Merge branch 'master' into dr-support-pip-cm

comfy/k_diffusion/sa_solver.py

@@ -0,0 +1,121 @@
+# SA-Solver: Stochastic Adams Solver (NeurIPS 2023, arXiv:2309.05019)
+# Conference: https://proceedings.neurips.cc/paper_files/paper/2023/file/f4a6806490d31216a3ba667eb240c897-Paper-Conference.pdf
+# Codebase ref: https://github.com/scxue/SA-Solver
+
+import math
+from typing import Union, Callable
+import torch
+
+
+def compute_exponential_coeffs(s: torch.Tensor, t: torch.Tensor, solver_order: int, tau_t: float) -> torch.Tensor:
+    """Compute (1 + tau^2) * integral of exp((1 + tau^2) * x) * x^p dx from s to t with exp((1 + tau^2) * t) factored out, using integration by parts.
+
+    Integral of exp((1 + tau^2) * x) * x^p dx
+        = product_terms[p] - (p / (1 + tau^2)) * integral of exp((1 + tau^2) * x) * x^(p-1) dx,
+    with base case p=0 where integral equals product_terms[0].
+
+    where
+        product_terms[p] = x^p * exp((1 + tau^2) * x) / (1 + tau^2).
+
+    Construct a recursive coefficient matrix following the above recursive relation to compute all integral terms up to p = (solver_order - 1).
+    Return coefficients used by the SA-Solver in data prediction mode.
+
+    Args:
+        s: Start time s.
+        t: End time t.
+        solver_order: Current order of the solver.
+        tau_t: Stochastic strength parameter in the SDE.
+
+    Returns:
+        Exponential coefficients used in data prediction, with exp((1 + tau^2) * t) factored out, ordered from p=0 to p=solver_order−1, shape (solver_order,).
+    """
+    tau_mul = 1 + tau_t ** 2
+    h = t - s
+    p = torch.arange(solver_order, dtype=s.dtype, device=s.device)
+
+    # product_terms after factoring out exp((1 + tau^2) * t)
+    # Includes (1 + tau^2) factor from outside the integral
+    product_terms_factored = (t ** p - s ** p * (-tau_mul * h).exp())
+
+    # Lower triangular recursive coefficient matrix
+    # Accumulates recursive coefficients based on p / (1 + tau^2)
+    recursive_depth_mat = p.unsqueeze(1) - p.unsqueeze(0)
+    log_factorial = (p + 1).lgamma()
+    recursive_coeff_mat = log_factorial.unsqueeze(1) - log_factorial.unsqueeze(0)
+    if tau_t > 0:
+        recursive_coeff_mat = recursive_coeff_mat - (recursive_depth_mat * math.log(tau_mul))
+    signs = torch.where(recursive_depth_mat % 2 == 0, 1.0, -1.0)
+    recursive_coeff_mat = (recursive_coeff_mat.exp() * signs).tril()
+
+    return recursive_coeff_mat @ product_terms_factored
+
+
+def compute_simple_stochastic_adams_b_coeffs(sigma_next: torch.Tensor, curr_lambdas: torch.Tensor, lambda_s: torch.Tensor, lambda_t: torch.Tensor, tau_t: float, is_corrector_step: bool = False) -> torch.Tensor:
+    """Compute simple order-2 b coefficients from SA-Solver paper (Appendix D. Implementation Details)."""
+    tau_mul = 1 + tau_t ** 2
+    h = lambda_t - lambda_s
+    alpha_t = sigma_next * lambda_t.exp()
+    if is_corrector_step:
