[Diffusion][NPU]Add attention backends for diffusion models for Ascend NPU (#23482)

Co-authored-by: Napkin-AI <arseniy.mironov.dev@gmail.com>

docs/diffusion/compatibility_matrix.md

@@ -16,29 +16,29 @@ default parameters when initializing and generating videos.
 
 ### Video Generation Models
 
-| Model Name                   | Hugging Face Model ID                             | Resolutions          | TeaCache | Sliding Tile Attn | Sage Attn | Video Sparse Attention (VSA) | Sparse Linear Attention (SLA) | Sage Sparse Linear Attention (SageSLA) | Sparse Video Gen 2 (SVG2) |
+| Model Name                   | Hugging Face Model ID                             | Resolutions          | TeaCache | Sliding Tile Attn | Sage Attn | Video Sparse Attention (VSA) | Sparse Linear Attention (SLA) | Sage Sparse Linear Attention (SageSLA) | Sparse Video Gen 2 (SVG2) | Laser Attn | Block Sparse Attn | Rain Fusion Attn |
-|:-----------------------------|:--------------------------------------------------|:---------------------|:--------:|:-----------------:|:---------:|:----------------------------:|:-----------------------------:|:--------------------------------------:|:-------------------------:|
+|:-----------------------------|:--------------------------------------------------|:---------------------|:--------:|:-----------------:|:---------:|:----------------------------:|:-----------------------------:|:--------------------------------------:|:-------------------------:|:-------------------------:|:-------------------------:|:-------------------------:|
-| FastWan2.1 T2V 1.3B          | `FastVideo/FastWan2.1-T2V-1.3B-Diffusers`         | 480p                 |    ⭕     |         ⭕         |     ⭕     |              ✅               |               ❌               |                   ❌                    |             ❌             |
+| FastWan2.1 T2V 1.3B          | `FastVideo/FastWan2.1-T2V-1.3B-Diffusers`         | 480p                 |    ⭕     |         ⭕         |     ⭕     |              ✅               |               ❌               |                   ❌                    |             ❌             |               ❌               |                   ❌                    |             ❌             |
-| FastWan2.2 TI2V 5B Full Attn | `FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers` | 720p                 |    ⭕     |         ⭕         |     ⭕     |              ✅               |               ❌               |                   ❌                    |             ❌             |
+| FastWan2.2 TI2V 5B Full Attn | `FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers` | 720p                 |    ⭕     |         ⭕         |     ⭕     |              ✅               |               ❌               |                   ❌                    |             ❌             |               ❌               |                   ❌                    |             ❌             |
-| Wan2.2 TI2V 5B               | `Wan-AI/Wan2.2-TI2V-5B-Diffusers`                 | 720p                 |    ⭕     |         ⭕         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ❌             |
+| Wan2.2 TI2V 5B               | `Wan-AI/Wan2.2-TI2V-5B-Diffusers`                 | 720p                 |    ⭕     |         ⭕         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ❌             |               ✅               |                   ✅                    |             ✅             |
-| Wan2.2 T2V A14B              | `Wan-AI/Wan2.2-T2V-A14B-Diffusers`                | 480p<br>720p         |    ❌     |         ❌         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ❌             |
+| Wan2.2 T2V A14B              | `Wan-AI/Wan2.2-T2V-A14B-Diffusers`                | 480p<br>720p         |    ❌     |         ❌         |     ✅     |              ⭕               |               ✅               |                   ❌                    |             ✅             |
-| Wan2.2 I2V A14B              | `Wan-AI/Wan2.2-I2V-A14B-Diffusers`                | 480p<br>720p         |    ❌     |         ❌         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ❌             |
+| Wan2.2 I2V A14B              | `Wan-AI/Wan2.2-I2V-A14B-Diffusers`                | 480p<br>720p         |    ❌     |         ❌         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ❌             |               ✅               |                   ✅                    |             ✅             |
-| HunyuanVideo                 | `hunyuanvideo-community/HunyuanVideo`             | 720×1280<br>544×960  |    ❌     |         ✅         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ✅             |
+| HunyuanVideo                 | `hunyuanvideo-community/HunyuanVideo`             | 720×1280<br>544×960  |    ❌     |         ✅         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ✅             |               ❌               |                   ❌                    |             ❌             |
-| FastHunyuan                  | `FastVideo/FastHunyuan-diffusers`                 | 720×1280<br>544×960  |    ❌     |         ✅         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ✅             |
+| FastHunyuan                  | `FastVideo/FastHunyuan-diffusers`                 | 720×1280<br>544×960  |    ❌     |         ✅         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ✅             |               ❌               |                   ❌                    |             ❌             |
-| Wan2.1 T2V 1.3B              | `Wan-AI/Wan2.1-T2V-1.3B-Diffusers`                | 480p                 |    ✅     |         ✅         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ✅             |
+| Wan2.1 T2V 1.3B              | `Wan-AI/Wan2.1-T2V-1.3B-Diffusers`                | 480p                 |    ✅     |         ✅         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ✅             |               ✅               |                   ✅                    |             ✅             |
-| Wan2.1 T2V 14B               | `Wan-AI/Wan2.1-T2V-14B-Diffusers`                 | 480p, 720p           |    ✅     |         ✅         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ✅             |
+| Wan2.1 T2V 14B               | `Wan-AI/Wan2.1-T2V-14B-Diffusers`                 | 480p, 720p           |    ✅     |         ✅         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ✅             |               ✅               |                   ✅                    |             ✅             |
-| Wan2.1 I2V 480P              | `Wan-AI/Wan2.1-I2V-14B-480P-Diffusers`            | 480p                 |    ✅     |         ✅         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ✅             |
+| Wan2.1 I2V 480P              | `Wan-AI/Wan2.1-I2V-14B-480P-Diffusers`            | 480p                 |    ✅     |         ✅         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ✅             |               ✅               |                   ✅                    |             ✅             |
-| Wan2.1 I2V 720P              | `Wan-AI/Wan2.1-I2V-14B-720P-Diffusers`            | 720p                 |    ✅     |         ✅         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ✅             |
+| Wan2.1 I2V 720P              | `Wan-AI/Wan2.1-I2V-14B-720P-Diffusers`            | 720p                 |    ✅     |         ✅         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ✅             |               ✅               |                   ✅                    |             ✅             |
 | TurboWan2.1 T2V 1.3B         | `IPostYellow/TurboWan2.1-T2V-1.3B-Diffusers`      | 480p                 |    ✅     |         ❌         |     ❌     |              ❌               |               ✅               |                   ✅                    |             ⭕             |
-| TurboWan2.1 T2V 14B          | `IPostYellow/TurboWan2.1-T2V-14B-Diffusers`       | 480p                 |    ✅     |         ❌         |     ❌     |              ❌               |               ✅               |                   ✅                    |             ⭕             |
+| TurboWan2.1 T2V 14B          | `IPostYellow/TurboWan2.1-T2V-14B-Diffusers`       | 480p                 |    ✅     |         ❌         |     ❌     |              ❌               |               ✅               |                   ✅                    |             ⭕             |               ❌               |                   ❌                    |             ❌             |
 | TurboWan2.1 T2V 14B 720P     | `IPostYellow/TurboWan2.1-T2V-14B-720P-Diffusers`  | 720p                 |    ✅     |         ❌         |     ❌     |              ❌               |               ✅               |                   ✅                    |             ⭕             |
 | TurboWan2.2 I2V A14B         | `IPostYellow/TurboWan2.2-I2V-A14B-Diffusers`      | 720p                 |    ✅     |         ❌         |     ❌     |              ❌               |               ✅               |                   ✅                    |             ⭕             |
-| Wan2.1 Fun 1.3B InP          | `weizhou03/Wan2.1-Fun-1.3B-InP-Diffusers`          | 480p                 |    ✅     |         ✅         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ✅             |
+| Wan2.1 Fun 1.3B InP          | `weizhou03/Wan2.1-Fun-1.3B-InP-Diffusers`          | 480p                 |    ✅     |         ✅         |     ✅     |              ⭕               |               ❌               |                   ❌                    |             ✅             |               ❌               |                   ❌                    |             ❌             |
-| Helios Base                  | `BestWishYsh/Helios-Base`                          | 720p                 |    ❌     |         ❌         |     ❌     |              ❌               |               ❌               |                   ❌                    |             ❌             |
+| Helios Base                  | `BestWishYsh/Helios-Base`                          | 720p                 |    ❌     |         ❌         |     ❌     |              ❌               |               ❌               |                   ❌                    |             ❌             |               ❌               |                   ❌                    |             ❌             |
 | Helios Mid                   | `BestWishYsh/Helios-Mid`                           | 720p                 |    ❌     |         ❌         |     ❌     |              ❌               |               ❌               |                   ❌                    |             ❌             |
-| Helios Distilled             | `BestWishYsh/Helios-Distilled`                     | 720p                 |    ❌     |         ❌         |     ❌     |              ❌               |               ❌               |                   ❌                    |             ❌             |
+| Helios Distilled             | `BestWishYsh/Helios-Distilled`                     | 720p                 |    ❌     |         ❌         |     ❌     |              ❌               |               ❌               |                   ❌                    |             ❌             |               ❌               |                   ❌                    |             ❌             |
-| LTX-2 (one/two-stage/TI2V)   | `Lightricks/LTX-2`                                | 768×512<br>1536×1024 |    ❌     |         ❌         |     ❌     |              ❌               |               ❌               |                   ❌                    |             ❌             |
+| LTX-2 (one/two-stage/TI2V)   | `Lightricks/LTX-2`                                | 768×512<br>1536×1024 |    ❌     |         ❌         |     ❌     |              ❌               |               ❌               |                   ❌                    |             ❌             |               ❌               |                   ❌                    |             ❌             |
-| LTX-2.3 (one/two-stage/TI2V/HQ) | `Lightricks/LTX-2.3`                           | 768×512<br>1536×1024<br>1920×1088 (HQ default) |    ❌     |         ❌         |     ❌     |              ❌               |               ❌               |                   ❌                    |             ❌             |
+| LTX-2.3 (one/two-stage/TI2V/HQ) | `Lightricks/LTX-2.3`                           | 768×512<br>1536×1024<br>1920×1088 (HQ default) |    ❌     |         ❌         |     ❌     |              ❌               |               ❌               |                   ❌                    |               ❌               |                   ❌                    |             ❌             |             ❌             |
 
