up

Author: sayakpaul

docs/source/en/_toctree.yml

@@ -525,6 +525,8 @@
         title: Kandinsky 2.2
       - local: api/pipelines/kandinsky3
         title: Kandinsky 3
+      - local: api/pipelines/kandinsky5
+        title: Kandinsky 5
       - local: api/pipelines/kolors
         title: Kolors
       - local: api/pipelines/latent_consistency_models

docs/source/en/api/models/chroma_transformer.md

@@ -12,7 +12,7 @@ specific language governing permissions and limitations under the License.
 
 # ChromaTransformer2DModel
 
-A modified flux Transformer model from [Chroma](https://huggingface.co/lodestones/Chroma)
+A modified flux Transformer model from [Chroma](https://huggingface.co/lodestones/Chroma1-HD)
 
 ## ChromaTransformer2DModel
 

docs/source/en/api/pipelines/chroma.md

@@ -19,20 +19,21 @@ specific language governing permissions and limitations under the License.
 
 Chroma is a text to image generation model based on Flux.
 
-Original model checkpoints for Chroma can be found [here](https://huggingface.co/lodestones/Chroma).
+Original model checkpoints for Chroma can be found here:
+* High-resolution finetune: [lodestones/Chroma1-HD](https://huggingface.co/lodestones/Chroma1-HD)
+* Base model: [lodestones/Chroma1-Base](https://huggingface.co/lodestones/Chroma1-Base)
+* Original repo with progress checkpoints: [lodestones/Chroma](https://huggingface.co/lodestones/Chroma) (loading this repo with `from_pretrained` will load a Diffusers-compatible version of the `unlocked-v37` checkpoint)
 
 > [!TIP]
 > Chroma can use all the same optimizations as Flux.
 