+        # Simplified 1-step (order-2) corrector
+        b_1 = alpha_t * (0.5 * tau_mul * h)
+        b_2 = alpha_t * (-h * tau_mul).expm1().neg() - b_1
+    else:
+        # Simplified 2-step predictor
+        b_2 = alpha_t * (0.5 * tau_mul * h ** 2) / (curr_lambdas[-2] - lambda_s)
+        b_1 = alpha_t * (-h * tau_mul).expm1().neg() - b_2
+    return torch.stack([b_2, b_1])
+
+
+def compute_stochastic_adams_b_coeffs(sigma_next: torch.Tensor, curr_lambdas: torch.Tensor, lambda_s: torch.Tensor, lambda_t: torch.Tensor, tau_t: float, simple_order_2: bool = False, is_corrector_step: bool = False) -> torch.Tensor:
+    """Compute b_i coefficients for the SA-Solver (see eqs. 15 and 18).
+
+    The solver order corresponds to the number of input lambdas (half-logSNR points).
+
+    Args:
+        sigma_next: Sigma at end time t.
+        curr_lambdas: Lambda time points used to construct the Lagrange basis, shape (N,).
+        lambda_s: Lambda at start time s.
+        lambda_t: Lambda at end time t.
+        tau_t: Stochastic strength parameter in the SDE.
+        simple_order_2: Whether to enable the simple order-2 scheme.
+        is_corrector_step: Flag for corrector step in simple order-2 mode.
+
+    Returns:
+        b_i coefficients for the SA-Solver, shape (N,), where N is the solver order.
+    """
+    num_timesteps = curr_lambdas.shape[0]
+
+    if simple_order_2 and num_timesteps == 2:
+        return compute_simple_stochastic_adams_b_coeffs(sigma_next, curr_lambdas, lambda_s, lambda_t, tau_t, is_corrector_step)
+
+    # Compute coefficients by solving a linear system from Lagrange basis interpolation
+    exp_integral_coeffs = compute_exponential_coeffs(lambda_s, lambda_t, num_timesteps, tau_t)
+    vandermonde_matrix_T = torch.vander(curr_lambdas, num_timesteps, increasing=True).T
+    lagrange_integrals = torch.linalg.solve(vandermonde_matrix_T, exp_integral_coeffs)
+
+    # (sigma_t * exp(-tau^2 * lambda_t)) * exp((1 + tau^2) * lambda_t)
+    # = sigma_t * exp(lambda_t) = alpha_t
+    # exp((1 + tau^2) * lambda_t) is extracted from the integral
+    alpha_t = sigma_next * lambda_t.exp()
+    return alpha_t * lagrange_integrals
+
+
+def get_tau_interval_func(start_sigma: float, end_sigma: float, eta: float = 1.0) -> Callable[[Union[torch.Tensor, float]], float]:
+    """Return a function that controls the stochasticity of SA-Solver.
+
+    When eta = 0, SA-Solver runs as ODE. The official approach uses
+    time t to determine the SDE interval, while here we use sigma instead.
+
+    See:
+        https://github.com/scxue/SA-Solver/blob/main/README.md
+    """
+
+    def tau_func(sigma: Union[torch.Tensor, float]) -> float:
+        if eta <= 0:
+            return 0.0  # ODE
+
+        if isinstance(sigma, torch.Tensor):
+            sigma = sigma.item()
+        return eta if start_sigma >= sigma >= end_sigma else 0.0
+
+    return tau_func