 **Note**:
 

docs/diffusion/performance/attention_backends.md

@@ -36,6 +36,9 @@ For SGLang-native pipelines, the CLI accepts the lowercase names of `AttentionBa
 | `sla_attn` | `SLA_ATTN` | Sparse Linear Attention. Requires `SpargeAttn`. Install with `pip install git+https://github.com/thu-ml/SpargeAttn.git --no-build-isolation`. |
 | `sage_sla_attn` | `SAGE_SLA_ATTN` | SageAttention + Sparse Linear Attention. Requires `SpargeAttn` (same install as SLA). |
 | `sparse_video_gen_2_attn` | `SPARSE_VIDEO_GEN_2_ATTN` | Requires `svg`. See installation instructions at https://github.com/svg-project/Sparse-VideoGen. |
+| `laser_attn` | `LASER_ATTN` | Requires `attentions` which can be installed with `sgl_kernel_npu`; available only for NPU. |
+| `block_sparse_attn` | `BLOCK_SPARSE_ATTN` | Requires `attentions` which can be installed with `sgl_kernel_npu`; available only for NPU. |
+| `rain_fusion_attn` | `RAIN_FUSION_ATTN` | Requires `attentions` which can be installed with `sgl_kernel_npu`; available only for NPU. |
 
 ## Selection priority
 
@@ -88,6 +91,12 @@ Some backends require additional configuration. You can pass these parameters vi
 | `spatial_layer` | `int` | Number of spatial layers. | `1` |
 | `st_layer` | `int` | Number of spatiotemporal layers. | `1` |
 
+**Block Sparse attention ( `block_sparse_attn` ) and  Rain Fusion attention ( `rain_fusion_attn` )**
+| Parameter | Type | Description | Default |
+| :--- | :--- | :--- | :--- |
+| `skip_first_steps` | `int` | Number of steps to use laser attention before switching to sparse attention. | `10` |
+| `sparsity` | `float` | The sparsity coefficient must be in the range (0, 1). | `0.2` |
+
 ## Platform support matrix
 
 | Backend | CUDA | ROCm | XPU | MUSA | MPS | NPU | Notes |
@@ -104,6 +113,9 @@ Some backends require additional configuration. You can pass these parameters vi
 | `aiter` | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | Requires `aiter`. |
 | `aiter_sage` | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | Requires `aiter`. |
 | `sparse_video_gen_2_attn` | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | CUDA-only. Requires `svg`. |
+| `laser_attn` | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | NPU-only. Requires `attentions` from `sgl_kernel_npu`. Uses SDPA if seqlen is less than 2048. |
+| `block_sparse_attn` | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | NPU-only. Requires `attentions` from `sgl_kernel_npu`.  Configure via `--attention-backend-config`. |
+| `rain_fusion_attn` | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | NPU-only. Requires `attentions` from `sgl_kernel_npu`. Configure via `--attention-backend-config`. |
 
 ## Usage
 

docs_new/docs/sglang-diffusion/attention_backends.mdx

@@ -98,6 +98,21 @@ For SGLang-native pipelines, the CLI accepts the lowercase names of `AttentionBa
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)", whiteSpace: "nowrap"}}>`SPARSE_VIDEO_GEN_2_ATTN`</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>Requires <code>svg</code>. See installation instructions at https://github.com/svg-project/Sparse-VideoGen.</td>
     </tr>
+<tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><code>laser_attn</code></td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)", whiteSpace: "nowrap"}}><code>LASER_ATTN</code></td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>Requires <code>attentions</code> which can be installed with <code>sgl_kernel_npu</code>; available only for NPU.</td>
+    </tr>
+<tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><code>block_sparse_attn</code></td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)", whiteSpace: "nowrap"}}><code>BLOCK_SPARSE_ATTN</code></td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>Requires <code>attentions</code> which can be installed with <code>sgl_kernel_npu</code>; available only for NPU.</td>
+    </tr>
+<tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><code>rain_fusion_attn</code></td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)", whiteSpace: "nowrap"}}><code>RAIN_FUSION_ATTN</code></td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>Requires <code>attentions</code> which can be installed with <code>sgl_kernel_npu</code>; available only for NPU.</td>
+    </tr>
 </tbody>
 </table>
 
