diff --git a/xfuser/model_executor/layers/ltx2/__init__.py b/xfuser/model_executor/layers/ltx2/__init__.py
new file mode 100644
index 00000000..e69de29b
diff --git a/xfuser/model_executor/layers/ltx2/diffusion_decoder.py b/xfuser/model_executor/layers/ltx2/diffusion_decoder.py
new file mode 100644
index 00000000..bf8ab442
--- /dev/null
+++ b/xfuser/model_executor/layers/ltx2/diffusion_decoder.py
@@ -0,0 +1,268 @@
+"""Tile-parallel diffusion VAE decode for LTX-2.5 across 2/4/8 GPUs.
+
+The stock ``LTX2VideoDiffusionDecoderModel.tiled_decode`` runs a triple-nested loop over
+``(temporal × height × width)`` tiles, each independently calling ``forward_stage_4``
++ ``denoise`` (8 neighbourhood-attention blocks each).
+
+``xFuserLTX2VideoDiffusionDecoderWrapper`` distributes those tiles across the SP group
+via round-robin ownership, then does one small all_reduce for shape metadata followed by
+per-tile broadcasts to reassemble the full video on every rank.  Communication is a
+single collective per tile.
+
+Tile geometry for the default 1024×1536×121 config: 12 tiles, adequate for 2/4/8 GPUs.
+"""
+from __future__ import annotations
+
+import math
+
+import torch
+import torch.distributed
+
+from diffusers import LTX2VideoDiffusionDecoderModel
+from diffusers.models.autoencoders.ltx2_diffusion_decoder import _tile_intervals
+from diffusers.utils.torch_utils import randn_tensor
+
+from xfuser.core.distributed import (
+    get_sequence_parallel_rank,
+    get_sequence_parallel_world_size,
+    get_sp_group,
+)
+
+
+class xFuserLTX2VideoDiffusionDecoderWrapper(LTX2VideoDiffusionDecoderModel):
+    """``LTX2VideoDiffusionDecoderModel`` with tile-parallel decode across the SP group.
+    Drop-in replacement: use ``from_pretrained`` on this class exactly as on the base
+    class.  The runner activates tile-parallel decode by:
+    1. Setting ``decoder._parallel_decode = True`` (done when ``--use_parallel_vae`` is on).
+    2. Calling ``decoder.enable_tiling()`` so that ``decode()`` dispatches to ``tiled_decode``.
+    When ``_parallel_decode`` is False, or ``sp_world_size == 1``, or the tile count is
+    less than 2, ``tiled_decode`` delegates to the stock implementation unchanged.
+    """
+
+    # Set to True by the runner when --use_parallel_vae is on.
+    _parallel_decode: bool = False
+
+    def tiled_decode(
+        self,
+        z: torch.Tensor,
+        generator: torch.Generator | None = None,
+        num_inference_steps: int | None = None,
+    ) -> torch.Tensor:
+        """Decode with tiles distributed across SP ranks.
+        Each rank owns a round-robin subset of tiles.  Owned tiles are computed locally;
+        all ranks gather all tiles via a shape-metadata all_reduce + per-tile broadcast,
+        then run the stock blend/assembly loop to produce the full video.
+        Noise determinism:
+        - *Shipping 1-step x0 path*: each tile is seeded from ``base_seed XOR f(t,h,w)``,
+          making noise rank-invariant without requiring non-owners to know tile shapes.
+          Output differs from single-GPU stock (different per-tile seeds) but is an
+          equally valid sample.
+        - *Multi-step path*: the full noise canvas is drawn identically on all ranks from
+          a shared generator seeded by ``base_seed``, then each owner slices its region.