comfy/k_diffusion/sampling.py

@@ -9,6 +9,7 @@ from tqdm.auto import trange, tqdm
 
 from . import utils
 from . import deis
+from . import sa_solver
 import comfy.model_patcher
 import comfy.model_sampling
 
@@ -1648,3 +1649,113 @@ def sample_seeds_3(model, x, sigmas, extra_args=None, callback=None, disable=Non
             if inject_noise:
                 x = x + sigmas[i + 1] * (noise_coeff_3 * noise_1 + noise_coeff_2 * noise_2 + noise_coeff_1 * noise_3) * s_noise
     return x
+
+
+@torch.no_grad()
+def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=False, tau_func=None, s_noise=1.0, noise_sampler=None, predictor_order=3, corrector_order=4, use_pece=False, simple_order_2=False):
+    """Stochastic Adams Solver with predictor-corrector method (NeurIPS 2023)."""
+    if len(sigmas) <= 1:
+        return x
+    extra_args = {} if extra_args is None else extra_args
+    seed = extra_args.get("seed", None)
+    noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
+    s_in = x.new_ones([x.shape[0]])
+
+    model_sampling = model.inner_model.model_patcher.get_model_object("model_sampling")
+    sigmas = offset_first_sigma_for_snr(sigmas, model_sampling)
+    lambdas = sigma_to_half_log_snr(sigmas, model_sampling=model_sampling)
+
+    if tau_func is None:
+        # Use default interval for stochastic sampling
+        start_sigma = model_sampling.percent_to_sigma(0.2)
+        end_sigma = model_sampling.percent_to_sigma(0.8)
+        tau_func = sa_solver.get_tau_interval_func(start_sigma, end_sigma, eta=1.0)
+
+    max_used_order = max(predictor_order, corrector_order)
+    x_pred = x  # x: current state, x_pred: predicted next state
+
+    h = 0.0
+    tau_t = 0.0
+    noise = 0.0
+    pred_list = []
+
+    # Lower order near the end to improve stability
+    lower_order_to_end = sigmas[-1].item() == 0
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        # Evaluation
+        denoised = model(x_pred, sigmas[i] * s_in, **extra_args)
+        if callback is not None:
+            callback({"x": x_pred, "i": i, "sigma": sigmas[i], "sigma_hat": sigmas[i], "denoised": denoised})
+        pred_list.append(denoised)
+        pred_list = pred_list[-max_used_order:]
+
+        predictor_order_used = min(predictor_order, len(pred_list))
+        if i == 0 or (sigmas[i + 1] == 0 and not use_pece):
+            corrector_order_used = 0
+        else:
+            corrector_order_used = min(corrector_order, len(pred_list))
+
+        if lower_order_to_end:
+            predictor_order_used = min(predictor_order_used, len(sigmas) - 2 - i)
+            corrector_order_used = min(corrector_order_used, len(sigmas) - 1 - i)
+
+        # Corrector
+        if corrector_order_used == 0:
+            # Update by the predicted state
+            x = x_pred
+        else:
+            curr_lambdas = lambdas[i - corrector_order_used + 1:i + 1]
+            b_coeffs = sa_solver.compute_stochastic_adams_b_coeffs(
+                sigmas[i],
+                curr_lambdas,
+                lambdas[i - 1],
+                lambdas[i],
+                tau_t,
+                simple_order_2,
+                is_corrector_step=True,
+            )
+            pred_mat = torch.stack(pred_list[-corrector_order_used:], dim=1)    # (B, K, ...)
+            corr_res = torch.tensordot(pred_mat, b_coeffs, dims=([1], [0]))  # (B, ...)
+            x = sigmas[i] / sigmas[i - 1] * (-(tau_t ** 2) * h).exp() * x + corr_res
+
+            if tau_t > 0 and s_noise > 0:
+                # The noise from the previous predictor step
+                x = x + noise
+
+            if use_pece:
+                # Evaluate the corrected state
+                denoised = model(x, sigmas[i] * s_in, **extra_args)
+                pred_list[-1] = denoised
+
+        # Predictor
+        if sigmas[i + 1] == 0:
+            # Denoising step
+            x = denoised
+        else:
+            tau_t = tau_func(sigmas[i + 1])
+            curr_lambdas = lambdas[i - predictor_order_used + 1:i + 1]
+            b_coeffs = sa_solver.compute_stochastic_adams_b_coeffs(
+                sigmas[i + 1],
+                curr_lambdas,
+                lambdas[i],
+                lambdas[i + 1],
+                tau_t,
+                simple_order_2,
+                is_corrector_step=False,
+            )
+            pred_mat = torch.stack(pred_list[-predictor_order_used:], dim=1)    # (B, K, ...)
+            pred_res = torch.tensordot(pred_mat, b_coeffs, dims=([1], [0]))  # (B, ...)
+            h = lambdas[i + 1] - lambdas[i]
+            x_pred = sigmas[i + 1] / sigmas[i] * (-(tau_t ** 2) * h).exp() * x + pred_res
+
+            if tau_t > 0 and s_noise > 0:
+                noise = noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * tau_t ** 2 * h).expm1().neg().sqrt() * s_noise
+                x_pred = x_pred + noise
+    return x
+
+
+@torch.no_grad()
+def sample_sa_solver_pece(model, x, sigmas, extra_args=None, callback=None, disable=False, tau_func=None, s_noise=1.0, noise_sampler=None, predictor_order=3, corrector_order=4, simple_order_2=False):
+    """Stochastic Adams Solver with PECE (Predict–Evaluate–Correct–Evaluate) mode (NeurIPS 2023)."""
+    return sample_sa_solver(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, tau_func=tau_func, s_noise=s_noise, noise_sampler=noise_sampler, predictor_order=predictor_order, corrector_order=corrector_order, use_pece=True, simple_order_2=simple_order_2)

comfy/ldm/chroma/model.py

@@ -254,13 +254,12 @@ class Chroma(nn.Module):
 
     def forward(self, x, timestep, context, guidance, control=None, transformer_options={}, **kwargs):
         bs, c, h, w = x.shape
-        patch_size = 2
-        x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
 
-        img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
+        img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=self.patch_size, pw=self.patch_size)
 
-        h_len = ((h + (patch_size // 2)) // patch_size)
-        w_len = ((w + (patch_size // 2)) // patch_size)
+        h_len = ((h + (self.patch_size // 2)) // self.patch_size)
+        w_len = ((w + (self.patch_size // 2)) // self.patch_size)
         img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
         img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1)
         img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
@@ -268,4 +267,4 @@ class Chroma(nn.Module):
 
         txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
         out = self.forward_orig(img, img_ids, context, txt_ids, timestep, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None))
-        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)[:,:,:h,:w]
+        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=self.patch_size, pw=self.patch_size)[:,:,:h,:w]