 ## Inference
 
-The Diffusers version of Chroma is based on the [`unlocked-v37`](https://huggingface.co/lodestones/Chroma/blob/main/chroma-unlocked-v37.safetensors) version of the original model, which is available in the [Chroma repository](https://huggingface.co/lodestones/Chroma).
-
 ```python
 import torch
 from diffusers import ChromaPipeline
 
-pipe = ChromaPipeline.from_pretrained("lodestones/Chroma", torch_dtype=torch.bfloat16)
+pipe = ChromaPipeline.from_pretrained("lodestones/Chroma1-HD", torch_dtype=torch.bfloat16)
 pipe.enable_model_cpu_offload()
 
 prompt = [
@@ -63,10 +64,10 @@ Then run the following example
 import torch
 from diffusers import ChromaTransformer2DModel, ChromaPipeline
 
-model_id = "lodestones/Chroma"
+model_id = "lodestones/Chroma1-HD"
 dtype = torch.bfloat16
 
-transformer = ChromaTransformer2DModel.from_single_file("https://huggingface.co/lodestones/Chroma/blob/main/chroma-unlocked-v37.safetensors", torch_dtype=dtype)
+transformer = ChromaTransformer2DModel.from_single_file("https://huggingface.co/lodestones/Chroma1-HD/blob/main/Chroma1-HD.safetensors", torch_dtype=dtype)
 
 pipe = ChromaPipeline.from_pretrained(model_id, transformer=transformer, torch_dtype=dtype)
 pipe.enable_model_cpu_offload()

docs/source/en/api/pipelines/kandinsky5.md

@@ -0,0 +1,149 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Kandinsky 5.0
+
+Kandinsky 5.0 is created by the Kandinsky team: Alexey Letunovskiy, Maria Kovaleva, Ivan Kirillov, Lev Novitskiy, Denis Koposov, Dmitrii Mikhailov, Anna Averchenkova, Andrey Shutkin, Julia Agafonova, Olga Kim, Anastasiia Kargapoltseva, Nikita Kiselev, Anna Dmitrienko,  Anastasia Maltseva, Kirill Chernyshev, Ilia Vasiliev, Viacheslav Vasilev, Vladimir Polovnikov, Yury Kolabushin, Alexander Belykh, Mikhail Mamaev, Anastasia Aliaskina, Tatiana Nikulina, Polina Gavrilova, Vladimir Arkhipkin, Vladimir Korviakov, Nikolai Gerasimenko, Denis Parkhomenko, Denis Dimitrov
+
+
+Kandinsky 5.0 is a family of diffusion models for Video & Image generation. Kandinsky 5.0 T2V Lite is a lightweight video generation model (2B parameters) that ranks #1 among open-source models in its class. It outperforms larger models and offers the best understanding of Russian concepts in the open-source ecosystem.
+
+The model introduces several key innovations:
+- **Latent diffusion pipeline** with **Flow Matching** for improved training stability
+- **Diffusion Transformer (DiT)** as the main generative backbone with cross-attention to text embeddings
+- Dual text encoding using **Qwen2.5-VL** and **CLIP** for comprehensive text understanding
+- **HunyuanVideo 3D VAE** for efficient video encoding and decoding
+- **Sparse attention mechanisms** (NABLA) for efficient long-sequence processing
+
+The original codebase can be found at [ai-forever/Kandinsky-5](https://github.com/ai-forever/Kandinsky-5).
+
+> [!TIP]
+> Check out the [AI Forever](https://huggingface.co/ai-forever) organization on the Hub for the official model checkpoints for text-to-video generation, including pretrained, SFT, no-CFG, and distilled variants.
+
+## Available Models
+
+Kandinsky 5.0 T2V Lite comes in several variants optimized for different use cases:
+
+| model_id | Description | Use Cases |
+|------------|-------------|-----------|
+| **ai-forever/Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers** | 5 second Supervised Fine-Tuned model | Highest generation quality |
+| **ai-forever/Kandinsky-5.0-T2V-Lite-sft-10s-Diffusers** | 10 second Supervised Fine-Tuned model | Highest generation quality |
+| **ai-forever/Kandinsky-5.0-T2V-Lite-nocfg-5s-Diffusers** | 5 second Classifier-Free Guidance distilled | 2× faster inference |
+| **ai-forever/Kandinsky-5.0-T2V-Lite-nocfg-10s-Diffusers** | 10 second Classifier-Free Guidance distilled | 2× faster inference |
+| **ai-forever/Kandinsky-5.0-T2V-Lite-distilled16steps-5s-Diffusers** | 5 second Diffusion distilled to 16 steps | 6× faster inference, minimal quality loss |