+        """
+        sp_world_size = get_sequence_parallel_world_size()
+        if not self._parallel_decode or sp_world_size <= 1:
+            return super().tiled_decode(z, generator=generator,
+                                        num_inference_steps=num_inference_steps)
+
+        sp_group = get_sp_group()
+        sp_rank = get_sequence_parallel_rank()
+
+        # Tile geometry
+        decoder = self.decoder
+        num_inference_steps = num_inference_steps or decoder.default_num_inference_steps
+        batch_size = z.shape[0]
+        patch_size = decoder.patch_size
+
+        upsample_stride = decoder.upsamples[-1].stride
+        scale_t = upsample_stride[0]
+        scale_h = upsample_stride[1] * patch_size
+        scale_w = upsample_stride[2] * patch_size
+
+        tile_t  = self.tile_sample_min_num_frames // scale_t
+        stride_t = self.tile_sample_stride_num_frames // scale_t
+        tile_h  = self.tile_sample_min_height // scale_h
+        stride_h = self.tile_sample_stride_height // scale_h
+        tile_w  = self.tile_sample_min_width // scale_w
+        stride_w = self.tile_sample_stride_width // scale_w
+
+        min_sizes = [
+            max(k4, -(-k5 // s))
+            for k4, k5, s in zip(
+                self.config.decoder_stage_kernels[-1],
+                self.config.decoder_stage5_kernel,
+                upsample_stride,
+            )
+        ]
+
+        # Cheap deterministic stages run on the full volume on every rank.
+        # This avoids communicating the intermediate features tensor and keeps the
+        # tile-grid computation identical across ranks.
+        features = decoder.forward_stages_1_to_3(z)
+        ghost_frames = decoder.trailing_pad_latent_frames * math.prod(
+            up.stride[0] for up in decoder.upsamples[:-1]
+        )
+        num_frames = features.shape[1] - ghost_frames
+        height, width = features.shape[2], features.shape[3]
+
+        temporal_tiles = _tile_intervals(num_frames, tile_t, stride_t, min_sizes[0])
+        height_tiles   = _tile_intervals(height,      tile_h, stride_h, min_sizes[1])
+        width_tiles    = _tile_intervals(width,        tile_w, stride_w, min_sizes[2])
+
+        blend_frames = (tile_t - stride_t) * scale_t
+        blend_height = (tile_h - stride_h) * scale_h
+        blend_width  = (tile_w - stride_w) * scale_w
+
+        n_tiles = len(temporal_tiles) * len(height_tiles) * len(width_tiles)
+
+        # Fall back to stock when there are too few tiles to distribute.
+        if n_tiles < 2:
+            return super().tiled_decode(z, generator=generator,
+                                        num_inference_steps=num_inference_steps)
+
+        # Noise setup
+        single_step_x0 = (num_inference_steps == 1 and decoder.model_output_type == "x0")
+
+        # Draw one integer from the shared generator to derive a base seed.
+        # Generator state is guaranteed identical across SP ranks at decode entry, so
+        # all ranks draw the same value, advancing the generator identically.
+        if generator is not None:
+            base_seed_t = torch.randint(0, 2**31, (1,), generator=generator, device=z.device)
+            base_seed = int(base_seed_t.item())
+        else:
+            # generator=None: use a fixed seed (in practice the runner always provides one)
+            base_seed = 0
+
+        x_t_full: torch.Tensor | None = None
+        if not single_step_x0:
+            pixel_frames = num_frames * scale_t - (1 if scale_t == 2 else 0)
+            shared_gen = torch.Generator(device=z.device).manual_seed(base_seed)
+            x_t_full = randn_tensor(
+                (batch_size, decoder.out_channels,
+                 pixel_frames, height * scale_h, width * scale_w),
+                generator=shared_gen,
+                device=z.device,
+                dtype=z.dtype,
+            )
+
+        # Compute owned tiles
+        # shape_table[i] = (T_px, H_px, W_px) for tile i; zero for non-owned tiles.
+        # After all_reduce(SUM) every rank knows all tile pixel shapes.
+        shape_table = torch.zeros(n_tiles, 3, dtype=torch.int64, device=z.device)
+        tile_outputs: dict[int, torch.Tensor] = {}
+
+        tile_idx = 0
+        for t_idx, (t0, t1) in enumerate(temporal_tiles):
+            is_origin   = (t0 == 0)
+            is_trailing = (t1 == num_frames)
+            # Trailing temporal tile carries ghost frames into forward_stage_4.
+            feature_t1 = features.shape[1] if is_trailing else t1
+
+            for h_idx, (h0, h1) in enumerate(height_tiles):
+                for w_idx, (w0, w1) in enumerate(width_tiles):
+                    owner = tile_idx % sp_world_size
+
+                    if owner == sp_rank:
+                        context = decoder.forward_stage_4(
+                            features[:, t0:feature_t1, h0:h1, w0:w1],
+                            drop_leading_frame=is_origin,
+                            crop_trailing_ghost=is_trailing,
+                        )
+                        tile_pixel_shape = (
+                            batch_size,
+                            decoder.out_channels,
+                            context.shape[1],
+                            context.shape[2] * patch_size,
+                            context.shape[3] * patch_size,
+                        )
+
+                        if single_step_x0:
+                            # Per-tile deterministic seeding: owner-only draw, no
+                            # cross-rank RNG synchronisation needed.