@@ -290,6 +305,39 @@ Some backends require additional configuration. You can pass these parameters vi
   </tbody>
 </table>
 
+**Block Sparse Attention (`block_sparse_attn`)**
+
+<table style={{width: "100%", borderCollapse: "collapse", tableLayout: "fixed"}}>
+  <colgroup>
+    <col style={{width: "20%"}} />
+    <col style={{width: "16%"}} />
+    <col style={{width: "46%"}} />
+    <col style={{width: "18%"}} />
+  </colgroup>
+  <thead>
+    <tr style={{borderBottom: "2px solid #d55816"}}>
+      <th style={{textAlign: "left", padding: "10px 12px", fontWeight: 700, whiteSpace: "nowrap", backgroundColor: "rgba(255,255,255,0.02)"}}>Parameter</th>
+      <th style={{textAlign: "left", padding: "10px 12px", fontWeight: 700, whiteSpace: "nowrap", backgroundColor: "rgba(255,255,255,0.05)"}}>Type</th>
+      <th style={{textAlign: "left", padding: "10px 12px", fontWeight: 700, whiteSpace: "nowrap", backgroundColor: "rgba(255,255,255,0.02)"}}>Description</th>
+      <th style={{textAlign: "left", padding: "10px 12px", fontWeight: 700, whiteSpace: "nowrap", backgroundColor: "rgba(255,255,255,0.05)"}}>Default</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>`skip_first_steps`</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>`int`</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>Number of steps to use laser attention before switching to sparse attention.</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>`10`</td>
+    </tr>
+    <tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>`sparsity`</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>`float`</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}> The sparsity coefficient must be in the range (0, 1).</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>`0.2`</td>
+    </tr>
+  </tbody>
+</table>
+
 ## Platform support matrix
 
 <table style={{width: "100%", borderCollapse: "collapse", tableLayout: "fixed"}}>
@@ -434,6 +482,36 @@ Some backends require additional configuration. You can pass these parameters vi
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>CUDA-only. Requires <code>svg</code>.</td>
     </tr>
+<tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><code>laser_attn</code></td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>NPU-only. Requires <code>attentions</code> from <code>sgl_kernel_npu</code>. Uses SDPA if seqlen less than 2048.</td>
+    </tr>
+<tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><code>block_sparse_attn</code></td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>NPU-only. Requires <code>attentions</code> from <code>sgl_kernel_npu</code>. Configuration via <code>--attention-backend-config</code>.</td>
+    </tr>
+<tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><code>rain_fusion_attn</code></td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>NPU-only. Requires <code>attentions</code> from <code>sgl_kernel_npu</code> Configuration via <code>--attention-backend-config</code>.</td>
+    </tr>
 </tbody>
 </table>
 

docs_new/docs/sglang-diffusion/compatibility_matrix.mdx

@@ -27,6 +27,9 @@ Optimization columns are abbreviated to keep the matrix readable:
 - `SLA` = Sparse Linear Attention
 - `SageSLA` = Sage Sparse Linear Attention
 - `SVG2` = Sparse Video Gen 2
+- `LA` = Laser Attention
+- `BSA` = Block Sparse Attention
+- `RF` = Rain Fusion Attention
 
 <table style={{width: "100%", borderCollapse: "collapse", tableLayout: "fixed"}}>
   <colgroup>
@@ -40,6 +43,9 @@ Optimization columns are abbreviated to keep the matrix readable:
     <col style={{width: "5%"}} />
     <col style={{width: "5%"}} />
     <col style={{width: "5%"}} />
+    <col style={{width: "5%"}} />
+    <col style={{width: "5%"}} />
+    <col style={{width: "5%"}} />
   </colgroup>
   <thead>
     <tr style={{borderBottom: "2px solid #d55816"}}>
@@ -53,6 +59,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <th style={{textAlign: "center", padding: "10px 6px", fontWeight: 700, whiteSpace: "normal", backgroundColor: "rgba(255,255,255,0.05)"}}><abbr title="Sparse Linear Attention">SLA</abbr></th>
       <th style={{textAlign: "center", padding: "10px 6px", fontWeight: 700, whiteSpace: "normal", backgroundColor: "rgba(255,255,255,0.02)"}}><abbr title="Sage Sparse Linear Attention">SageSLA</abbr></th>
       <th style={{textAlign: "center", padding: "10px 6px", fontWeight: 700, whiteSpace: "normal", backgroundColor: "rgba(255,255,255,0.05)"}}><abbr title="Sparse Video Gen 2">SVG2</abbr></th>
+      <th style={{textAlign: "center", padding: "10px 6px", fontWeight: 700, whiteSpace: "normal", backgroundColor: "rgba(255,255,255,0.02)"}}><abbr title="Laser Attention">LA</abbr></th>
+      <th style={{textAlign: "center", padding: "10px 6px", fontWeight: 700, whiteSpace: "normal", backgroundColor: "rgba(255,255,255,0.02)"}}><abbr title="Block Sparse Attention">BSA</abbr></th>
+      <th style={{textAlign: "center", padding: "10px 6px", fontWeight: 700, whiteSpace: "normal", backgroundColor: "rgba(255,255,255,0.02)"}}><abbr title="Rain Fusion Attention">RF</abbr></th>
     </tr>
   </thead>
   <tbody>
@@ -67,6 +76,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
     </tr>
     <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>FastWan2.2 TI2V 5B Full Attn</td>
@@ -79,6 +91,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
     </tr>
     <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>Wan2.2 TI2V 5B</td>
@@ -91,6 +106,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>✅</td>
     </tr>
     <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>Wan2.2 T2V A14B</td>
@@ -103,6 +121,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>✅</td>
     </tr>
     <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>Wan2.2 I2V A14B</td>
@@ -115,6 +136,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>✅</td>
     </tr>
     <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>HunyuanVideo</td>
@@ -127,6 +151,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
     </tr>
     <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>FastHunyuan</td>
@@ -139,6 +166,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
     </tr>
     <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>Wan2.1 T2V 1.3B</td>
@@ -151,6 +181,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>✅</td>
     </tr>
     <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>Wan2.1 T2V 14B</td>
@@ -163,6 +196,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>✅</td>
     </tr>
     <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>Wan2.1 I2V 480P</td>
@@ -175,6 +211,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>✅</td>
     </tr>
     <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>Wan2.1 I2V 720P</td>
@@ -187,6 +226,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>✅</td>
     </tr>
     <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>TurboWan2.1 T2V 1.3B</td>
@@ -199,6 +241,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>✅</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>⭕</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
     </tr>
     <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>TurboWan2.1 T2V 14B</td>
@@ -211,6 +256,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>✅</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>⭕</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
     </tr>
     <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>TurboWan2.1 T2V 14B 720P</td>
@@ -223,6 +271,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>✅</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>⭕</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
     </tr>
     <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>TurboWan2.2 I2V A14B</td>
@@ -235,6 +286,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>✅</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>⭕</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
     </tr>
 <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>Wan2.1 Fun 1.3B InP</td>
@@ -247,6 +301,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>✅</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
     </tr>
 <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>Helios Base</td>
@@ -259,6 +316,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
     </tr>
 <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>Helios Mid</td>
@@ -271,6 +331,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
     </tr>
 <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>Helios Distilled</td>
@@ -283,6 +346,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
     </tr>
 <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>LTX-2 (one/two-stage/TI2V)</td>
@@ -295,6 +361,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
     </tr>
 <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>LTX-2.3 (one/two-stage/TI2V/HQ)</td>
@@ -307,6 +376,9 @@ Optimization columns are abbreviated to keep the matrix readable:
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>❌</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>❌</td>
     </tr>
 </tbody>
 </table>