+| **ai-forever/Kandinsky-5.0-T2V-Lite-distilled16steps-10s-Diffusers** | 10 second Diffusion distilled to 16 steps | 6× faster inference, minimal quality loss |
+| **ai-forever/Kandinsky-5.0-T2V-Lite-pretrain-5s-Diffusers** | 5 second Base pretrained model | Research and fine-tuning |
+| **ai-forever/Kandinsky-5.0-T2V-Lite-pretrain-10s-Diffusers** | 10 second Base pretrained model | Research and fine-tuning |
+
+All models are available in 5-second and 10-second video generation versions.
+
+## Kandinsky5T2VPipeline
+
+[[autodoc]] Kandinsky5T2VPipeline
+    - all
+    - __call__
+
+## Usage Examples
+
+### Basic Text-to-Video Generation
+
+```python
+import torch
+from diffusers import Kandinsky5T2VPipeline
+from diffusers.utils import export_to_video
+
+# Load the pipeline
+model_id = "ai-forever/Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers"
+pipe = Kandinsky5T2VPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16)
+pipe = pipe.to("cuda")
+
+# Generate video
+prompt = "A cat and a dog baking a cake together in a kitchen."
+negative_prompt = "Static, 2D cartoon, cartoon, 2d animation, paintings, images, worst quality, low quality, ugly, deformed, walking backwards"
+
+output = pipe(
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=512,
+    width=768,
+    num_frames=121,  # ~5 seconds at 24fps
+    num_inference_steps=50,
+    guidance_scale=5.0,
+).frames[0]
+
+export_to_video(output, "output.mp4", fps=24, quality=9)
+```
+
+### 10 second Models
+**⚠️ Warning!** all 10 second models should be used with Flex attention and max-autotune-no-cudagraphs compilation:
+
+```python
+pipe = Kandinsky5T2VPipeline.from_pretrained(
+    "ai-forever/Kandinsky-5.0-T2V-Lite-sft-10s-Diffusers", 
+    torch_dtype=torch.bfloat16
+)
+pipe = pipe.to("cuda")
+
+pipe.transformer.set_attention_backend(
+    "flex"
+)                                       # <--- Set attention backend to Flex
+pipe.transformer.compile(
+    mode="max-autotune-no-cudagraphs", 
+    dynamic=True
+)                                       # <--- Compile with max-autotune-no-cudagraphs
+
+prompt = "A cat and a dog baking a cake together in a kitchen."
+negative_prompt = "Static, 2D cartoon, cartoon, 2d animation, paintings, images, worst quality, low quality, ugly, deformed, walking backwards"
+
+output = pipe(
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=512,
+    width=768,
+    num_frames=241,
+    num_inference_steps=50,
+    guidance_scale=5.0,
+).frames[0]
+
+export_to_video(output, "output.mp4", fps=24, quality=9)
+```
+
+### Diffusion Distilled model
+**⚠️ Warning!** all nocfg and diffusion distilled models should be inferred without CFG (```guidance_scale=1.0```):
+
+```python
+model_id = "ai-forever/Kandinsky-5.0-T2V-Lite-distilled16steps-5s-Diffusers"
+pipe = Kandinsky5T2VPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16)
+pipe = pipe.to("cuda")
+
+output = pipe(
+    prompt="A beautiful sunset over mountains",
+    num_inference_steps=16,  # <--- Model is distilled in 16 steps
+    guidance_scale=1.0,      # <--- no CFG
+).frames[0]
+
+export_to_video(output, "output.mp4", fps=24, quality=9)
+```
+
+
+## Citation
+```bibtex
+@misc{kandinsky2025,
+    author = {Alexey Letunovskiy and Maria Kovaleva and Ivan Kirillov and Lev Novitskiy and Denis Koposov and
+              Dmitrii Mikhailov and Anna Averchenkova and Andrey Shutkin and Julia Agafonova and Olga Kim and
+              Anastasiia Kargapoltseva and Nikita Kiselev and Vladimir Arkhipkin and Vladimir Korviakov and
+              Nikolai Gerasimenko and Denis Parkhomenko and Anna Dmitrienko and Anastasia Maltseva and
+              Kirill Chernyshev and Ilia Vasiliev and Viacheslav Vasilev and Vladimir Polovnikov and
+              Yury Kolabushin and Alexander Belykh and Mikhail Mamaev and Anastasia Aliaskina and
+              Tatiana Nikulina and Polina Gavrilova and Denis Dimitrov},
+    title = {Kandinsky 5.0: A family of diffusion models for Video & Image generation},