+                            # Small integer hash keeps seeds well-separated.
+                            tile_seed = (base_seed ^ (t_idx * 0x3_D6F1
+                                                      + h_idx * 0x1_E3
+                                                      + w_idx * 0x7)) & 0x7FFF_FFFF
+                            tile_gen = torch.Generator(device=z.device).manual_seed(tile_seed)
+                            x_t = randn_tensor(tile_pixel_shape, generator=tile_gen,
+                                               device=z.device, dtype=z.dtype)
+                        else:
+                            # Slice from the shared noise canvas (drawn identically on all
+                            # ranks above).
+                            pixel_t0 = (t0 * scale_t
+                                        - (1 if not is_origin and scale_t == 2 else 0))
+                            x_t = x_t_full[
+                                :, :,
+                                pixel_t0 : pixel_t0 + tile_pixel_shape[2],
+                                h0 * scale_h : h0 * scale_h + tile_pixel_shape[3],
+                                w0 * scale_w : w0 * scale_w + tile_pixel_shape[4],
+                            ]
+
+                        out = decoder.denoise(context, x_t, num_inference_steps)
+                        tile_outputs[tile_idx] = out
+                        shape_table[tile_idx, 0] = out.shape[2]  # T_px
+                        shape_table[tile_idx, 1] = out.shape[3]  # H_px
+                        shape_table[tile_idx, 2] = out.shape[4]  # W_px
+
+                    tile_idx += 1
+
+        # Gather: shape metadata + per-tile broadcast
+        # all_reduce(SUM): safe because each element is non-zero on exactly one rank.
+        torch.distributed.all_reduce(
+            shape_table,
+            op=torch.distributed.ReduceOp.SUM,
+            group=sp_group.device_group,
+        )
+
+        gathered: dict[int, torch.Tensor] = {}
+        for i in range(n_tiles):
+            owner = i % sp_world_size
+            T_px = int(shape_table[i, 0])
+            H_px = int(shape_table[i, 1])
+            W_px = int(shape_table[i, 2])
+            if sp_rank == owner:
+                buf = tile_outputs[i]
+            else:
+                buf = torch.empty(
+                    batch_size, decoder.out_channels, T_px, H_px, W_px,
+                    dtype=z.dtype, device=z.device,
+                )
+            # src is the SP-group-local rank (0 … sp_world_size-1).
+            sp_group.broadcast(buf, src=owner)
+            gathered[i] = buf
+
+        # Assembly — identical to stock tiled_decode
+        frame_groups: list[torch.Tensor] = []
+        tile_idx = 0
+        for _t_idx, (_t0, _t1) in enumerate(temporal_tiles):
+            rows: list[list[torch.Tensor]] = []
+            for _h_idx, _ in enumerate(height_tiles):
+                row: list[torch.Tensor] = []
+                for _w_idx, _ in enumerate(width_tiles):
+                    row.append(gathered[tile_idx])
+                    tile_idx += 1
+                rows.append(row)
+
+            result_rows = []
+            for i, row in enumerate(rows):
+                result_row = []
+                for j, tile in enumerate(row):
+                    if i > 0:
+                        tile = self.blend_v(rows[i - 1][j], tile, blend_height)
+                    if j > 0:
+                        tile = self.blend_h(row[j - 1], tile, blend_width)
+                    keep_height = stride_h * scale_h if i < len(rows) - 1 else tile.shape[3]
+                    keep_width  = stride_w * scale_w if j < len(row)  - 1 else tile.shape[4]
+                    result_row.append(tile[:, :, :, :keep_height, :keep_width])
+                result_rows.append(torch.cat(result_row, dim=4))
+            frame_groups.append(torch.cat(result_rows, dim=3))
+
+        result = []
+        for k, group in enumerate(frame_groups):
+            if k > 0:
+                group = self.blend_t(frame_groups[k - 1], group, blend_frames)
+            if k < len(frame_groups) - 1:
+                keep_frames = stride_t * scale_t - (1 if k == 0 and scale_t == 2 else 0)
+                group = group[:, :, :keep_frames]
+            result.append(group)
+        return torch.cat(result, dim=2)
\ No newline at end of file