python/sglang/multimodal_gen/configs/models/adapter/base.py

@@ -27,6 +27,9 @@ class AdapterArchConfig(ArchConfig):
             AttentionBackendEnum.VIDEO_SPARSE_ATTN,
             AttentionBackendEnum.VMOBA_ATTN,
             AttentionBackendEnum.SAGE_ATTN_3,
+            AttentionBackendEnum.LASER_ATTN,
+            AttentionBackendEnum.BLOCK_SPARSE_ATTN,
+            AttentionBackendEnum.RAIN_FUSION_ATTN,
         }
     )
 

python/sglang/multimodal_gen/configs/models/dits/base.py

@@ -35,6 +35,9 @@ class DiTArchConfig(ArchConfig):
             AttentionBackendEnum.SPARSE_VIDEO_GEN_2_ATTN,
             AttentionBackendEnum.VMOBA_ATTN,
             AttentionBackendEnum.SAGE_ATTN_3,
+            AttentionBackendEnum.LASER_ATTN,
+            AttentionBackendEnum.BLOCK_SPARSE_ATTN,
+            AttentionBackendEnum.RAIN_FUSION_ATTN,
         }
     )
 

python/sglang/multimodal_gen/runtime/layers/attention/backends/block_sparse_attn.py

@@ -0,0 +1,279 @@
+from dataclasses import dataclass
+from typing import Any
+
+import attentions  # noqa: F401
+import torch
+
+from sglang.multimodal_gen.runtime.layers.attention.backends.attention_backend import (
+    AttentionBackend,
+    AttentionImpl,
+    AttentionMetadata,
+    AttentionMetadataBuilder,
+)
+from sglang.multimodal_gen.runtime.layers.attention.backends.laser_attn import (
+    LaserAttentionBackend,
+)
+from sglang.multimodal_gen.runtime.platforms import AttentionBackendEnum
+from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
+
+logger = init_logger(__name__)
+BSA_BLOCK_SIZE = 128
+
+
+class BlockSparseAttentionBackend(AttentionBackend):
+
+    accept_output_buffer: bool = True
+
+    @staticmethod
+    def get_supported_head_sizes() -> list[int]:
+        return [32, 64, 96, 128]
+
+    @staticmethod
+    def get_enum() -> AttentionBackendEnum:
+        return AttentionBackendEnum.BLOCK_SPARSE_ATTN
+
+    @staticmethod
+    def get_impl_cls() -> type["BlockSparseAttentionImpl"]:
+        return BlockSparseAttentionImpl
+
+    @staticmethod
+    def get_metadata_cls() -> type["BlockSparseAttentionMetadata"]:
+        return BlockSparseAttentionMetadata
+
+    @staticmethod
+    def get_builder_cls() -> type["BlockSparseAttentionMetadataBuilder"]:
+        return BlockSparseAttentionMetadataBuilder
+
+
+@dataclass
+class BlockSparseAttentionMetadata(AttentionMetadata):
+    current_timestep: int
+    skip_first_steps: int
+    sparsity: float
+    block_frame_stride: int
+
+
+class BlockSparseAttentionMetadataBuilder(AttentionMetadataBuilder):
+    def __init__(self) -> None:
+        pass
+
+    def prepare(self) -> None:
+        pass
+
+    def build(
+        self,
+        current_timestep: int,
+        skip_first_steps: int,
+        sparsity: float,
+        raw_latent_shape: list[int],
+        patch_size: tuple[int, int, int],
+        **kwargs: dict[str, Any],
+    ) -> BlockSparseAttentionMetadata:
+        """
+        Builds BlockSparseAttention metadata.
+
+        Args:
+            current_timestep: The current diffusion timestep.
+            skip_first_steps: Number of initial timesteps to skip before applying
+                sparsity. Must be non‑negative.
+            sparsity: Fraction of tokens to drop (block‑wise) in the block sparse
+                attention mechanism. Must be in the range [0.0, 1.0).
+            raw_latent_shape: Shape of the latent tensor before patching.
+            patch_size: Patch size as (T, height, width). Only the height
+                and width components are used to divide the latent dimensions.
+            **kwargs: Additional keyword arguments (ignored, but accepted for
+                compatibility with base class or calling conventions).
+
+        Returns:
+            BlockSparseAttentionMetadata
+        Note:
+            The `block_frame_stride` is needed to set the first blocks to be non‑sparse.
+        """
+        if not (skip_first_steps >= 0 and 0.0 <= sparsity < 1.0):
+            raise ValueError(
+                (
+                    "Invalid attention metadata values."
+                    f"Sparsity should be in [0, 1), skip_first_steps should be non-negative."
+                    f"Got sparsity={sparsity}, skip_first_steps={skip_first_steps}"
+                )
+            )
+
+        if sparsity == 0.0:
+            logger.warning(
+                (
+                    "Sparsity is set to 0.0, which means no tokens will be dropped."
+                    "For better performance use Laser Attention or increase sparsity."
+                )
+            )
+
+        if len(raw_latent_shape) >= 5:
+            latent_height, latent_width = raw_latent_shape[3:5]
+        else:
+            latent_height, latent_width = raw_latent_shape[-2:]
+
+        latent_height //= patch_size[1]
+        latent_width //= patch_size[2]
+
+        frame_stride = latent_height * latent_width
+        block_frame_stride = (frame_stride + BSA_BLOCK_SIZE - 1) // BSA_BLOCK_SIZE
+
+        return BlockSparseAttentionMetadata(
+            current_timestep=current_timestep,
+            skip_first_steps=skip_first_steps,
+            sparsity=sparsity,
+            block_frame_stride=block_frame_stride,
+        )
+
+
+class BlockSparseAttentionImpl(AttentionImpl):
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        causal: bool,
+        softmax_scale: float,
+        num_kv_heads: int | None = None,
+        prefix: str = "",
+        **extra_impl_args,
+    ) -> None:
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.num_heads = num_heads
+        self.num_kv_heads = num_kv_heads or num_heads
+        self.block_size = BSA_BLOCK_SIZE
+        self.stride = 8
+        self.default_tokens = 214748647
+
+        self.laser_attn_impl = LaserAttentionBackend.get_impl_cls()(
+            num_heads,
+            head_size,
+            causal,
+            softmax_scale,
+            num_kv_heads,
+            prefix,
+            **extra_impl_args,
+        )
+
+    def _get_estimate_mask(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        sparsity: float,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        return torch.ops.attentions.sparse_block_estimate(
+            query=query,
+            key=key,
+            actual_seq_lengths=None,
+            actual_seq_lengths_kv=None,
+            input_layout="BNSD",
+            stride=self.stride,
+            sparse_size=self.block_size,
+            num_heads=query.shape[1],
+            num_key_value_heads=key.shape[1],
+            scale_value=self.softmax_scale / self.stride,
+            threshold=1.0,
+            causal=self.causal,
+            keep_sink=True,
+            keep_recent=True,
+            row_sparse=1.0 - sparsity,
+        )
+
+    def _block_sparse_attention(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        smask: torch.Tensor,
+        sct: torch.Tensor,
+    ) -> torch.Tensor:
+        return torch.ops.attentions.block_sparse_attention(
+            query=query,
+            key=key,
+            value=value,
+            sparse_mask=smask,
+            sparse_count_table=sct,
+            input_layout="BNSD",
+            sparse_size=self.block_size,
+            num_heads=query.shape[1],
+            num_key_value_heads=key.shape[1],
+            scale_value=self.softmax_scale,
+            causal=self.causal,
+            inner_precise=1,
+            pre_tokens=self.default_tokens,
+            next_tokens=self.default_tokens,
+            actual_seq_lengths=None,
+            actual_seq_lengths_kv=None,
+        )
+
+    def _get_smask(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        block_frame_stride: int,
+        sparsity: float,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        smask, sct = self._get_estimate_mask(
+            query,
+            key,
+            sparsity,
+        )
+
+        seq_len = smask.shape[2]
+
+        # Set the first blocks to be non-sparse to ensure the quality of the first few steps
+        smask[:, :, :block_frame_stride, :seq_len] = 1
+        smask[:, :, :seq_len, :block_frame_stride] = 1
+        smask = smask.to(torch.int8)
+        sct = smask.sum(dim=-1, dtype=torch.int32)
+        return smask, sct
+
+    def _adaptive_block_sparse_attention(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        block_frame_stride: int,
+        sparsity: float,
+    ) -> torch.Tensor:
+        # TODO Currently implementation for BSND input layout has quality issues
+        # When the implementation is improved, transposes can be removed
+        q = query.permute(0, 2, 1, 3).contiguous()
+        k = key.permute(0, 2, 1, 3).contiguous()
+        v = value.permute(0, 2, 1, 3).contiguous()
+
+        smask, sct = self._get_smask(
+            q,
+            k,
+            block_frame_stride,
+            sparsity,
+        )
+        output = self._block_sparse_attention(q, k, v, smask, sct)
+        output = output.permute(0, 2, 1, 3).contiguous()
+
+        return output
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        if attn_metadata.current_timestep < attn_metadata.skip_first_steps:
+            output = self.laser_attn_impl.forward(
+                query,
+                key,
+                value,
+                attn_metadata,
+            )
+        else:
+            output = self._adaptive_block_sparse_attention(
+                query,
+                key,
+                value,
+                attn_metadata.block_frame_stride,
+                attn_metadata.sparsity,
+            )
+
+        return output