+    howpublished = {\url{https://github.com/ai-forever/Kandinsky-5}},
+    year = 2025
+}
+```

docs/source/en/optimization/attention_backends.md

@@ -21,6 +21,7 @@ Refer to the table below for an overview of the available attention families and
 | attention family | main feature |
 |---|---|
 | FlashAttention | minimizes memory reads/writes through tiling and recomputation |
+| AI Tensor Engine for ROCm | FlashAttention implementation optimized for AMD ROCm accelerators |
 | SageAttention | quantizes attention to int8 |
 | PyTorch native | built-in PyTorch implementation using [scaled_dot_product_attention](./fp16#scaled-dot-product-attention) |
 | xFormers | memory-efficient attention with support for various attention kernels |
@@ -139,6 +140,7 @@ Refer to the table below for a complete list of available attention backends and
 | `_native_xla` | [PyTorch native](https://docs.pytorch.org/docs/stable/generated/torch.nn.attention.SDPBackend.html#torch.nn.attention.SDPBackend) | XLA-optimized attention |
 | `flash` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | FlashAttention-2 |
 | `flash_varlen` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | Variable length FlashAttention |
+| `aiter` | [AI Tensor Engine for ROCm](https://github.com/ROCm/aiter) | FlashAttention for AMD ROCm |
 | `_flash_3` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | FlashAttention-3 |
 | `_flash_varlen_3` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | Variable length FlashAttention-3 |
 | `_flash_3_hub` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | FlashAttention-3 from kernels |

src/diffusers/loaders/lora_conversion_utils.py

@@ -1977,14 +1977,34 @@ def get_alpha_scales(down_weight, alpha_key):
                     "time_projection.1.diff_b"
                 )
 
-        if any("head.head" in k for k in state_dict):
-            converted_state_dict["proj_out.lora_A.weight"] = original_state_dict.pop(
-                f"head.head.{lora_down_key}.weight"
-            )
-            converted_state_dict["proj_out.lora_B.weight"] = original_state_dict.pop(f"head.head.{lora_up_key}.weight")
+        if any("head.head" in k for k in original_state_dict):
+            if any(f"head.head.{lora_down_key}.weight" in k for k in state_dict):
+                converted_state_dict["proj_out.lora_A.weight"] = original_state_dict.pop(
+                    f"head.head.{lora_down_key}.weight"
+                )
+            if any(f"head.head.{lora_up_key}.weight" in k for k in state_dict):
+                converted_state_dict["proj_out.lora_B.weight"] = original_state_dict.pop(
+                    f"head.head.{lora_up_key}.weight"
+                )
             if "head.head.diff_b" in original_state_dict:
                 converted_state_dict["proj_out.lora_B.bias"] = original_state_dict.pop("head.head.diff_b")
 
+            # Notes: https://huggingface.co/lightx2v/Wan2.2-Distill-Loras
+            # This is my (sayakpaul) assumption that this particular key belongs to the down matrix.
+            # Since for this particular LoRA, we don't have the corresponding up matrix, I will use
+            # an identity.
+            if any("head.head" in k and k.endswith(".diff") for k in state_dict):
+                if f"head.head.{lora_down_key}.weight" in state_dict:
+                    logger.info(
+                        f"The state dict seems to be have both `head.head.diff` and `head.head.{lora_down_key}.weight` keys, which is unexpected."
+                    )
+                converted_state_dict["proj_out.lora_A.weight"] = original_state_dict.pop("head.head.diff")
+                down_matrix_head = converted_state_dict["proj_out.lora_A.weight"]
+                up_matrix_shape = (down_matrix_head.shape[0], converted_state_dict["proj_out.lora_B.bias"].shape[0])
+                converted_state_dict["proj_out.lora_B.weight"] = torch.eye(