comfy/ldm/pixart/pixartms.py

@@ -1,256 +1,256 @@
-# Based on:
-# https://github.com/PixArt-alpha/PixArt-alpha [Apache 2.0 license]
-# https://github.com/PixArt-alpha/PixArt-sigma [Apache 2.0 license]
-import torch
-import torch.nn as nn
-
-from .blocks import (
-    t2i_modulate,
-    CaptionEmbedder,
-    AttentionKVCompress,
-    MultiHeadCrossAttention,
-    T2IFinalLayer,
-    SizeEmbedder,
-)
-from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder, PatchEmbed, Mlp, get_1d_sincos_pos_embed_from_grid_torch
-
-
-def get_2d_sincos_pos_embed_torch(embed_dim, w, h, pe_interpolation=1.0, base_size=16, device=None, dtype=torch.float32):
-    grid_h, grid_w = torch.meshgrid(
-        torch.arange(h, device=device, dtype=dtype) / (h/base_size) / pe_interpolation,
-        torch.arange(w, device=device, dtype=dtype) / (w/base_size) / pe_interpolation,
-        indexing='ij'
-    )
-    emb_h = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_h, device=device, dtype=dtype)
-    emb_w = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_w, device=device, dtype=dtype)
-    emb = torch.cat([emb_w, emb_h], dim=1)  # (H*W, D)
-    return emb
-
-class PixArtMSBlock(nn.Module):
-    """
-    A PixArt block with adaptive layer norm zero (adaLN-Zero) conditioning.
-    """
-    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, drop_path=0., input_size=None,
-                 sampling=None, sr_ratio=1, qk_norm=False, dtype=None, device=None, operations=None, **block_kwargs):
-        super().__init__()
-        self.hidden_size = hidden_size
-        self.norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.attn = AttentionKVCompress(
-            hidden_size, num_heads=num_heads, qkv_bias=True, sampling=sampling, sr_ratio=sr_ratio,
-            qk_norm=qk_norm, dtype=dtype, device=device, operations=operations, **block_kwargs
-        )
-        self.cross_attn = MultiHeadCrossAttention(
-            hidden_size, num_heads, dtype=dtype, device=device, operations=operations, **block_kwargs
-        )
-        self.norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        # to be compatible with lower version pytorch
-        approx_gelu = lambda: nn.GELU(approximate="tanh")
-        self.mlp = Mlp(
-            in_features=hidden_size, hidden_features=int(hidden_size * mlp_ratio), act_layer=approx_gelu,
-            dtype=dtype, device=device, operations=operations
-        )
-        self.scale_shift_table = nn.Parameter(torch.randn(6, hidden_size) / hidden_size ** 0.5)
-
-    def forward(self, x, y, t, mask=None, HW=None, **kwargs):
-        B, N, C = x.shape
-
-        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None].to(dtype=x.dtype, device=x.device) + t.reshape(B, 6, -1)).chunk(6, dim=1)
-        x = x + (gate_msa * self.attn(t2i_modulate(self.norm1(x), shift_msa, scale_msa), HW=HW))
-        x = x + self.cross_attn(x, y, mask)
-        x = x + (gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp)))
-
-        return x
-
-
-### Core PixArt Model ###
-class PixArtMS(nn.Module):
-    """
-    Diffusion model with a Transformer backbone.
-    """
-    def __init__(
-            self,
-            input_size=32,
-            patch_size=2,
-            in_channels=4,
-            hidden_size=1152,
-            depth=28,
-            num_heads=16,
-            mlp_ratio=4.0,
-            class_dropout_prob=0.1,
-            learn_sigma=True,
-            pred_sigma=True,
-            drop_path: float = 0.,
-            caption_channels=4096,
-            pe_interpolation=None,
-            pe_precision=None,
-            config=None,
-            model_max_length=120,
-            micro_condition=True,
-            qk_norm=False,
-            kv_compress_config=None,
-            dtype=None,
-            device=None,
-            operations=None,
-            **kwargs,
-    ):
-        nn.Module.__init__(self)
-        self.dtype = dtype
-        self.pred_sigma = pred_sigma
-        self.in_channels = in_channels
-        self.out_channels = in_channels * 2 if pred_sigma else in_channels
-        self.patch_size = patch_size
-        self.num_heads = num_heads
-        self.pe_interpolation = pe_interpolation
-        self.pe_precision = pe_precision
-        self.hidden_size = hidden_size
-        self.depth = depth
-
-        approx_gelu = lambda: nn.GELU(approximate="tanh")
-        self.t_block = nn.Sequential(
-            nn.SiLU(),
-            operations.Linear(hidden_size, 6 * hidden_size, bias=True, dtype=dtype, device=device)
-        )
-        self.x_embedder = PatchEmbed(
-            patch_size=patch_size,
-            in_chans=in_channels,
-            embed_dim=hidden_size,
-            bias=True,
-            dtype=dtype,
-            device=device,
-            operations=operations
-        )
-        self.t_embedder = TimestepEmbedder(
-            hidden_size, dtype=dtype, device=device, operations=operations,
-        )
-        self.y_embedder = CaptionEmbedder(
-            in_channels=caption_channels, hidden_size=hidden_size, uncond_prob=class_dropout_prob,
-            act_layer=approx_gelu, token_num=model_max_length,
-            dtype=dtype, device=device, operations=operations,
-        )
-
-        self.micro_conditioning = micro_condition
-        if self.micro_conditioning:
-            self.csize_embedder = SizeEmbedder(hidden_size//3, dtype=dtype, device=device, operations=operations)
-            self.ar_embedder = SizeEmbedder(hidden_size//3, dtype=dtype, device=device, operations=operations)
-
-        # For fixed sin-cos embedding:
-        # num_patches = (input_size // patch_size) * (input_size // patch_size)
-        # self.base_size = input_size // self.patch_size
-        # self.register_buffer("pos_embed", torch.zeros(1, num_patches, hidden_size))
-
-        drop_path = [x.item() for x in torch.linspace(0, drop_path, depth)]  # stochastic depth decay rule
-        if kv_compress_config is None:
-            kv_compress_config = {
-                'sampling': None,
-                'scale_factor': 1,
-                'kv_compress_layer': [],
-            }
-        self.blocks = nn.ModuleList([
-            PixArtMSBlock(
-                hidden_size, num_heads, mlp_ratio=mlp_ratio, drop_path=drop_path[i],
-                sampling=kv_compress_config['sampling'],
-                sr_ratio=int(kv_compress_config['scale_factor']) if i in kv_compress_config['kv_compress_layer'] else 1,
-                qk_norm=qk_norm,
-                dtype=dtype,
-                device=device,
-                operations=operations,
-            )
-            for i in range(depth)
-        ])
-        self.final_layer = T2IFinalLayer(
-            hidden_size, patch_size, self.out_channels, dtype=dtype, device=device, operations=operations
-        )
-
-    def forward_orig(self, x, timestep, y, mask=None, c_size=None, c_ar=None, **kwargs):
-        """
-        Original forward pass of PixArt.
-        x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
-        t: (N,) tensor of diffusion timesteps
-        y: (N, 1, 120, C) conditioning
-        ar: (N, 1): aspect ratio
-        cs: (N ,2) size conditioning for height/width
-        """
-        B, C, H, W = x.shape
-        c_res = (H + W) // 2
-        pe_interpolation = self.pe_interpolation
-        if pe_interpolation is None or self.pe_precision is not None:
-            # calculate pe_interpolation on-the-fly
-            pe_interpolation = round(c_res / (512/8.0), self.pe_precision or 0)
-
-        pos_embed = get_2d_sincos_pos_embed_torch(
-            self.hidden_size,
-            h=(H // self.patch_size),
-            w=(W // self.patch_size),
-            pe_interpolation=pe_interpolation,
-            base_size=((round(c_res / 64) * 64) // self.patch_size),
-            device=x.device,
-            dtype=x.dtype,
-        ).unsqueeze(0)
-
-        x = self.x_embedder(x) + pos_embed  # (N, T, D), where T = H * W / patch_size ** 2
-        t = self.t_embedder(timestep, x.dtype)  # (N, D)
-
-        if self.micro_conditioning and (c_size is not None and c_ar is not None):
-            bs = x.shape[0]
-            c_size = self.csize_embedder(c_size, bs)  # (N, D)
-            c_ar = self.ar_embedder(c_ar, bs)  # (N, D)
-            t = t + torch.cat([c_size, c_ar], dim=1)
-
-        t0 = self.t_block(t)
-        y = self.y_embedder(y, self.training)  # (N, D)
-
-        if mask is not None:
-            if mask.shape[0] != y.shape[0]:
-                mask = mask.repeat(y.shape[0] // mask.shape[0], 1)
-            mask = mask.squeeze(1).squeeze(1)
-            y = y.squeeze(1).masked_select(mask.unsqueeze(-1) != 0).view(1, -1, x.shape[-1])
-            y_lens = mask.sum(dim=1).tolist()
-        else:
-            y_lens = None
-            y = y.squeeze(1).view(1, -1, x.shape[-1])
-        for block in self.blocks:
-            x = block(x, y, t0, y_lens, (H, W), **kwargs)  # (N, T, D)
-
-        x = self.final_layer(x, t)  # (N, T, patch_size ** 2 * out_channels)
-        x = self.unpatchify(x, H, W)  # (N, out_channels, H, W)
-
-        return x
-
-    def forward(self, x, timesteps, context, c_size=None, c_ar=None, **kwargs):
-        B, C, H, W = x.shape
-
-        # Fallback for missing microconds
-        if self.micro_conditioning:
-            if c_size is None:
-                c_size = torch.tensor([H*8, W*8], dtype=x.dtype, device=x.device).repeat(B, 1)
-
-            if c_ar is None:
-                c_ar = torch.tensor([H/W], dtype=x.dtype, device=x.device).repeat(B, 1)
-
-        ## Still accepts the input w/o that dim but returns garbage
-        if len(context.shape) == 3:
-            context = context.unsqueeze(1)
-
-        ## run original forward pass
-        out = self.forward_orig(x, timesteps, context, c_size=c_size, c_ar=c_ar)
-
-        ## only return EPS
-        if self.pred_sigma:
-            return out[:, :self.in_channels]
-        return out
-
-    def unpatchify(self, x, h, w):
-        """
-        x: (N, T, patch_size**2 * C)
-        imgs: (N, H, W, C)
-        """
-        c = self.out_channels
-        p = self.x_embedder.patch_size[0]
-        h = h // self.patch_size
-        w = w // self.patch_size
-        assert h * w == x.shape[1]
-
-        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
+# Based on:
+# https://github.com/PixArt-alpha/PixArt-alpha [Apache 2.0 license]
+# https://github.com/PixArt-alpha/PixArt-sigma [Apache 2.0 license]
+import torch
+import torch.nn as nn
+
+from .blocks import (
+    t2i_modulate,
+    CaptionEmbedder,
+    AttentionKVCompress,
+    MultiHeadCrossAttention,
+    T2IFinalLayer,
+    SizeEmbedder,
+)
+from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder, PatchEmbed, Mlp, get_1d_sincos_pos_embed_from_grid_torch
+
+
+def get_2d_sincos_pos_embed_torch(embed_dim, w, h, pe_interpolation=1.0, base_size=16, device=None, dtype=torch.float32):
+    grid_h, grid_w = torch.meshgrid(
+        torch.arange(h, device=device, dtype=dtype) / (h/base_size) / pe_interpolation,
+        torch.arange(w, device=device, dtype=dtype) / (w/base_size) / pe_interpolation,
+        indexing='ij'
+    )
+    emb_h = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_h, device=device, dtype=dtype)
+    emb_w = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_w, device=device, dtype=dtype)
+    emb = torch.cat([emb_w, emb_h], dim=1)  # (H*W, D)
+    return emb
+
+class PixArtMSBlock(nn.Module):
+    """
+    A PixArt block with adaptive layer norm zero (adaLN-Zero) conditioning.
+    """
+    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, drop_path=0., input_size=None,
+                 sampling=None, sr_ratio=1, qk_norm=False, dtype=None, device=None, operations=None, **block_kwargs):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.attn = AttentionKVCompress(
+            hidden_size, num_heads=num_heads, qkv_bias=True, sampling=sampling, sr_ratio=sr_ratio,
+            qk_norm=qk_norm, dtype=dtype, device=device, operations=operations, **block_kwargs
+        )
+        self.cross_attn = MultiHeadCrossAttention(
+            hidden_size, num_heads, dtype=dtype, device=device, operations=operations, **block_kwargs
+        )
+        self.norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        # to be compatible with lower version pytorch
+        approx_gelu = lambda: nn.GELU(approximate="tanh")
+        self.mlp = Mlp(
+            in_features=hidden_size, hidden_features=int(hidden_size * mlp_ratio), act_layer=approx_gelu,
+            dtype=dtype, device=device, operations=operations
+        )
+        self.scale_shift_table = nn.Parameter(torch.randn(6, hidden_size) / hidden_size ** 0.5)
+
+    def forward(self, x, y, t, mask=None, HW=None, **kwargs):
+        B, N, C = x.shape
+
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None].to(dtype=x.dtype, device=x.device) + t.reshape(B, 6, -1)).chunk(6, dim=1)
+        x = x + (gate_msa * self.attn(t2i_modulate(self.norm1(x), shift_msa, scale_msa), HW=HW))
+        x = x + self.cross_attn(x, y, mask)
+        x = x + (gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp)))
+
+        return x
+
+
+### Core PixArt Model ###
+class PixArtMS(nn.Module):
+    """
+    Diffusion model with a Transformer backbone.
+    """
+    def __init__(
+            self,
+            input_size=32,
+            patch_size=2,
+            in_channels=4,
+            hidden_size=1152,
+            depth=28,
+            num_heads=16,
+            mlp_ratio=4.0,
+            class_dropout_prob=0.1,
+            learn_sigma=True,
+            pred_sigma=True,
+            drop_path: float = 0.,
+            caption_channels=4096,
+            pe_interpolation=None,
+            pe_precision=None,
+            config=None,
+            model_max_length=120,
+            micro_condition=True,
+            qk_norm=False,
+            kv_compress_config=None,
+            dtype=None,
+            device=None,
+            operations=None,
+            **kwargs,
+    ):
+        nn.Module.__init__(self)
+        self.dtype = dtype
+        self.pred_sigma = pred_sigma
+        self.in_channels = in_channels
+        self.out_channels = in_channels * 2 if pred_sigma else in_channels
+        self.patch_size = patch_size
+        self.num_heads = num_heads
+        self.pe_interpolation = pe_interpolation
+        self.pe_precision = pe_precision
+        self.hidden_size = hidden_size
+        self.depth = depth
+
+        approx_gelu = lambda: nn.GELU(approximate="tanh")
+        self.t_block = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(hidden_size, 6 * hidden_size, bias=True, dtype=dtype, device=device)
+        )
+        self.x_embedder = PatchEmbed(
+            patch_size=patch_size,
+            in_chans=in_channels,
+            embed_dim=hidden_size,
+            bias=True,
+            dtype=dtype,
+            device=device,
+            operations=operations
+        )
+        self.t_embedder = TimestepEmbedder(
+            hidden_size, dtype=dtype, device=device, operations=operations,
+        )
+        self.y_embedder = CaptionEmbedder(
+            in_channels=caption_channels, hidden_size=hidden_size, uncond_prob=class_dropout_prob,
+            act_layer=approx_gelu, token_num=model_max_length,
+            dtype=dtype, device=device, operations=operations,