python/sglang/multimodal_gen/runtime/layers/attention/backends/laser_attn.py

@@ -0,0 +1,191 @@
+import torch
+
+from sglang.multimodal_gen.runtime.layers.attention.backends.attention_backend import (
+    AttentionBackend,
+    AttentionImpl,
+    AttentionMetadata,
+)
+from sglang.multimodal_gen.runtime.layers.attention.backends.sdpa import SDPABackend
+from sglang.multimodal_gen.runtime.platforms import AttentionBackendEnum
+from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
+
+# Import to use torch.ops.attentions, install package with sgl_kernel_npu
+try:
+    import attentions  # noqa: F401
+except ImportError as e:
+    raise ImportError(
+        (
+            "The required 'attentions' package is not installed."
+            "The package can be installed with sgl_kernel_npu"
+        )
+    ) from e
+
+logger = init_logger(__name__)
+
+
+class LaserAttentionBackend(AttentionBackend):
+
+    accept_output_buffer: bool = True
+
+    @staticmethod
+    def get_supported_head_sizes() -> list[int]:
+        return [32, 64, 96, 128]
+
+    @staticmethod
+    def get_enum() -> AttentionBackendEnum:
+        return AttentionBackendEnum.LASER_ATTN
+
+    @staticmethod
+    def get_impl_cls() -> type["LaserAttentionImpl"]:
+        return LaserAttentionImpl
+
+
+class LaserAttentionImpl(AttentionImpl):
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        causal: bool,
+        softmax_scale: float,
+        num_kv_heads: int | None = None,
+        prefix: str = "",
+        **extra_impl_args,
+    ) -> None:
+        self.softmax_scale = softmax_scale
+
+        # After preprocess input layout should be BNSD.
+        self.seqlen_base = 256
+        self.seqlen_index = 2
+        self.dim_index = 3
+        self.dim_base = 128
+        self.max_token = 2**31 - 1
+        self.seq_len_pad_base = 256
+
+        # the laser attention operator has issues with small seq_len
+        self.min_seqlen = 2048
+        self.sdpa_impl = SDPABackend.get_impl_cls()(
+            num_heads,
+            head_size,
+            causal,
+            softmax_scale,
+            num_kv_heads,
+            prefix,
+            **extra_impl_args,
+        )
+
+    def _pad(self, input_tensor: torch.Tensor) -> torch.Tensor:
+        """
+        Pad the input tensor along the sequence length and head dimension.
+        to multiples of base values. self.seqlen_index and self.dim_index should be positive integers.
+        """
+
+        seq_len = input_tensor.size(self.seqlen_index)
+        head_dim = input_tensor.size(self.dim_index)
+
+        pad_seq = 0
+        if seq_len % self.seqlen_base != 0:
+            pad_seq = ((seq_len // self.seqlen_base) + 1) * self.seqlen_base - seq_len
+
+        pad_dim = 0
+        if head_dim % self.dim_base != 0:
+            pad_dim = ((head_dim // self.dim_base) + 1) * self.dim_base - head_dim
+
+        if pad_seq == 0 and pad_dim == 0:
+            return input_tensor
+
+        pad_list = [0] * (2 * input_tensor.ndim)
+
+        pad_list[len(pad_list) - 2 * self.seqlen_index - 1] = pad_seq
+        pad_list[len(pad_list) - 2 * self.dim_index - 1] = pad_dim
+
+        return torch.nn.functional.pad(input_tensor, pad_list)
+
+    def _la_preprocess_input(
+        self, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        # Currently BSND input layout is not supported
+        q = query.transpose(1, 2)
+        k = key.transpose(1, 2)
+        v = value.transpose(1, 2)
+
+        if q.dtype != torch.float16:
+            q = q.to(torch.float16)
+            k = k.to(torch.float16)
+            v = v.to(torch.float16)
+
+        q = self._pad(q)
+        k = self._pad(k)
+        v = self._pad(v)
+
+        return q, k, v
+
+    def _la_postprocess_output(
+        self,
+        attention_out: torch.Tensor,
+        dtype: torch.dtype,
+        qseqlen: int,
+        head_dim: int,
+    ) -> torch.Tensor:
+        if dtype != attention_out.dtype:
+            attention_out = attention_out.to(dtype)
+
+        attention_out = attention_out[:, :, :qseqlen, :head_dim]
+        attention_out = attention_out.transpose(1, 2).contiguous()
+        return attention_out
+
+    def _laser_attention(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        head_num: int,
+        pre_tokens: int,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        return torch.ops.attentions.la(
+            query=query,
+            key=key,
+            value=value,
+            atten_mask=None,
+            alibi_mask=None,
+            drop_mask=None,
+            scale_value=self.softmax_scale,
+            head_num=head_num,
+            input_layout="BNSD",
+            keep_prob=1.0,
+            pre_tokens=pre_tokens,
+            next_tokens=1,
+            is_highPrecision=True,
+        )
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        q_seqlen, head_dim = query.shape[1], query.shape[3]
+        kv_seqlen = key.shape[1]
+
+        if q_seqlen < self.min_seqlen or kv_seqlen != q_seqlen:
+            output = self.sdpa_impl.forward(
+                query,
+                key,
+                value,
+                attn_metadata,
+            )
+        else:
+            pre_tokens = self.max_token
+            if kv_seqlen % self.seq_len_pad_base != 0:
+                pre_tokens = (
+                    kv_seqlen // self.seq_len_pad_base + 1
+                ) * self.seq_len_pad_base - kv_seqlen
+
+            q, k, v = self._la_preprocess_input(query, key, value)
+            _, la_output = self._laser_attention(q, k, v, q.shape[1], pre_tokens)
+            output = self._la_postprocess_output(
+                la_output, query.dtype, q_seqlen, head_dim
+            )
+
+        return output

python/sglang/multimodal_gen/runtime/layers/attention/backends/rain_fusion_attn.py

@@ -0,0 +1,414 @@
+import math
+from dataclasses import dataclass
+from typing import Any, List, Optional
+
+import attentions  # noqa: F401
+import torch
+from einops import rearrange
+
+from sglang.multimodal_gen.runtime.layers.attention.backends.attention_backend import (
+    AttentionBackend,
+    AttentionImpl,
+    AttentionMetadata,
+    AttentionMetadataBuilder,
+)
+from sglang.multimodal_gen.runtime.layers.attention.backends.laser_attn import (
+    LaserAttentionBackend,
+)
+from sglang.multimodal_gen.runtime.platforms import AttentionBackendEnum
+from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
+
+logger = init_logger(__name__)
+
+
+class RainFusionAttentionBackend(AttentionBackend):
+
+    accept_output_buffer: bool = True
+
+    @staticmethod
+    def get_supported_head_sizes() -> list[int]:
+        return [32, 64, 96, 128]
+
+    @staticmethod
+    def get_enum() -> AttentionBackendEnum:
+        return AttentionBackendEnum.RAIN_FUSION_ATTN
+
+    @staticmethod
+    def get_impl_cls() -> type["RainFusionAttentionImpl"]:
+        return RainFusionAttentionImpl
+
+    @staticmethod
+    def get_metadata_cls() -> type["RainFusionAttentionMetadata"]:
+        return RainFusionAttentionMetadata
+
+    @staticmethod
+    def get_builder_cls() -> type["RainFusionAttentionMetadataBuilder"]:
+        return RainFusionAttentionMetadataBuilder
+
+
+@dataclass
+class RainFusionAttentionMetadata(AttentionMetadata):
+    current_timestep: int
+    skip_first_steps: int
+    sparsity: float
+    latent_shape: list[int]
+
+
+class RainFusionAttentionMetadataBuilder(AttentionMetadataBuilder):
+    def __init__(self) -> None:
+        pass
+
+    def prepare(self) -> None:
+        pass
+
+    def build(
+        self,
+        current_timestep: int,
+        skip_first_steps: int,
+        sparsity: float,
+        raw_latent_shape: list[int],
+        patch_size: tuple[int, int, int],
+        **kwargs: dict[str, Any],
+    ) -> RainFusionAttentionMetadata:
+        if not (skip_first_steps >= 0 and 0.0 <= sparsity < 1.0):
+            raise ValueError(
+                (
+                    "Invalid attention metadata values."
+                    f"Sparsity should be in [0, 1), skip_first_steps should be non-negative."
+                    f"Got sparsity={sparsity}, skip_first_steps={skip_first_steps}"
+                )
+            )
+
+        if sparsity == 0.0:
+            logger.warning(
+                (
+                    "Sparsity is set to 0.0, which means no tokens will be dropped."
+                    "For better performance use Laser Attention or increase sparsity."
+                )
+            )
+
+        latent_shape = raw_latent_shape[-3:]
+        latent_shape = [latent_shape[i] // patch_size[i] for i in range(3)]
+
+        return RainFusionAttentionMetadata(
+            current_timestep=current_timestep,
+            skip_first_steps=skip_first_steps,
+            sparsity=sparsity,
+            latent_shape=latent_shape,
+        )
+
+
+class RainFusionAttentionImpl(AttentionImpl):
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        causal: bool,
+        softmax_scale: float,
+        num_kv_heads: int | None = None,
+        prefix: str = "",
+        **extra_impl_args,
+    ) -> None:
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.block_size = 128
+        self.inner_precise = 0
+
+        self.laser_attn_impl = LaserAttentionBackend.get_impl_cls()(
+            num_heads,
+            head_size,
+            causal,
+            softmax_scale,
+            num_kv_heads,
+            prefix,
+            **extra_impl_args,
+        )
+
+    def _avgpool(
+        self, input_tensor: torch.Tensor, pool_size: int = 128
+    ) -> torch.Tensor:
+        batch, seqlen, heads, dim = input_tensor.shape
+        x = input_tensor.permute(0, 2, 3, 1).reshape(batch * heads, dim, seqlen)
+
+        pooled = torch.nn.functional.avg_pool1d(
+            x, kernel_size=pool_size, stride=pool_size, ceil_mode=True
+        )
+        out = pooled.reshape(batch, heads, dim, -1).permute(0, 3, 1, 2).contiguous()
+
+        return out
+
+    def _get_mask_index(self, mask: torch.Tensor) -> torch.Tensor:
+        batch_size, num_heads, seq_len, _ = mask.shape
+
+        mask_reshaped = mask.reshape(-1, seq_len)
+        row_indices = torch.arange(
+            seq_len, device=mask.device, dtype=torch.float32
+        ).unsqueeze(0)
+
+        sorted_vals = torch.where(mask_reshaped, row_indices, seq_len)
+        sorted_vals, _ = torch.sort(sorted_vals, dim=-1)
+        valid_count = mask_reshaped.sum(dim=-1, keepdim=True)
+        keep_mask = row_indices < valid_count
+        result = torch.where(keep_mask, sorted_vals, -1)
+
+        pos_matrix = result.reshape(batch_size, num_heads, seq_len, seq_len).to(
+            torch.int64
+        )
+        return pos_matrix
+
+    def _get_blockwise_mask(
+        self,
+        qkv_pool: torch.Tensor,
+        sparsity: float,
+        scale: float,
+        pool_size: int,
+        latent_shape: tuple,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        first_frame_len = latent_shape[1] * latent_shape[2]
+
+        query_pool, key_pool, value_pool = torch.chunk(qkv_pool, 3, dim=0)
+        attn_scores = (
+            query_pool.permute(0, 2, 1, 3) @ key_pool.permute(0, 2, 3, 1) * scale
+        )
+
+        keep_len = math.ceil(attn_scores.shape[-1] * (1 - sparsity))
+
+        topk_values, _ = torch.topk(attn_scores, k=keep_len, dim=-1)
+        mask = attn_scores >= topk_values[..., -1:]
+
+        firstframe_block_num = (first_frame_len + pool_size - 1) // pool_size
+        if firstframe_block_num > 0:
+            mask[:, :, :firstframe_block_num, :] = True
+            mask[:, :, :, :firstframe_block_num] = True
+
+        select_idx = self._get_mask_index(mask)
+        select_idx = select_idx[0].transpose(0, 1)
+        select_num_idx = mask[0].transpose(0, 1).sum(dim=-1)
+        return select_idx, select_num_idx
+
+    def _rearrange_with_remaining(
+        self, tensor: torch.Tensor, latent_shape: tuple[int, int, int]
+    ) -> torch.Tensor:
+        """
+        b (f hn hb wn wb) n d -> b (f hn wn hb wb) n d
+        or
+        b n (f hn hb wn wb) d -> b n (f hn wn hb wb) d
+        """
+        tq, hq, wq = latent_shape
+        first_frame_len, frame_num = hq * wq, tq
+
+        b, s, n, d = tensor.shape
+
+        if (hq % 8 != 0) or (wq % 8 != 0):
+            tensor_first = tensor[:, :first_frame_len, :, :]
+            tensor = tensor[:, first_frame_len:, :, :]
+            tensor_hwt = rearrange(
+                tensor, "b (f h w) n d -> b f h w n d", f=frame_num - 1, h=hq, w=wq
+            )
+            if hq % 8 != 0:
+                tensor_hwt, tensor_h_r = torch.split(tensor_hwt, hq - (hq % 8), dim=2)
+                tensor_h_r = tensor_h_r.reshape(b, frame_num - 1, -1, n, d)
+            if wq % 8 != 0:
+                tensor_hwt, tensor_w_r = torch.split(tensor_hwt, wq - (wq % 8), dim=3)
+                tensor_w_r = tensor_w_r.reshape(b, frame_num - 1, -1, n, d)
+            tensor_hwt = rearrange(
+                tensor_hwt,
+                "b f (hn hb) (wn wb) n d -> b f (hn wn hb wb) n d",
+                f=frame_num - 1,
+                hb=8,
+                wb=8,
+                hn=hq // 8,
+                wn=wq // 8,
+            )
+            if hq % 8 != 0:
+                tensor_hwt = torch.cat((tensor_hwt, tensor_h_r), dim=2)
+            if wq % 8 != 0:
+                tensor_hwt = torch.cat((tensor_hwt, tensor_w_r), dim=2)
+            tensor_hwt = tensor_hwt.reshape(b, -1, n, d)
+            tensor_hwt = torch.cat([tensor_first, tensor_hwt], dim=1)
+        else:
+            tensor_hwt = rearrange(
+                tensor,
+                "b (f hn hb wn wb) n d -> b (f hn wn hb wb) n d",
+                f=frame_num,
+                hb=8,
+                wb=8,
+                hn=hq // 8,
+                wn=wq // 8,
+            )
+
+        return tensor_hwt
+
+    def _inv_rearrange_with_remaining(
+        self, tensor: torch.Tensor, latent_shape: tuple[int, int, int]
+    ) -> torch.Tensor:
+        tq, hq, wq = latent_shape
+        first_frame_len, frame_num = hq * wq, tq
+
+        b, s, n, d = tensor.shape
+
+        if (hq % 8 != 0) or (wq % 8 != 0):
+            tensor_first = tensor[:, :first_frame_len, :, :]
+            tensor = tensor[:, first_frame_len:, :, :]
+            tensor_hwt = rearrange(
+                tensor, "b (f h w) n d -> b f h w n d", f=frame_num - 1, h=hq, w=wq
+            )
+            if hq % 8 != 0:
+                tensor_hwt, tensor_h_r = torch.split(tensor_hwt, hq - (hq % 8), dim=2)
+            if wq % 8 != 0:
+                tensor_hwt, tensor_w_r = torch.split(tensor_hwt, wq - (wq % 8), dim=3)
+            tensor_hwt = tensor_hwt.reshape(b, frame_num - 1, -1, n, d)
+            tensor_hwt = rearrange(
+                tensor_hwt,
+                "b f (hn wn hb wb) n d -> b f (hn hb) (wn wb) n d",
+                f=frame_num - 1,
+                hb=8,
+                wb=8,
+                hn=hq // 8,
+                wn=wq // 8,
+            )
+            if wq % 8 != 0:
+                tensor_hwt = torch.cat((tensor_hwt, tensor_w_r), dim=3)
+            if hq % 8 != 0:
+                tensor_hwt = torch.cat((tensor_hwt, tensor_h_r), dim=2)
+            tensor_hwt = tensor_hwt.reshape(b, -1, n, d)
+            tensor_hwt = torch.cat([tensor_first, tensor_hwt], dim=1)
+        else:
+            tensor_hwt = rearrange(
+                tensor,
+                "b (f hn wn hb wb) n h -> b (f hn hb wn wb) n h",
+                f=frame_num,
+                hb=8,
+                wb=8,
+                hn=hq // 8,
+                wn=wq // 8,
+            )
+
+        return tensor_hwt
+
+    def _do_tensor_rearrange_pooling(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        pool_size: int,
+        latent_shape: tuple[int, int, int],
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Tensor block rearrangement + pooling operation
+        """
+        tensor = torch.cat((query, key, value), dim=0)
+
+        tensor = self._rearrange_with_remaining(tensor, latent_shape)
+        tensor_pool = self._avgpool(tensor, pool_size)
+
+        query_, key_, value_ = torch.chunk(tensor, 3, dim=0)
+        return query_, key_, value_, tensor_pool
+
+    def _rain_fusion_attention(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        select_idx: torch.Tensor,
+        select_num_idx: torch.Tensor,
+        blockshape: List[int],
+        scale: float = 1.0,
+        head_num: int = 1,
+        input_layout: str = "TND",
+        actual_seq_lengths=Optional[torch.Tensor],
+        actual_seq_lengths_kv=Optional[torch.Tensor],
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        return torch.ops.attentions.rainfusionattention(
+            query=query,
+            key=key,
+            value=value,
+            select_idx=select_idx,
+            select_num_idx=select_num_idx,
+            blockshape=blockshape,
+            attn_mask=None,
+            actual_seq_qlen=actual_seq_lengths,
+            actual_seq_kvlen=actual_seq_lengths_kv,
+            block_table=None,
+            q_input_layout=input_layout,
+            kv_input_layout=input_layout,
+            head_num=head_num,
+            mask_type=0,
+            scale=scale,
+            inner_precise=self.inner_precise,
+            block_size=0,
+        )
+
+    def _rain_fusion_sparse_attention(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        latent_shape: tuple[int, int, int],
+        sparsity: float,
+    ):
+        q, k, v, qkv_pool = self._do_tensor_rearrange_pooling(
+            query, key, value, self.block_size, latent_shape
+        )
+
+        select_idx, select_num_idx = self._get_blockwise_mask(
+            qkv_pool,
+            sparsity,
+            self.softmax_scale,
+            self.block_size,
+            latent_shape,
+        )
+
+        batch_size, seqlen_q, head_num, head_dim = q.shape
+        seqlen_kv = k.shape[1]
+
+        layout = "TND"
+        q = q.reshape(-1, head_num, head_dim)
+        k = k.reshape(-1, head_num, head_dim)
+        v = v.reshape(-1, head_num, head_dim)
+
+        actual_seq_lengths = [seqlen_q] * batch_size
+        actual_seq_lengths_kv = [seqlen_kv] * batch_size
+
+        out, _ = self._rain_fusion_attention(
+            q,
+            k,
+            v,
+            scale=self.softmax_scale,
+            head_num=head_num,
+            input_layout=layout,
+            select_idx=select_idx,
+            select_num_idx=select_num_idx,
+            blockshape=[self.block_size, self.block_size],
+            actual_seq_lengths=actual_seq_lengths,
+            actual_seq_lengths_kv=actual_seq_lengths_kv,
+        )
+
+        out = out.reshape(batch_size, seqlen_q, head_num, head_dim)
+        out = self._inv_rearrange_with_remaining(out, latent_shape)
+        return out
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        if attn_metadata.current_timestep < attn_metadata.skip_first_steps:
+            output = self.laser_attn_impl.forward(
+                query,
+                key,
+                value,
+                attn_metadata,
+            )
+        else:
+            output = self._rain_fusion_sparse_attention(
+                query,
+                key,
+                value,
+                attn_metadata.latent_shape,
+                attn_metadata.sparsity,
+            )
+        return output