+                    *up_matrix_shape, dtype=down_matrix_head.dtype, device=down_matrix_head.device
+                ).T
+
         for text_time in ["text_embedding", "time_embedding"]:
             if any(text_time in k for k in original_state_dict):
                 for b_n in [0, 2]:

src/diffusers/models/attention_dispatch.py

@@ -28,6 +28,8 @@
 
 from ..utils import (
     get_logger,
+    is_aiter_available,
+    is_aiter_version,
     is_flash_attn_3_available,
     is_flash_attn_available,
     is_flash_attn_version,
@@ -48,13 +50,15 @@
     from ._modeling_parallel import ParallelConfig
 
 _REQUIRED_FLASH_VERSION = "2.6.3"
+_REQUIRED_AITER_VERSION = "0.1.5"
 _REQUIRED_SAGE_VERSION = "2.1.1"
 _REQUIRED_FLEX_VERSION = "2.5.0"
 _REQUIRED_XLA_VERSION = "2.2"
 _REQUIRED_XFORMERS_VERSION = "0.0.29"
 
 _CAN_USE_FLASH_ATTN = is_flash_attn_available() and is_flash_attn_version(">=", _REQUIRED_FLASH_VERSION)
 _CAN_USE_FLASH_ATTN_3 = is_flash_attn_3_available()
+_CAN_USE_AITER_ATTN = is_aiter_available() and is_aiter_version(">=", _REQUIRED_AITER_VERSION)
 _CAN_USE_SAGE_ATTN = is_sageattention_available() and is_sageattention_version(">=", _REQUIRED_SAGE_VERSION)
 _CAN_USE_FLEX_ATTN = is_torch_version(">=", _REQUIRED_FLEX_VERSION)
 _CAN_USE_NPU_ATTN = is_torch_npu_available()
@@ -79,6 +83,11 @@
     flash_attn_3_func = None
     flash_attn_3_varlen_func = None
 
+if _CAN_USE_AITER_ATTN:
+    from aiter import flash_attn_func as aiter_flash_attn_func
+else:
+    aiter_flash_attn_func = None
+
 if _CAN_USE_SAGE_ATTN:
     from sageattention import (
         sageattn,
@@ -167,6 +176,9 @@ class AttentionBackendName(str, Enum):
     _FLASH_3_HUB = "_flash_3_hub"
     # _FLASH_VARLEN_3_HUB = "_flash_varlen_3_hub"  # not supported yet.
 
+    # `aiter`
+    AITER = "aiter"
+
     # PyTorch native
     FLEX = "flex"
     NATIVE = "native"
@@ -418,6 +430,12 @@ def _check_attention_backend_requirements(backend: AttentionBackendName) -> None
                 f"Backend '{backend.value}' is not usable because the `kernels` package isn't available. Please install it with `pip install kernels`."
             )
 
+    elif backend == AttentionBackendName.AITER:
+        if not _CAN_USE_AITER_ATTN:
+            raise RuntimeError(
+                f"Aiter Attention backend '{backend.value}' is not usable because of missing package or the version is too old. Please install `aiter>={_REQUIRED_AITER_VERSION}`."
+            )
+
     elif backend in [
         AttentionBackendName.SAGE,
         AttentionBackendName.SAGE_VARLEN,
@@ -1425,6 +1443,47 @@ def _flash_varlen_attention_3(
     return (out, lse) if return_lse else out
 
 
+@_AttentionBackendRegistry.register(
+    AttentionBackendName.AITER,
+    constraints=[_check_device_cuda, _check_qkv_dtype_bf16_or_fp16, _check_shape],
+)
+def _aiter_flash_attention(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    dropout_p: float = 0.0,
+    is_causal: bool = False,
+    scale: Optional[float] = None,
+    return_lse: bool = False,
+    _parallel_config: Optional["ParallelConfig"] = None,
+) -> torch.Tensor:
+    if not return_lse and torch.is_grad_enabled():
+        # aiter requires return_lse=True by assertion when gradients are enabled.
+        out, lse, *_ = aiter_flash_attn_func(
+            q=query,
+            k=key,
+            v=value,
+            dropout_p=dropout_p,
+            softmax_scale=scale,
+            causal=is_causal,
+            return_lse=True,
+        )
+    else:
+        out = aiter_flash_attn_func(
+            q=query,
+            k=key,
+            v=value,
+            dropout_p=dropout_p,
+            softmax_scale=scale,
+            causal=is_causal,
+            return_lse=return_lse,
+        )
+        if return_lse:
+            out, lse, *_ = out
+
+    return (out, lse) if return_lse else out
+
+
 @_AttentionBackendRegistry.register(
     AttentionBackendName.FLEX,
     constraints=[_check_attn_mask_or_causal, _check_device, _check_shape],