+        )
+
+        self.micro_conditioning = micro_condition
+        if self.micro_conditioning:
+            self.csize_embedder = SizeEmbedder(hidden_size//3, dtype=dtype, device=device, operations=operations)
+            self.ar_embedder = SizeEmbedder(hidden_size//3, dtype=dtype, device=device, operations=operations)
+
+        # For fixed sin-cos embedding:
+        # num_patches = (input_size // patch_size) * (input_size // patch_size)
+        # self.base_size = input_size // self.patch_size
+        # self.register_buffer("pos_embed", torch.zeros(1, num_patches, hidden_size))
+
+        drop_path = [x.item() for x in torch.linspace(0, drop_path, depth)]  # stochastic depth decay rule
+        if kv_compress_config is None:
+            kv_compress_config = {
+                'sampling': None,
+                'scale_factor': 1,
+                'kv_compress_layer': [],
+            }
+        self.blocks = nn.ModuleList([
+            PixArtMSBlock(
+                hidden_size, num_heads, mlp_ratio=mlp_ratio, drop_path=drop_path[i],
+                sampling=kv_compress_config['sampling'],
+                sr_ratio=int(kv_compress_config['scale_factor']) if i in kv_compress_config['kv_compress_layer'] else 1,
+                qk_norm=qk_norm,
+                dtype=dtype,
+                device=device,
+                operations=operations,
+            )
+            for i in range(depth)
+        ])
+        self.final_layer = T2IFinalLayer(
+            hidden_size, patch_size, self.out_channels, dtype=dtype, device=device, operations=operations
+        )
+
+    def forward_orig(self, x, timestep, y, mask=None, c_size=None, c_ar=None, **kwargs):
+        """
+        Original forward pass of PixArt.
+        x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
+        t: (N,) tensor of diffusion timesteps
+        y: (N, 1, 120, C) conditioning
+        ar: (N, 1): aspect ratio
+        cs: (N ,2) size conditioning for height/width
+        """
+        B, C, H, W = x.shape
+        c_res = (H + W) // 2
+        pe_interpolation = self.pe_interpolation
+        if pe_interpolation is None or self.pe_precision is not None:
+            # calculate pe_interpolation on-the-fly
+            pe_interpolation = round(c_res / (512/8.0), self.pe_precision or 0)
+
+        pos_embed = get_2d_sincos_pos_embed_torch(
+            self.hidden_size,
+            h=(H // self.patch_size),
+            w=(W // self.patch_size),
+            pe_interpolation=pe_interpolation,
+            base_size=((round(c_res / 64) * 64) // self.patch_size),
+            device=x.device,
+            dtype=x.dtype,
+        ).unsqueeze(0)
+
+        x = self.x_embedder(x) + pos_embed  # (N, T, D), where T = H * W / patch_size ** 2
+        t = self.t_embedder(timestep, x.dtype)  # (N, D)
+
+        if self.micro_conditioning and (c_size is not None and c_ar is not None):
+            bs = x.shape[0]
+            c_size = self.csize_embedder(c_size, bs)  # (N, D)
+            c_ar = self.ar_embedder(c_ar, bs)  # (N, D)
+            t = t + torch.cat([c_size, c_ar], dim=1)
+
+        t0 = self.t_block(t)
+        y = self.y_embedder(y, self.training)  # (N, D)
+
+        if mask is not None:
+            if mask.shape[0] != y.shape[0]:
+                mask = mask.repeat(y.shape[0] // mask.shape[0], 1)
+            mask = mask.squeeze(1).squeeze(1)
+            y = y.squeeze(1).masked_select(mask.unsqueeze(-1) != 0).view(1, -1, x.shape[-1])
+            y_lens = mask.sum(dim=1).tolist()
+        else:
+            y_lens = None
+            y = y.squeeze(1).view(1, -1, x.shape[-1])
+        for block in self.blocks:
+            x = block(x, y, t0, y_lens, (H, W), **kwargs)  # (N, T, D)
+
+        x = self.final_layer(x, t)  # (N, T, patch_size ** 2 * out_channels)
+        x = self.unpatchify(x, H, W)  # (N, out_channels, H, W)
+
+        return x
+
+    def forward(self, x, timesteps, context, c_size=None, c_ar=None, **kwargs):
+        B, C, H, W = x.shape
+
+        # Fallback for missing microconds
+        if self.micro_conditioning:
+            if c_size is None:
+                c_size = torch.tensor([H*8, W*8], dtype=x.dtype, device=x.device).repeat(B, 1)
+
+            if c_ar is None:
+                c_ar = torch.tensor([H/W], dtype=x.dtype, device=x.device).repeat(B, 1)
+
+        ## Still accepts the input w/o that dim but returns garbage
+        if len(context.shape) == 3:
+            context = context.unsqueeze(1)
+
+        ## run original forward pass
+        out = self.forward_orig(x, timesteps, context, c_size=c_size, c_ar=c_ar)
+
+        ## only return EPS
+        if self.pred_sigma:
+            return out[:, :self.in_channels]
+        return out
+
+    def unpatchify(self, x, h, w):
+        """
+        x: (N, T, patch_size**2 * C)
+        imgs: (N, H, W, C)
+        """
+        c = self.out_channels
+        p = self.x_embedder.patch_size[0]
+        h = h // self.patch_size
+        w = w // self.patch_size
+        assert h * w == x.shape[1]
+
+        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

comfy/samplers.py

@@ -720,7 +720,7 @@ KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_c
                   "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_2s_ancestral_cfg_pp", "dpmpp_sde", "dpmpp_sde_gpu",
                   "dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm",
                   "ipndm", "ipndm_v", "deis", "res_multistep", "res_multistep_cfg_pp", "res_multistep_ancestral", "res_multistep_ancestral_cfg_pp",
-                  "gradient_estimation", "gradient_estimation_cfg_pp", "er_sde", "seeds_2", "seeds_3"]
+                  "gradient_estimation", "gradient_estimation_cfg_pp", "er_sde", "seeds_2", "seeds_3", "sa_solver", "sa_solver_pece"]
 
 class KSAMPLER(Sampler):
     def __init__(self, sampler_function, extra_options={}, inpaint_options={}):

comfy/text_encoders/pixart_t5.py

@@ -1,42 +1,42 @@
-import os
-
-from comfy import sd1_clip
-import comfy.text_encoders.t5
-import comfy.text_encoders.sd3_clip
-from comfy.sd1_clip import gen_empty_tokens
-
-from transformers import T5TokenizerFast
-
-class T5XXLModel(comfy.text_encoders.sd3_clip.T5XXLModel):
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
-
-    def gen_empty_tokens(self, special_tokens, *args, **kwargs):
-        # PixArt expects the negative to be all pad tokens
-        special_tokens = special_tokens.copy()
-        special_tokens.pop("end")
-        return gen_empty_tokens(special_tokens, *args, **kwargs)
-
-class PixArtT5XXL(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None, model_options={}):
-        super().__init__(device=device, dtype=dtype, name="t5xxl", clip_model=T5XXLModel, model_options=model_options)
-
-class T5XXLTokenizer(sd1_clip.SDTokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer")
-        super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_data=tokenizer_data) # no padding
-
-class PixArtTokenizer(sd1_clip.SD1Tokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer)
-
-def pixart_te(dtype_t5=None, t5xxl_scaled_fp8=None):
-    class PixArtTEModel_(PixArtT5XXL):
-        def __init__(self, device="cpu", dtype=None, model_options={}):
-            if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
-                model_options = model_options.copy()
-                model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
-            if dtype is None:
-                dtype = dtype_t5
-            super().__init__(device=device, dtype=dtype, model_options=model_options)
-    return PixArtTEModel_
+import os
+
+from comfy import sd1_clip
+import comfy.text_encoders.t5
+import comfy.text_encoders.sd3_clip
+from comfy.sd1_clip import gen_empty_tokens
+
+from transformers import T5TokenizerFast
+
+class T5XXLModel(comfy.text_encoders.sd3_clip.T5XXLModel):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+    def gen_empty_tokens(self, special_tokens, *args, **kwargs):
+        # PixArt expects the negative to be all pad tokens
+        special_tokens = special_tokens.copy()
+        special_tokens.pop("end")
+        return gen_empty_tokens(special_tokens, *args, **kwargs)
+
+class PixArtT5XXL(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, name="t5xxl", clip_model=T5XXLModel, model_options=model_options)
+
+class T5XXLTokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer")
+        super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_data=tokenizer_data) # no padding
+
+class PixArtTokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer)
+
+def pixart_te(dtype_t5=None, t5xxl_scaled_fp8=None):
+    class PixArtTEModel_(PixArtT5XXL):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
+                model_options = model_options.copy()
+                model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
+            if dtype is None:
+                dtype = dtype_t5
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return PixArtTEModel_

comfy/utils.py

@@ -998,11 +998,12 @@ def set_progress_bar_global_hook(function):
     PROGRESS_BAR_HOOK = function
 
 class ProgressBar:
-    def __init__(self, total):
+    def __init__(self, total, node_id=None):
         global PROGRESS_BAR_HOOK
         self.total = total
         self.current = 0
         self.hook = PROGRESS_BAR_HOOK
+        self.node_id = node_id
 
     def update_absolute(self, value, total=None, preview=None):
         if total is not None:
@@ -1011,7 +1012,7 @@ class ProgressBar:
             value = self.total
         self.current = value
         if self.hook is not None:
-            self.hook(self.current, self.total, preview)
+            self.hook(self.current, self.total, preview, node_id=self.node_id)
 
     def update(self, value):
         self.update_absolute(self.current + value)