python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py

@@ -1582,6 +1582,35 @@ class DenoisingStage(PipelineStage, RolloutDenoisingMixin):
             attn_metadata = self.attn_metadata_builder.build(
                 raw_latent_shape=batch.raw_latent_shape
             )
+        elif self.attn_backend.get_enum() in [
+            AttentionBackendEnum.BLOCK_SPARSE_ATTN,
+            AttentionBackendEnum.RAIN_FUSION_ATTN,
+        ]:
+            sparse_config = server_args.attention_backend_config
+
+            current_timestep = i
+            skip_first_steps = sparse_config.get("skip_first_steps", 10)
+            sparsity = sparse_config.get("sparsity", 0.2)
+
+            raw_latent_shape = batch.raw_latent_shape
+            patch_size = server_args.pipeline_config.dit_config.patch_size
+
+            if isinstance(patch_size, int):
+                patch_size_t = getattr(
+                    server_args.pipeline_config.dit_config, "patch_size_t", None
+                )
+                if patch_size_t is not None:
+                    patch_size = (patch_size_t, patch_size, patch_size)
+                else:
+                    patch_size = (patch_size, patch_size, patch_size)
+
+            attn_metadata = self.attn_metadata_builder.build(
+                current_timestep=current_timestep,
+                skip_first_steps=skip_first_steps,
+                sparsity=sparsity,
+                raw_latent_shape=raw_latent_shape,
+                patch_size=patch_size,
+            )
         else:
             # attn_metadata can be None for SDPA attention backend
             return None

python/sglang/multimodal_gen/runtime/platforms/interface.py

@@ -38,6 +38,9 @@ class AttentionBackendEnum(enum.Enum):
     AITER_SAGE = enum.auto()
     SLA_ATTN = enum.auto()
     SAGE_SLA_ATTN = enum.auto()
+    LASER_ATTN = enum.auto()
+    BLOCK_SPARSE_ATTN = enum.auto()
+    RAIN_FUSION_ATTN = enum.auto()
     NO_ATTENTION = enum.auto()
 
     def __str__(self):
@@ -52,6 +55,9 @@ class AttentionBackendEnum(enum.Enum):
             AttentionBackendEnum.VMOBA_ATTN,
             AttentionBackendEnum.SLA_ATTN,
             AttentionBackendEnum.SAGE_SLA_ATTN,
+            AttentionBackendEnum.LASER_ATTN,
+            AttentionBackendEnum.BLOCK_SPARSE_ATTN,
+            AttentionBackendEnum.RAIN_FUSION_ATTN,
         }
 
 

python/sglang/multimodal_gen/runtime/platforms/npu.py

@@ -120,6 +120,57 @@ class NPUPlatformBase(Platform):
             logger.info("Using Ascend Flash Attention backend.")
             return "sglang.multimodal_gen.runtime.layers.attention.backends.ascend_fa.AscendFABackend"
 
+        elif selected_backend == AttentionBackendEnum.LASER_ATTN:
+            try:
+                from sglang.multimodal_gen.runtime.layers.attention.backends.laser_attn import (  # noqa: F401
+                    LaserAttentionBackend,
+                )
+
+                logger.info("Using Laser Attention backend")
+
+                return "sglang.multimodal_gen.runtime.layers.attention.backends.laser_attn.LaserAttentionBackend"
+            except ImportError as e:
+                logger.error(f"Failed to import Laser Attention backend: {e}")
+                raise ImportError(
+                    "Laser Attention backend is not installed. "
+                    "It requires the `attentions` module which can be installed along with sgl_kernel_npu. "
+                    "Manual installation from source is required. See https://github.com/sgl-project/sgl-kernel-npu."
+                ) from e
+
+        elif selected_backend == AttentionBackendEnum.BLOCK_SPARSE_ATTN:
+            try:
+                from sglang.multimodal_gen.runtime.layers.attention.backends.block_sparse_attn import (  # noqa: F401
+                    BlockSparseAttentionBackend,
+                )
+
+                logger.info("Using Block Sparse Attention backend")
+
+                return "sglang.multimodal_gen.runtime.layers.attention.backends.block_sparse_attn.BlockSparseAttentionBackend"
+            except ImportError as e:
+                logger.error(f"Failed to import Block Sparse Attention backend: {e}")
+                raise ImportError(
+                    "Block Sparse Attention backend is not installed. "
+                    "It requires the `attentions` module which can be installed along with sgl_kernel_npu. "
+                    "Manual installation from source is required. See https://github.com/sgl-project/sgl-kernel-npu."
+                ) from e
+
+        elif selected_backend == AttentionBackendEnum.RAIN_FUSION_ATTN:
+            try:
+                from sglang.multimodal_gen.runtime.layers.attention.backends.rain_fusion_attn import (  # noqa: F401
+                    RainFusionAttentionBackend,
+                )
+
+                logger.info("Using Rain Fusion Attention backend")
+
+                return "sglang.multimodal_gen.runtime.layers.attention.backends.rain_fusion_attn.RainFusionAttentionBackend"
+            except ImportError as e:
+                logger.error(f"Failed to import Rain Fusion Attention backend: {e}")
+                raise ImportError(
+                    "Rain Fusion Attention backend is not installed. "
+                    "It requires the `attentions` module which can be installed along with sgl_kernel_npu. "
+                    "Manual installation from source is required. See https://github.com/sgl-project/sgl-kernel-npu."
+                ) from e
+
         logger.info("Using Torch SDPA backend.")
         return (
             "sglang.multimodal_gen.runtime.layers.attention.backends.sdpa.SDPABackend"