diff --git a/xfuser/model_executor/layers/ltx2_na3d_eager_attn.py b/xfuser/model_executor/layers/ltx2/na3d_eager_attn.py
similarity index 100%
rename from xfuser/model_executor/layers/ltx2_na3d_eager_attn.py
rename to xfuser/model_executor/layers/ltx2/na3d_eager_attn.py
diff --git a/xfuser/model_executor/layers/norms.py b/xfuser/model_executor/layers/norms.py
new file mode 100644
index 00000000..43975fec
--- /dev/null
+++ b/xfuser/model_executor/layers/norms.py
@@ -0,0 +1,87 @@
+from __future__ import annotations
+
+import torch
+
+try:
+    import aiter as _aiter
+
+    # Wrapping aiter.rms_norm as a non-mutating custom op fixes graph capture
+    @torch.library.custom_op("xfuser::aiter_rms_norm", mutates_args=())
+    def _aiter_rms_norm(x: torch.Tensor, w: torch.Tensor, eps: float) -> torch.Tensor:
+        return _aiter.rms_norm(x, w, eps)
+
+    @_aiter_rms_norm.register_fake
+    def _(x, w, eps):
+        return torch.empty_like(x)
+
+    @_aiter_rms_norm.register_autograd
+    def _(ctx, grad_output):
+        # Inference only, backward is never called; return identity gradient.
+        return grad_output, None, None
+
+    _HAS_AITER = True
+except ImportError:
+    _HAS_AITER = False
+
+
+class _AITERRMSNorm(torch.nn.Module):
+    """Drop-in for diffusers RMSNorm using AITER's CK-Tile kernel.
+    Diffusers RMSNorm up-casts to float32 internally; AITER's kernel works
+    natively in bfloat16 and is faster.  Handles elementwise_affine=False
+    (no learned weight) via a registered ones buffer so AITER always receives a
+    weight tensor.
+    """
+
+    def __init__(self, weight: torch.nn.Parameter | None, eps: float, dim: int) -> None:
+        super().__init__()
+        if weight is not None:
+            self.weight = weight
+            self._use_ones = False
+        else:
+            self.register_buffer("_ones_weight", torch.ones(dim))
+            self._use_ones = True
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        shape = x.shape
+        w = self._ones_weight if self._use_ones else self.weight
+        out = _aiter_rms_norm(x.view(-1, shape[-1]), w, self.eps)
+        return out.view(shape)
+
+
+def _replace_rms_norms_with_aiter(model: torch.nn.Module) -> None:
+    """Walk ``model`` and replace diffusers RMSNorm modules with :class:`_AITERRMSNorm`.
+    Only replaces ``diffusers.models.normalization.RMSNorm`` (Python impl,
+    float32 up-cast), leaving ``torch.nn.RMSNorm`` (C++ impl) untouched.
+    Must be called BEFORE ``model.to(device)`` so buffers are placed correctly.
+    """
+    if not _HAS_AITER:
+        return
+    try:
+        from diffusers.models.normalization import RMSNorm as _DiffusersRMSNorm
+    except ImportError:
+        return
+
+    replacements: list[tuple[torch.nn.Module, str, _AITERRMSNorm]] = []
+    for _parent_name, parent_module in model.named_modules():
+        for child_name, child_module in parent_module.named_children():
+            if type(child_module) is _DiffusersRMSNorm and hasattr(child_module, "eps"):
+                if child_module.weight is not None:
+                    dim = child_module.weight.shape[0]
+                elif hasattr(child_module, "dim"):
+                    dim = (
+                        int(child_module.dim[0])
+                        if hasattr(child_module.dim, "__len__")
+                        else int(child_module.dim)
+                    )
+                else:
+                    continue
+                replacements.append(
+                    (
+                        parent_module,
+                        child_name,
+                        _AITERRMSNorm(child_module.weight, child_module.eps, dim),
+                    )
+                )
+    for parent, name, replacement in replacements:
+        setattr(parent, name, replacement)
\ No newline at end of file
diff --git a/xfuser/model_executor/models/runner_models/ltx.py b/xfuser/model_executor/models/runner_models/ltx.py
index 3b367fe2..92f63a57 100644
--- a/xfuser/model_executor/models/runner_models/ltx.py
+++ b/xfuser/model_executor/models/runner_models/ltx.py
@@ -445,6 +445,7 @@ class _xFuserLTX25VideoModelBase(xFuserModel):
         ulysses_degree=True,
         ring_degree=True,
         enable_tiling=True,
+        use_parallel_vae=True,
         use_fp8_gemms=True,
         use_fp4_gemms=True,
     )
@@ -497,12 +498,15 @@ def _load_model(self) -> DiffusionPipeline:
 
