BATCH_LIMIT], v[i : i + SDP_BATCH_LIMIT], attn_mask=mask, dropout_p=0.0, is_causal=False).transpose(1, 2).reshape(-1, q.shape[2], heads * dim_head)
     return out
 
 

comfy/sd.py

@@ -171,6 +171,7 @@ class VAE:
         self.downscale_ratio = 8
         self.upscale_ratio = 8
         self.latent_channels = 4
+        self.latent_dim = 2
         self.output_channels = 3
         self.process_input = lambda image: image * 2.0 - 1.0
         self.process_output = lambda image: torch.clamp((image + 1.0) / 2.0, min=0.0, max=1.0)
@@ -240,16 +241,22 @@ class VAE:
                 self.output_channels = 2
                 self.upscale_ratio = 2048
                 self.downscale_ratio =  2048
+                self.latent_dim = 1
                 self.process_output = lambda audio: audio
                 self.process_input = lambda audio: audio
                 self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32]
-            elif "blocks.2.blocks.3.stack.5.weight" in sd or "decoder.blocks.2.blocks.3.stack.5.weight" in sd: #genmo mochi vae
+            elif "blocks.2.blocks.3.stack.5.weight" in sd or "decoder.blocks.2.blocks.3.stack.5.weight" in sd or "layers.4.layers.1.attn_block.attn.qkv.weight" in sd or "encoder.layers.4.layers.1.attn_block.attn.qkv.weight": #genmo mochi vae
                 if "blocks.2.blocks.3.stack.5.weight" in sd:
                     sd = comfy.utils.state_dict_prefix_replace(sd, {"": "decoder."})
+                if "layers.4.layers.1.attn_block.attn.qkv.weight" in sd:
+                    sd = comfy.utils.state_dict_prefix_replace(sd, {"": "encoder."})
                 self.first_stage_model = comfy.ldm.genmo.vae.model.VideoVAE()
                 self.latent_channels = 12
+                self.latent_dim = 3
                 self.memory_used_decode = lambda shape, dtype: (1000 * shape[2] * shape[3] * shape[4] * (6 * 8 * 8)) * model_management.dtype_size(dtype)
+                self.memory_used_encode = lambda shape, dtype: (1.5 * max(shape[2], 7) * shape[3] * shape[4] * (6 * 8 * 8)) * model_management.dtype_size(dtype)
                 self.upscale_ratio = (lambda a: max(0, a * 6 - 5), 8, 8)
+                self.working_dtypes = [torch.float16, torch.float32]
             else:
                 logging.warning("WARNING: No VAE weights detected, VAE not initalized.")
                 self.first_stage_model = None
@@ -361,17 +368,22 @@ class VAE:
 
     def encode(self, pixel_samples):
         pixel_samples = self.vae_encode_crop_pixels(pixel_samples)
-        pixel_samples = pixel_samples.movedim(-1,1)
+        pixel_samples = pixel_samples.movedim(-1, 1)
+        if self.latent_dim == 3:
+            pixel_samples = pixel_samples.movedim(1, 0).unsqueeze(0)
         try:
             memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype)
             model_management.load_models_gpu([self.patcher], memory_required=memory_used)
             free_memory = model_management.get_free_memory(self.device)
             batch_number = int(free_memory / max(1, memory_used))
             batch_number = max(1, batch_number)
-            samples = torch.empty((pixel_samples.shape[0], self.latent_channels) + tuple(map(lambda a: a // self.downscale_ratio, pixel_samples.shape[2:])), device=self.output_device)
+            samples = None
             for x in range(0, pixel_samples.shape[0], batch_number):
-                pixels_in = self.process_input(pixel_samples[x:x+batch_number]).to(self.vae_dtype).to(self.device)
+                pixels_in = self.process_input(pixel_samples[x:x + batch_number]).to(self.vae_dtype).to(self.device)
-                samples[x:x+batch_number] = self.first_stage_model.encode(pixels_in).to(self.output_device).float()
+                out = self.first_stage_model.encode(pixels_in).to(self.output_device).float()
+                if samples is None:
+                    samples = torch.empty((pixel_samples.shape[0],) + tuple(out.shape[1:]), device=self.output_device)
+                samples[x:x + batch_number] = out
 
         except model_management.OOM_EXCEPTION as e:
             logging.warning("Warning: Ran out of memory when regular VAE encoding, retrying with tiled VAE encoding.")