         # Diffusion decoder — replaces convolutional VAE decode for both distilled
         # and full model pipelines.
-        from diffusers import LTX2VideoDiffusionDecoderModel
         from diffusers.pipelines.ltx2.pipeline_ltx2_diffusion_decode import (
             LTX2VideoDiffusionDecodePipeline,
         )
 
-        diff_decoder = LTX2VideoDiffusionDecoderModel.from_pretrained(
+        from xfuser.model_executor.layers.ltx2.diffusion_decoder import (
+            xFuserLTX2VideoDiffusionDecoderWrapper,
+        )
+
+        diff_decoder = xFuserLTX2VideoDiffusionDecoderWrapper.from_pretrained(
             self.settings.model_name,
             subfolder="diffusion_decoder",
             torch_dtype=torch.bfloat16,
@@ -518,13 +522,13 @@ def _load_model(self) -> DiffusionPipeline:
             # NATTEN is not available
             # Fall back to tiled PyTorch SDPA
             # Works on CUDA, ROCm and CPU. Ported from LTX-2 EagerSdpaAttention.
-            from xfuser.model_executor.layers.ltx2_na3d_eager_attn import (
+            from xfuser.model_executor.layers.ltx2.na3d_eager_attn import (
                 LTX2VideoVaeEagerSdpaAttnProcessor,
             )
 
             diff_decoder.set_attn_processor(LTX2VideoVaeEagerSdpaAttnProcessor())
             log(
-                "Diffusion decoder: NATTEN unavailable; using Triton na3d attention fallback."
+                "Diffusion decoder: NATTEN unavailable; using tiled PyTorch SDPA fallback."
             )
         self.decode_pipe = LTX2VideoDiffusionDecodePipeline(
             diffusion_decoder=diff_decoder,
@@ -535,9 +539,16 @@ def _load_model(self) -> DiffusionPipeline:
 
     def _enable_options(self) -> None:
         super()._enable_options()
-        if self.config.enable_tiling:
+        # Tiling on the diffusion decoder is enabled when either --enable_tiling is set
+        # (for single-GPU memory savings) or --use_parallel_vae is set (required for
+        # tiled_decode dispatch, which is what activates tile-parallel decode).
+        if self.config.enable_tiling or self.config.use_parallel_vae:
             self.pipe.vae.enable_tiling()
             self.decode_pipe.diffusion_decoder.enable_tiling()
+        # Gate tile-parallel decode behind --use_parallel_vae.  The wrapper's tiled_decode
+        # checks this flag before distributing tiles across SP ranks.
+        if self.config.use_parallel_vae:
+            self.decode_pipe.diffusion_decoder._parallel_decode = True
 
     def _preprocess_args_images(self, input_args: dict) -> dict:
         input_args = super()._preprocess_args_images(input_args)
diff --git a/xfuser/model_executor/models/transformers/transformer_ltx2.py b/xfuser/model_executor/models/transformers/transformer_ltx2.py
index fabc13a4..8cc21890 100644
--- a/xfuser/model_executor/models/transformers/transformer_ltx2.py
+++ b/xfuser/model_executor/models/transformers/transformer_ltx2.py
@@ -18,6 +18,7 @@
     AttentionMaskWithMeta,
     make_attn_mask_with_meta,
 )
+from xfuser.model_executor.layers.norms import _replace_rms_norms_with_aiter
 from xfuser.model_executor.layers.usp import USP, attention
 
 
@@ -352,6 +353,11 @@ def __init__(
         self._enc_mask_cache: dict = {}
         self._audio_enc_mask_cache: dict = {}
 
+        # If AITER is available, replace diffusers RMSNorm (slow float32 cast)
+        # with AITER RMSNorm.
+        # Called BEFORE device move so buffers are placed correctly by model.to().
+        _replace_rms_norms_with_aiter(self)
+
         if perturbed_attn:
             attn_processor_cls = xFuserLTX2PerturbedAttnProcessor
         else: