[Model] Support SenseNova-U1.5-8B-MoT and its distilled 8-step LoRA (#6516)

Signed-off-by: MrlixiangWE <mrdanaer@gmail.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

docs/models/supported_models.md

@@ -95,7 +95,7 @@ th {
 | `MiMoV2ASRForCausalLM` | MiMo-V2.5-ASR | `XiaomiMiMo/MiMo-V2.5-ASR` | ✅︎ | ✅︎ | | | — |
 | `Flux2Pipeline` | FLUX.2-dev | `black-forest-labs/FLUX.2-dev` | ✅︎ | ✅︎ | | | — |
 | `FishSpeechSlowARForConditionalGeneration` | Fish Speech S2 Pro | `fishaudio/s2-pro` | ✅︎ | ✅︎ | | ✅︎ | — |
-| `SenseNovaU1Pipeline` | SenseNova-U1 (DiT-only) | `SenseNova/SenseNova-U1-8B-MoT` | ✅︎ | | | | — |
+| `SenseNovaU1Pipeline` | SenseNova-U1 (DiT-only), SenseNova-U1.5 | `SenseNova/SenseNova-U1-8B-MoT`, `sensenova/SenseNova-U1.5-8B-MoT` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/SenseNova/SenseNova-U1.5.md) |
 | `LancePipeline` | Lance | `bytedance-research/Lance` | ✅︎ | | | | — |
 | `HunyuanVideo15Pipeline` | HunyuanVideo-1.5-T2V | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v`, `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v` | ✅︎ | ✅︎ | | | — |
 | `HunyuanVideo15ImageToVideoPipeline` | HunyuanVideo-1.5-I2V | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v`, `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_i2v` | ✅︎ | ✅︎ | | | — |

docs/user_guide/diffusion_features.md

@@ -138,7 +138,7 @@ The following tables show which models support each feature:
 | **Qwen-Image-Edit**      |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **Qwen-Image-Edit-2509** |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |        ✅ (decode)         |          ✅           |       ❌        |        ❌         |
 | **Qwen-Image-Layered**   |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
-| **SenseNova-U1**         |     ❌     |     ✅      |           ❌           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
+| **SenseNova-U1 / U1.5**  |     ❌     |     ✅      |           ❌           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **Stable-Diffusion-XL**  |     ❌     |     ❌      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **Stable-Diffusion3.5**  |     ❌     |     ✅      |           ❌           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **Z-Image**              |     ✅     |     ✅      |           ✅           |       ❓        |   ✅ (TP=2 only)   |          ❌          |   ✅    |             ❌             |      ✅ (decode)      |       ✅        |        ❌         |

recipes/README.md

@@ -78,6 +78,7 @@ recipes/
 | [`Qwen/Qwen3-Omni.md`](./Qwen/Qwen3-Omni.md) | Online serving for multimodal chat | 1x A100 80GB |
 | [`Qwen/Qwen3-TTS.md`](./Qwen/Qwen3-TTS.md) | Text-to-speech serving (CustomVoice / VoiceDesign / Base) | 1x H100/A100 80GB / 1x AMD MI300X |
 | [`SenseNova/SenseNova-U1.md`](./SenseNova/SenseNova-U1.md) | Unified image generation and understanding | 1x or 2x H200 / 1x AMD MI300X |
+| [`SenseNova/SenseNova-U1.5.md`](./SenseNova/SenseNova-U1.5.md) | Unified image generation and understanding, with an 8-step distilled LoRA | 1x A800 80GB |
 | [`StabilityAI/Stable-Audio-Open.md`](./StabilityAI/Stable-Audio-Open.md) | Offline + online text-to-audio generation (Stable Audio Open) | 1x RTX 4090 24GB / 1x AMD MI300X |
 | [`Tencent/Covo-Audio-Chat.md`](./Tencent/Covo-Audio-Chat.md) | Online serving for audio chat | 1x A100 80GB |
 | [`Tencent/HunyuanImage-3.0-Instruct.md`](./Tencent/HunyuanImage-3.0-Instruct.md) | DiT-only text-to-image serving and benchmark, including ModelOpt mixed FP8/NVFP4 | 4x H100/H800 80GB / 2x B200 |

recipes/SenseNova/SenseNova-U1.5.md

@@ -0,0 +1,159 @@
+# SenseNova-U1.5
+
+> Unified image generation and understanding, with an official 8-step distilled LoRA
+
+## Summary
+
+- Vendor: SenseNova
+- Model: `sensenova/SenseNova-U1.5-8B-MoT`
+- LoRA: `sensenova/SenseNova-U1.5-8B-MoT-LoRAs` (`SenseNova-U1.5-8B-MoT-LoRA-8step.safetensors`)
+- Task: text2img, img2img, img2text (visual understanding), text2text (chat)
+- Mode: Offline inference, Online serving (OpenAI-compatible API)
+- Maintainer: Community
+
+## When to use this recipe
+
+U1.5 runs on the same `SenseNovaU1Pipeline` as U1 — the checkpoint keeps
+`model_type: neo_chat`, so it resolves without `--model-class-name`. Relative to U1 the
+config flips two fields, `use_pixel_head` to `true` (the flow-matching head becomes a
+`ConvDecoder`) and `noise_scale_max_value` from 8.0 to 16.0; both are already read by
+`SenseNovaU1Config`. The checkpoint is 13 shards / 50.2 GB on disk, 30.3 GB of that fp32, and
+loads to roughly 34 GB in bf16.
+
+Use this recipe for U1.5 specifically, including its distilled few-step LoRA. For U1 see
+[SenseNova-U1](SenseNova-U1.md).
+
+## Hardware Support
+
+### GPU
+
+#### 1x A800 80GB
+
+- 1x NVIDIA A800 80GB PCIe with an Intel Xeon Gold 6336Y, driver 595.84, Linux
+- Python 3.12.13, vLLM 0.28.0, torch 2.13.0+cu130, CUDA 13.0, diffusers 0.40.0,
+  transformers 5.14.1, flashinfer 0.6.16.post3, BF16, TP=1
+- Peak GPU memory: 34.3 GB at 1024x1024, 36.4 GB at 1536x2720
+
+##### Text-to-Image
+
+```bash
+python examples/offline_inference/text_to_image/text_to_image.py \
+    --model sensenova/SenseNova-U1.5-8B-MoT \
+    --prompt "Close portrait of an elderly woman by a farmhouse window, warm natural light." \
+    --width 1024 --height 1024 \
+    --seed 42 --num-inference-steps 50 --cfg-scale 4.0 \
+    --extra-body '{"think": false, "cfg_norm": "none", "timestep_shift": 3.0, "t_eps": 0.02}' \
+    --output sensenova_u15_t2i.png
+```
+
+Think mode (`"think": true`) is recommended for higher image quality.
+
+##### Image-to-Image Editing
+
+```bash
+python examples/offline_inference/image_to_image/image_edit.py \
+    --model sensenova/SenseNova-U1.5-8B-MoT \
+    --prompt "Turn this into an oil painting" \
+    --image input.png --resolution 1024 \
+    --seed 42 --num-inference-steps 25 --cfg-scale 4.0 \
+    --extra-args '{"think": true, "img_cfg_scale": 1.0, "cfg_norm": "none", "timestep_shift": 3.0}' \
+    --output sensenova_u15_edit.png
+```
+
+##### 8-step distilled LoRA
+
+The distilled LoRA is fused into the generation tower at load time:
+
+```bash
+python examples/offline_inference/text_to_image/text_to_image.py \
+    --model sensenova/SenseNova-U1.5-8B-MoT \
+    --lora-path SenseNova-U1.5-8B-MoT-LoRA-8step.safetensors --lora-backend distill \
+    --prompt "Close portrait of an elderly woman by a farmhouse window, warm natural light." \
+    --width 1024 --height 1024 \
+    --seed 42 --num-inference-steps 8 --cfg-scale 1.0 \
+    --extra-body '{"think": false, "cfg_norm": "none", "timestep_shift": 3.0, "t_eps": 0.02}' \
+    --output sensenova_u15_lora8.png
+```
+
+**Use `--cfg-scale 1.0` with this LoRA.** It is distilled with DMD and runs without
+classifier-free guidance; the default `4.0` applies guidance twice and produces a
+blown-out, posterised image.
+
+##### Online serving
+
+```bash
+vllm serve sensenova/SenseNova-U1.5-8B-MoT --omni --port 8091
+
+python examples/online_serving/sensenova_u1/openai_chat_client.py \
+    -s http://127.0.0.1:8091 -m img2text -i input.png -p "Describe this image."
+```
+
+`-s` takes the base URL; the client appends `/v1` itself.
+
+#### Measured latency (1x A800 80GB, 25 steps, median of 3 after a warmup)
+
+| Resolution | Step latency | Total |
+| --- | --- | --- |
+| 1024x1024 | 209.5 ms | 5.24 s |
+| 1536x1536 | 458.5 ms | 11.46 s |
+
+At 1536x2720 with 50 steps, end-to-end is 43.0 s.
+
+#### Measured latency (1x NVIDIA H200 139GB, single run, no warmup)
+
+Reported by @hsliuustc0106 against PR head `29c090d`, on one reserved CUDA device.
+
+- Python 3.12.13, vLLM 0.28.0, torch 2.13.0+cu130, CUDA 13.0, diffusers 0.40.0,
+  transformers 5.14.1, flashinfer 0.6.16.post3, BF16, TP=1
+- `sensenova/SenseNova-U1.5-8B-MoT` with `SenseNova-U1.5-8B-MoT-LoRA-8step.safetensors`,
+  seed 42, 1024x1024
+
+| Case | Stage latency | Peak GPU memory |
+| --- | --- | --- |
+| 50 steps, cfg 4.0, think off | 9017 ms | 34384 MiB |
+| 8 steps, cfg 1.0, no LoRA control | 624.00 ms, 77.85 ms/step | 34376 MiB |
+| 8 steps, cfg 1.0, distilled LoRA | 623.85 ms, 77.98 ms/step | 34364 MiB |
+| 50 steps, cfg 4.0, think on, `VLLM_OMNI_SENSENOVA_PAGED_DECODE=1` | 81622 ms | 34594 MiB |
+
+Each row runs the Text-to-Image command above with `--num-inference-steps` and
+`--cfg-scale` as listed and `think` set through `--extra-body`; the LoRA row adds
+`--lora-path` and `--lora-backend distill`.
+
+Single runs without a preceding warmup, so the think-on row carries the first-request
+compile rather than a steady state. The distilled LoRA fused into 168 parameters, and its
+image differs from the same-seed no-LoRA control in 1,048,574 of 1,048,576 pixels, MAE 58.62.
+
+```bash
+pytest -m "cpu and not cuda" tests/diffusion/models/sensenova_u1/ tests/diffusion/lora/test_loader.py -q
+pytest -m "cpu and not cuda" tests/config/test_environment_variables.py tests/diffusion/test_diffusion_worker.py -q
+pytest -m cuda tests/diffusion/models/sensenova_u1/test_sensenova_u1_attention_gqa.py \
+    tests/diffusion/models/sensenova_u1/test_sensenova_u1_paged_decode.py -q --tb=short
+```
+
+71, 13 and 9 passed. Serving reached `/health`; `img2text` returned a 1921-character
+description and `text2text` the expected answer.
+
+#### Verification
+
+```bash
+pytest -q tests/diffusion/models/sensenova_u1/
+```
+
+## Notes
+
+- Both `use_pixel_head` and `noise_scale_max_value` come from `config.json`; no flag is needed.
+- The checkpoint carries no `configuration_neo_chat.py`, so the loader logs a few
+  "does not appear to have a file named configuration_neo_chat.py" errors during startup and
+  then proceeds on the in-tree config. Generation is unaffected.
+- The LoRA targets the generation tower only (`*_mot_gen`); understanding-tower weights are
+  untouched.
+- Autoregressive decode (think, text-to-text and image-to-text) runs on a paged K/V cache under
+  a captured CUDA graph. It falls back to the ordinary cache when the device or the bundled
+  `flash_attn_varlen_func` cannot support it; set `VLLM_OMNI_SENSENOVA_PAGED_DECODE=0` to force
+  that fallback.
+- The first request after startup costs about 0.7 s more than the steady state whether the paged
+  path is on or off. Measured on one A800 with the inductor, triton and vLLM compile caches all
+  cleared, median of three runs: 718 ms above steady with the path on, 679 ms with it off.
+- Each request captures its own graphs, and a think request captures twice because the sequence
+  grows past the 512 bucket, so the capture cost is paid per request rather than once at
+  startup.

tests/config/test_environment_variables.py

@@ -167,7 +167,7 @@ def test_inventory_matches_reviewed_snapshot_counts():
         EnvironmentVariableCategory.PUBLIC_OMNI: 24,
         EnvironmentVariableCategory.INHERITED_VLLM: 20,
         EnvironmentVariableCategory.PLATFORM_EXTERNAL: 27,
-        EnvironmentVariableCategory.MODEL_SPECIFIC: 56,
+        EnvironmentVariableCategory.MODEL_SPECIFIC: 57,
         EnvironmentVariableCategory.BENCHMARK_TRANSITIONAL: 20,
         EnvironmentVariableCategory.INTERNAL: 2,
     }
@@ -178,7 +178,7 @@ def test_inventory_matches_reviewed_snapshot_counts():
         if item.category is EnvironmentVariableCategory.MODEL_SPECIFIC
     )
     assert {disposition: disposition_counts[disposition] for disposition in ModelEnvironmentVariableDisposition} == {
-        ModelEnvironmentVariableDisposition.PROMOTE: 33,
+        ModelEnvironmentVariableDisposition.PROMOTE: 34,
         ModelEnvironmentVariableDisposition.REQUEST_SCOPE: 6,
         ModelEnvironmentVariableDisposition.EXTERNAL: 0,
         ModelEnvironmentVariableDisposition.INTERNALIZE: 11,

tests/diffusion/lora/test_loader.py

@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 from collections import defaultdict
 
 import pytest
@@ -106,7 +109,7 @@ def make_dummy_lora_state_dict(
     """
     key_a = f"{base_key}.{lora_a_suffix}"
     key_b = f"{base_key}.{lora_b_suffix}"
-    key_bias = f"{base_key}.{lora_bias_suffix}" if bias else None
+    key_bias = f"{base_key}.{lora_bias_suffix}"
 
     if weight_a is None:
         weight_a = torch.randn(RANK, HEAD_DIM)
@@ -436,3 +439,54 @@ def test_wan_load_lora_weights_targets_both_transformers(mocker: MockerFixture):
     assert_params_equal(pipeline.transformer, original_high)
     assert_params_equal(pipeline.transformer_2, original_low)
     assert "adapter0" not in pipeline.lora_loaded
+
+
+def test_prepare_lora_delta_disambiguates_prefix_overlapping_stacked_names():
+    """A fused name can be a prefix of another fused name.
+
+    SenseNova-U1 carries both `.qkv_proj` (understanding) and `.qkv_proj_mot_gen`
+    (generation). Selecting stacked mappings by substring matched both rules for
+    the generation weight and concatenated six shards into a delta twice as tall
+    as the parameter.
+    """
+    param_to_weight_names = {
+        ".qkv_proj_mot_gen": [".q_proj_mot_gen", ".k_proj_mot_gen", ".v_proj_mot_gen"],
+        ".qkv_proj": [".q_proj", ".k_proj", ".v_proj"],
+    }
+    state_dict = {}
+    for part in ("q", "k", "v"):
+        for tower in ("_mot_gen", ""):
+            base = f"layers.0.self_attn.{part}_proj{tower}"
+            state_dict[f"{base}.lora_A.weight"] = torch.ones(RANK, HEAD_DIM)
+            state_dict[f"{base}.lora_B.weight"] = torch.ones(HEAD_DIM, RANK)
+
+    for base_key in ("layers.0.self_attn.qkv_proj_mot_gen", "layers.0.self_attn.qkv_proj"):
+        delta, _ = _prepare_lora_delta(state_dict, base_key, param_to_weight_names)
+        assert delta.shape == (3 * HEAD_DIM, HEAD_DIM), base_key
+
+
+@pytest.mark.parametrize(
+    "fused, shards",
+    [
+        # qwen_image_transformer.py stacked_params_mapping
+        (".to_qkv", (".to_q", ".to_k", ".to_v")),
+        (".add_kv_proj", (".add_q_proj", ".add_k_proj", ".add_v_proj")),
+        # wan2_2_transformer.py stacked_params_mapping
+        (".attn1.to_qkv", (".attn1.to_q", ".attn1.to_k", ".attn1.to_v")),
+        # wan2_2_s2v_transformer.py stacked_params_mapping
+        (".self_attn.to_qkv", (".self_attn.q", ".self_attn.k", ".self_attn.v")),
+    ],
+)
+def test_prepare_lora_delta_unchanged_for_existing_fused_layouts(fused, shards):
+    """None of the fused names shipped today overlap by prefix, so matching on the
+    tail must select exactly the shards substring matching selected."""
+    param_to_weight_names = {fused: list(shards)}
+    base_key = f"blocks.0{fused}"
+    state_dict = {}
+    for shard in shards:
+        state_dict[f"blocks.0{shard}.lora_A.weight"] = torch.ones(RANK, HEAD_DIM)
+        state_dict[f"blocks.0{shard}.lora_B.weight"] = torch.ones(HEAD_DIM, RANK)
+
+    delta, used_keys = _prepare_lora_delta(state_dict, base_key, param_to_weight_names)
+    assert delta.shape == (len(shards) * HEAD_DIM, HEAD_DIM)
+    assert len(used_keys) == 2 * len(shards)

tests/diffusion/models/sensenova_u1/test_sensenova_u1_attention_gqa.py

@@ -0,0 +1,195 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""CPU tests for SenseNova-U1 attention head handling.
+
+The model used to expand K/V from 8 heads to 32 with ``repeat_interleave``
+before calling the attention backend, so the backend's own GQA dispatch never
+saw a GQA shape, and it passed an all-zeros mask during single-token decode,
+which only served to keep SDPA off its fused kernels. These pin both contracts:
+K/V must reach the backend uncompressed, decode must carry no mask, and the
+compressed path must compute the same thing the expanded one did.
+"""
+
+import pytest
+import torch
+
+from vllm_omni.diffusion.attention.backends.sdpa import SDPAImpl
+from vllm_omni.diffusion.models.sensenova_u1.sensenova_u1_transformer import (
+    SenseNovaU1Attention,
+    SenseNovaU1Model,
+)
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
+
+N_HEADS, N_KV_HEADS, HEAD_DIM, SEQ = 32, 8, 8, 6
+GROUPS = N_HEADS // N_KV_HEADS
+
+
+class _Recorder:
+    """Stands in for the Attention module and records what it was handed."""
+
+    def __init__(self):
+        self.calls: list[tuple[torch.Size, torch.Size, torch.Tensor | None]] = []
+
+    def __call__(self, query, key, value, attn_metadata):
+        mask = None if attn_metadata is None else attn_metadata.attn_mask
+        self.calls.append((query.shape, key.shape, mask))
+        return query
+
+
+class _AttnHost:
+    """Carries only the attributes ``_run_attn``/``_run_attn_bshd`` touch."""
+
+    num_kv_groups = GROUPS
+    _align_mask_dtype = staticmethod(SenseNovaU1Attention._align_mask_dtype)
+    _run_attn = SenseNovaU1Attention._run_attn
+    _run_attn_bshd = SenseNovaU1Attention._run_attn_bshd
+
+    def __init__(self):
+        self.attn = _Recorder()
+
+
+def test_bhsd_path_hands_the_backend_compressed_kv():
+    host = _AttnHost()
+    q = torch.randn(1, N_HEADS, SEQ, HEAD_DIM)
+    k = torch.randn(1, N_KV_HEADS, SEQ, HEAD_DIM)
+    host._run_attn(q, k, k.clone(), None)
+    _, key_shape, _ = host.attn.calls[-1]
+    # [B, S, H, D] after the transpose, so heads sit on dim 2.
+    assert key_shape[2] == N_KV_HEADS, f"K reached the backend with {key_shape[2]} heads"
+
+
+def test_bshd_path_hands_the_backend_compressed_kv():
+    host = _AttnHost()
+    q = torch.randn(1, SEQ, N_HEADS, HEAD_DIM)
+    k = torch.randn(1, SEQ, N_KV_HEADS, HEAD_DIM)
+    host._run_attn_bshd(q, k, k.clone(), None)
+    _, key_shape, _ = host.attn.calls[-1]
+    assert key_shape[2] == N_KV_HEADS, f"K reached the backend with {key_shape[2]} heads"
+
+
+def test_compressed_kv_computes_what_expansion_computed():
+    """The backend must return the same tensor either way, or this is a bugfix
+    dressed up as a perf change."""
+    torch.manual_seed(0)
+    q = torch.randn(1, SEQ, N_HEADS, HEAD_DIM)
+    k = torch.randn(1, SEQ, N_KV_HEADS, HEAD_DIM)
+    v = torch.randn(1, SEQ, N_KV_HEADS, HEAD_DIM)
+    impl = SDPAImpl(
+        num_heads=N_HEADS,
+        head_size=HEAD_DIM,
+        softmax_scale=HEAD_DIM**-0.5,
+        num_kv_heads=N_KV_HEADS,
+    )
+    compressed = impl._forward_impl(q, k, v)
+    expanded = impl._forward_impl(q, k.repeat_interleave(GROUPS, dim=2), v.repeat_interleave(GROUPS, dim=2))
+    torch.testing.assert_close(compressed, expanded, atol=1e-6, rtol=1e-6)
+
+
+class _MaskProbe(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.masks: list[torch.Tensor | None] = []
+
+    def forward(self, hidden_states, **kwargs):
+        self.masks.append(kwargs["attention_mask"]["full_attention"])
+        return hidden_states
+
+
+def _fake_rope(hidden_states, position_ids):
+    """The forward builds the three RoPE tables before the layer loop; the probe
+    ignores them, so any 2-tuple will do."""
+    return hidden_states, hidden_states
+
+
+class _ModelHost:
+    """Carries only the attributes ``SenseNovaU1Model.forward`` touches."""
+
+    forward = SenseNovaU1Model.forward
+
+    def __init__(self, probe):
+        self.layers = [probe]
+        self.norm = torch.nn.Identity()
+        self.norm_mot_gen = torch.nn.Identity()
+        self.rotary_emb = _fake_rope
+        self.rotary_emb_hw = _fake_rope
+
+
+def _run_model(seq_len: int):
+    probe = _MaskProbe()
+    _ModelHost(probe).forward(
+        inputs_embeds=torch.zeros(1, seq_len, 4),
+        indexes=torch.zeros(3, seq_len, dtype=torch.long),
+    )
+    return probe.masks[-1]
+
+
+def test_single_token_decode_carries_no_mask():
+    assert _run_model(1) is None, "decode still builds the all-zeros mask"
+
+
+def test_prefill_still_gets_a_causal_mask():
+    mask = _run_model(3)
+    assert mask is not None and mask.shape == (1, 1, 3, 3)
+    assert mask[0, 0, 0, 0] == 0.0
+    assert torch.isinf(mask[0, 0, 0, 1]) and mask[0, 0, 0, 1] < 0
+    assert mask[0, 0, 2, 1] == 0.0
+
+
+# ---------------------------------------------------------------------------
+# The decode path changes which SDPA kernel runs, so it is not bit-identical to
+# what it replaced. "Different" would be a regression if the new result were
+# worse, so pin the direction against an independent float64 reference. This
+# needs a fused GQA kernel, which only exists on CUDA.
+# ---------------------------------------------------------------------------
+
+
+def _float64_attention(q, k, v):
+    q64 = q.double()
+    k64 = k.double().repeat_interleave(GROUPS, dim=1)
+    v64 = v.double().repeat_interleave(GROUPS, dim=1)
+    scores = (q64 @ k64.transpose(-1, -2)) * (q.shape[-1] ** -0.5)
+    return torch.softmax(scores, dim=-1) @ v64
+
+
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required for fused GQA")
+@pytest.mark.cuda
+@pytest.mark.L4
+@pytest.mark.parametrize("kv_len", [271, 512, 2048])
+def test_decode_without_the_no_op_mask_is_closer_to_float64(kv_len):
+    """The old decode path expanded K/V and passed an all-zeros mask, which kept
+    SDPA off its fused kernels. Dropping the mask is only defensible if the
+    kernel it unlocks is at least as accurate."""
+    dev, dtype = "cuda", torch.bfloat16
+    heads, dim = 32, 128
+    new_errs: list[float] = []
+    old_errs: list[float] = []
+    trials = 8
+    for t in range(trials):
+        torch.manual_seed(2000 + t)
+        q = torch.randn(1, heads, 1, dim, device=dev, dtype=dtype)
+        k = torch.randn(1, heads // GROUPS, kv_len, dim, device=dev, dtype=dtype)
+        v = torch.randn(1, heads // GROUPS, kv_len, dim, device=dev, dtype=dtype)
+        ref = _float64_attention(q, k, v)
+
+        zeros_mask = torch.zeros(1, 1, 1, kv_len, device=dev, dtype=dtype)
+        old = torch.nn.functional.scaled_dot_product_attention(
+            q,
+            k.repeat_interleave(GROUPS, dim=1),
+            v.repeat_interleave(GROUPS, dim=1),
+            attn_mask=zeros_mask,
+        )
+        new = torch.nn.functional.scaled_dot_product_attention(q, k, v, enable_gqa=True)
+
+        new_errs.append((new.double() - ref).abs().mean().item())
+        old_errs.append((old.double() - ref).abs().mean().item())
+
+    # A single trial can tie either way depending on the GPU and the CUDA build,
+    # so gate on the aggregate: the maskless path must not be worse on average.
+    new_mean = sum(new_errs) / trials
+    old_mean = sum(old_errs) / trials
+    wins = sum(n <= o for n, o in zip(new_errs, old_errs))
+    assert new_mean <= old_mean, (
+        f"kv_len={kv_len}: maskless mean error {new_mean:.4e} is worse than masked {old_mean:.4e} "
+        f"({wins}/{trials} per-trial wins)"
+    )

tests/diffusion/models/sensenova_u1/test_sensenova_u1_distill_lora.py

@@ -0,0 +1,196 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""CPU tests for SenseNova-U1 distilled-LoRA fusion.
+
+The checkpoints ship unfused ``q_proj_mot_gen`` / ``gate_proj`` keys while the
+pipeline runs fused layers, so each delta lands in its own row slice.
+"""
+
+import pytest
+import torch
+import torch.nn as nn
+from safetensors.torch import save_file
+
+from vllm_omni.diffusion.models.sensenova_u1.pipeline_sensenova_u1 import SenseNovaU1Pipeline
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
+
+HIDDEN, KV, INTER, RANK, ALPHA = 8, 4, 16, 2, 4.0
+
+
+class _Layer(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.self_attn = nn.Module()
+        self.self_attn.qkv_proj_mot_gen = nn.Linear(HIDDEN, HIDDEN + 2 * KV, bias=False)
+        self.self_attn.o_proj_mot_gen = nn.Linear(HIDDEN, HIDDEN, bias=False)
+        self.mlp_mot_gen = nn.Module()
+        self.mlp_mot_gen.gate_up_proj = nn.Linear(HIDDEN, 2 * INTER, bias=False)
+        self.mlp_mot_gen.down_proj = nn.Linear(INTER, HIDDEN, bias=False)
+
+
+class _Stub(SenseNovaU1Pipeline):
+    """Only the parameter tree matters here, so skip the real __init__."""
+
+    def __init__(self):
+        nn.Module.__init__(self)
+        self.language_model = nn.Module()
+        self.language_model.model = nn.Module()
+        self.language_model.model.layers = nn.ModuleList([_Layer()])
+
+
+def _triple(out: int, in_: int, fill: float) -> dict[str, torch.Tensor]:
+    return {
+        "lora_down.weight": torch.full((RANK, in_), fill, dtype=torch.float32),
+        "lora_up.weight": torch.full((out, RANK), 1.0, dtype=torch.float32),
+        "alpha": torch.tensor(ALPHA),
+    }
+
+
+def _write(tmp_path, entries: dict[str, dict[str, torch.Tensor]]):
+    flat = {f"{mod}.{leaf}": t for mod, parts in entries.items() for leaf, t in parts.items()}
+    path = tmp_path / "lora.safetensors"
+    save_file(flat, str(path))
+    return str(path)
+
+
+def _prefix(i: int = 0) -> str:
+    return f"language_model.model.layers.{i}"
+
+
+def test_qkv_shards_land_in_their_own_row_slices(tmp_path):
+    pipe = _Stub()
+    qkv = pipe.language_model.model.layers[0].self_attn.qkv_proj_mot_gen.weight
+    with torch.no_grad():
+        qkv.zero_()
+    entries = {
+        f"{_prefix()}.self_attn.q_proj_mot_gen": _triple(HIDDEN, HIDDEN, 1.0),
+        f"{_prefix()}.self_attn.k_proj_mot_gen": _triple(KV, HIDDEN, 2.0),
+        f"{_prefix()}.self_attn.v_proj_mot_gen": _triple(KV, HIDDEN, 3.0),
+    }
+    pipe.load_lora_weights(_write(tmp_path, entries))
+
+    scale = ALPHA / RANK
+    assert torch.allclose(qkv[:HIDDEN], torch.full((HIDDEN, HIDDEN), scale * RANK * 1.0))
+    assert torch.allclose(qkv[HIDDEN : HIDDEN + KV], torch.full((KV, HIDDEN), scale * RANK * 2.0))
+    assert torch.allclose(qkv[HIDDEN + KV :], torch.full((KV, HIDDEN), scale * RANK * 3.0))
+
+
+def test_gate_and_up_land_in_their_own_halves(tmp_path):
+    pipe = _Stub()
+    gu = pipe.language_model.model.layers[0].mlp_mot_gen.gate_up_proj.weight
+    with torch.no_grad():
+        gu.zero_()
+    entries = {
+        f"{_prefix()}.mlp_mot_gen.gate_proj": _triple(INTER, HIDDEN, 1.0),
+        f"{_prefix()}.mlp_mot_gen.up_proj": _triple(INTER, HIDDEN, 5.0),
+    }
+    pipe.load_lora_weights(_write(tmp_path, entries))
+
+    scale = ALPHA / RANK
+    assert torch.allclose(gu[:INTER], torch.full((INTER, HIDDEN), scale * RANK * 1.0))
+    assert torch.allclose(gu[INTER:], torch.full((INTER, HIDDEN), scale * RANK * 5.0))
+
+
+def test_unfused_targets_are_added_whole(tmp_path):
+    pipe = _Stub()
+    o = pipe.language_model.model.layers[0].self_attn.o_proj_mot_gen.weight
+    before = o.detach().clone()
+    entries = {f"{_prefix()}.self_attn.o_proj_mot_gen": _triple(HIDDEN, HIDDEN, 1.0)}
+    pipe.load_lora_weights(_write(tmp_path, entries))
+    assert torch.allclose(o - before, torch.full((HIDDEN, HIDDEN), (ALPHA / RANK) * RANK))
+
+
+def test_lora_that_matches_nothing_raises_instead_of_no_op(tmp_path):
+    """Without a loader the worker only warned and generated without the
+    adapter, so a wrong path produced a plausible image and no error."""
+    pipe = _Stub()
+    entries = {f"{_prefix()}.self_attn.does_not_exist": _triple(HIDDEN, HIDDEN, 1.0)}
+    with pytest.raises(ValueError, match="matched no parameter"):
+        pipe.load_lora_weights(_write(tmp_path, entries))
+
+
+def test_shape_mismatch_raises(tmp_path):
+    pipe = _Stub()
+    entries = {f"{_prefix()}.self_attn.o_proj_mot_gen": _triple(HIDDEN + 1, HIDDEN, 1.0)}
+    with pytest.raises(RuntimeError, match="must match the size"):
+        pipe.load_lora_weights(_write(tmp_path, entries))
+
+
+def test_multiple_files_rejected(tmp_path):
+    pipe = _Stub()
+    path = _write(tmp_path, {f"{_prefix()}.self_attn.o_proj_mot_gen": _triple(HIDDEN, HIDDEN, 1.0)})
+    with pytest.raises(ValueError, match="exactly one"):
+        pipe.load_lora_weights([path, path])
+
+
+class _ShardedLayer(nn.Module):
+    """One rank of a two-way sharded fused QKV, with a vLLM-style weight loader.
+
+    The parameter holds half of each of q, k and v. The loader is handed the whole
+    layer's tensor and keeps only this rank's slice, which is what vLLM does.
+    """
+
+    def __init__(self):
+        super().__init__()
+        self.weight = nn.Parameter(torch.zeros(HIDDEN // 2 + KV, HIDDEN))
+        self.seen: list[tuple[str, tuple[int, ...]]] = []
+        self.weight.weight_loader = self._weight_loader
+
+    def _weight_loader(self, param, loaded_weight, shard_id):
+        self.seen.append((shard_id, tuple(loaded_weight.shape)))
+        sizes = {"q": HIDDEN // 2, "k": KV // 2, "v": KV // 2}
+        offsets = {"q": 0, "k": HIDDEN // 2, "v": HIDDEN // 2 + KV // 2}
+        rows = sizes[shard_id]
+        param.data[offsets[shard_id] : offsets[shard_id] + rows] = loaded_weight[:rows]
+
+
+class _ShardedStub(SenseNovaU1Pipeline):
+    def __init__(self):
+        nn.Module.__init__(self)
+        self.language_model = nn.Module()
+        self.language_model.model = nn.Module()
+        layer = nn.Module()
+        layer.self_attn = nn.Module()
+        layer.self_attn.qkv_proj_mot_gen = _ShardedLayer()
+        self.language_model.model.layers = nn.ModuleList([layer])
+
+
+def test_fused_delta_is_sharded_through_the_weight_loader(tmp_path):
+    """Under tensor parallelism the parameter is a rank-local shard while the
+    checkpoint delta spans the whole layer, so the delta has to go through the
+    layer's weight loader. Adding it directly raises a shape error."""
+    pipe = _ShardedStub()
+    layer = pipe.language_model.model.layers[0].self_attn.qkv_proj_mot_gen
+    entries = {
+        f"{_prefix()}.self_attn.q_proj_mot_gen": _triple(HIDDEN, HIDDEN, 1.0),
+        f"{_prefix()}.self_attn.k_proj_mot_gen": _triple(KV, HIDDEN, 2.0),
+        f"{_prefix()}.self_attn.v_proj_mot_gen": _triple(KV, HIDDEN, 3.0),
+    }
+    pipe.load_lora_weights(_write(tmp_path, entries))
+
+    # the loader saw the whole layer's tensors, not a pre-sharded slice
+    assert layer.seen == [("q", (HIDDEN, HIDDEN)), ("k", (KV, HIDDEN)), ("v", (KV, HIDDEN))]
+
+    scale = ALPHA / RANK * RANK
+    got = layer.weight
+    assert torch.allclose(got[: HIDDEN // 2], torch.full((HIDDEN // 2, HIDDEN), scale * 1.0))
+    assert torch.allclose(got[HIDDEN // 2 : HIDDEN // 2 + KV // 2], torch.full((KV // 2, HIDDEN), scale * 2.0))
+    assert torch.allclose(got[HIDDEN // 2 + KV // 2 :], torch.full((KV // 2, HIDDEN), scale * 3.0))
+
+
+def test_fusion_does_not_retain_the_adapter_tensors(tmp_path):
+    """Fusion is one-way and this pipeline has no unload path, so holding the
+    fp32 state dict would keep about 1.5 GiB alive for the shipped checkpoint."""
+    pipe = _Stub()
+    path = _write(tmp_path, {f"{_prefix()}.self_attn.o_proj_mot_gen": _triple(HIDDEN, HIDDEN, 1.0)})
+    pipe.load_lora_weights(path, adapter_name="distill")
+
+    kept = pipe.lora_loaded["distill"]
+    assert not isinstance(kept, dict)
+    assert not isinstance(kept, torch.Tensor)
+
+    # the sentinel still makes a second load a no-op
+    before = pipe.language_model.model.layers[0].self_attn.o_proj_mot_gen.weight.detach().clone()
+    pipe.load_lora_weights(path, adapter_name="distill")
+    torch.testing.assert_close(pipe.language_model.model.layers[0].self_attn.o_proj_mot_gen.weight, before)

tests/diffusion/models/sensenova_u1/test_sensenova_u1_paged_decode.py

@@ -0,0 +1,800 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""Tests for the paged decode cache and its CUDA graph.
+
+The cache exists so a decode step has static shapes and can be captured. That
+puts the burden on two things a normal test would not look at: the prefix must
+survive the hand-off from the ordinary cache, and every quantity that varies
+between steps must live in a device tensor. A Python-level index is correct
+eagerly and silently wrong once captured, because capture bakes it in -- which
+is exactly what the last test here pins.
+"""
+
+from types import SimpleNamespace
+
+import pytest
+import torch
+from vllm.lora.layers import BaseLayerWithLoRA
+
+import vllm_omni.diffusion.models.sensenova_u1.paged_decode as paged_decode
+from vllm_omni.diffusion.models.sensenova_u1.paged_decode import (
+    BLOCK_SIZE,
+    BUCKETS,
+    TAIL_STEP,
+    PagedDecodeCache,
+    _bucket_for,
+)
+from vllm_omni.diffusion.models.sensenova_u1.sensenova_u1_transformer import (
+    SenseNovaU1Attention,
+)
+from vllm_omni.platforms import current_omni_platform
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
+
+LAYERS, KV_HEADS, HEAD_DIM = 2, 4, 8
+N_HEADS = KV_HEADS * 2
+
+
+class _Layer:
+    def __init__(self, keys=None, values=None):
+        self.keys, self.values = keys, values
+
+
+class _Cache:
+    """Stands in for the transformers cache: layers with [B, H, S, D] tensors."""
+
+    def __init__(self, layers):
+        self.layers = layers
+
+
+def _dyn_cache(prefix, device=None):
+    torch.manual_seed(0)
+    return _Cache(
+        [
+            _Layer(
+                torch.randn(1, KV_HEADS, prefix, HEAD_DIM, device=device),
+                torch.randn(1, KV_HEADS, prefix, HEAD_DIM, device=device),
+            )
+            for _ in range(LAYERS)
+        ]
+    )
+
+
+def test_bucket_rounds_up_and_is_block_aligned():
+    assert _bucket_for(1) == 512
+    assert _bucket_for(512) == 512
+    assert _bucket_for(513) == 1024
+    assert _bucket_for(100_000) % BLOCK_SIZE == 0
+
+
+def test_past_the_last_bucket_the_schedule_grows_in_steps():
+    """Every bucket change reallocates the cache and re-captures the graph, so
+    the tail of the schedule must not step by BLOCK_SIZE. A think edit with two
+    2048x2048 inputs runs to about 9,100 tokens, so this regime is reachable,
+    and a request decodes at most 1024 steps.
+    """
+    last = BUCKETS[-1]
+    assert _bucket_for(last + 1) == last + TAIL_STEP
+    assert _bucket_for(100_000) % BLOCK_SIZE == 0
+    for start in (last + 1, last + TAIL_STEP - 1, last + TAIL_STEP + 7):
+        bucket, grows = _bucket_for(start), 0
+        for n in range(start, start + 1025):
+            if _bucket_for(n) != bucket:
+                bucket = _bucket_for(n)
+                grows += 1
+        assert grows <= 1, f"start {start}: {grows} reallocations over 1024 decode steps"
+
+
+def test_prefix_survives_the_hand_off():
+    dyn = _dyn_cache(37)
+    paged = PagedDecodeCache.from_dynamic_cache(dyn, LAYERS, torch.device("cpu"), torch.float32)
+    assert paged.length == 37
+    for i in range(LAYERS):
+        stored = paged.k[i].view(-1, KV_HEADS, HEAD_DIM)[:37]
+        torch.testing.assert_close(stored, dyn.layers[i].keys[0].transpose(0, 1))
+
+
+def test_write_back_restores_what_was_handed_over():
+    """Decode is paged but the stage after it reads the ordinary cache."""
+    dyn = _dyn_cache(21)
+    original = [layer.keys.clone() for layer in dyn.layers]
+    paged = PagedDecodeCache.from_dynamic_cache(dyn, LAYERS, torch.device("cpu"), torch.float32)
+    paged.to_dynamic_cache(dyn)
+    for before, layer in zip(original, dyn.layers):
+        torch.testing.assert_close(layer.keys, before)
+
+
+def test_growing_keeps_the_tokens_already_stored():
+    dyn = _dyn_cache(500)
+    paged = PagedDecodeCache.from_dynamic_cache(dyn, LAYERS, torch.device("cpu"), torch.float32)
+    assert paged.bucket == 512
+    before = [k.view(-1, KV_HEADS, HEAD_DIM)[:500].clone() for k in paged.k]
+    gen = paged.generation
+    assert paged.grow(513) is True
+    assert paged.bucket == 1024
+    assert paged.generation > gen, "a reallocation must invalidate captured graphs"
+    for old, new in zip(before, paged.k):
+        torch.testing.assert_close(new.view(-1, KV_HEADS, HEAD_DIM)[:500], old)
+
+
+def test_a_failed_grow_leaves_the_cache_on_its_old_buffers(monkeypatch):
+    """A half-applied resize would strand a capture on freed storage.
+
+    ``generation`` is half of the key a captured graph is stored under, so
+    publishing new K without publishing V, the block table, the bucket and the
+    generation leaves the old key selecting a graph whose recorded addresses
+    have just been freed. A later request replays it. Injecting a failure into
+    the last allocation is the cheapest way to reach that window.
+    """
+    dyn = _dyn_cache(500)
+    paged = PagedDecodeCache.from_dynamic_cache(dyn, LAYERS, torch.device("cpu"), torch.float32)
+    before_k, before_v = list(paged.k), list(paged.v)
+    before_table, before_bucket = paged.block_table, paged.bucket
+    before_generation, before_length = paged.generation, paged.length
+
+    real_zeros = torch.zeros
+    calls = []
+
+    def flaky(*args, **kwargs):
+        calls.append(1)
+        if len(calls) == 2 * LAYERS:  # the last of the two allocations per layer
+            raise RuntimeError("injected allocation failure")
+        return real_zeros(*args, **kwargs)
+
+    monkeypatch.setattr(torch, "zeros", flaky)
+    with pytest.raises(RuntimeError, match="injected"):
+        paged.grow(513)
+    monkeypatch.undo()
+
+    assert [id(t) for t in paged.k] == [id(t) for t in before_k], "K was published on its own"
+    assert [id(t) for t in paged.v] == [id(t) for t in before_v], "V was published on its own"
+    assert paged.block_table is before_table
+    assert paged.bucket == before_bucket
+    assert paged.generation == before_generation, "the graph key moved without new buffers"
+    assert paged.length == before_length
+
+
+def test_length_lives_in_device_tensors():
+    """Both the attended length and the write slot must be tensors.
+
+    A captured graph reads them at replay; a Python int would be frozen at
+    capture time and every replay would attend the wrong span and overwrite the
+    same slot.
+    """
+    dyn = _dyn_cache(10)
+    paged = PagedDecodeCache.from_dynamic_cache(dyn, LAYERS, torch.device("cpu"), torch.float32)
+    paged.set_length(15)
+    assert torch.is_tensor(paged.seqused) and int(paged.seqused) == 15
+    assert torch.is_tensor(paged.pos) and int(paged.pos) == 14
+
+
+# ---------------------------------------------------------------------------
+# The capture contract. These need CUDA and the bundled flash-attention kernel.
+# ---------------------------------------------------------------------------
+
+cuda_only = pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required")
+
+
+@cuda_only
+@pytest.mark.cuda
+@pytest.mark.L4
+def test_paged_attention_matches_sdpa_over_the_same_buffer():
+    import torch.nn.functional as F
+
+    from vllm_omni.diffusion.models.sensenova_u1.paged_decode import PagedDecodeCache
+
+    heads, kv_heads, dim, n = 8, 2, 64, 100
+    dev, dt = torch.device("cuda"), torch.bfloat16
+    torch.manual_seed(0)
+    dyn = _Cache(
+        [
+            _Layer(
+                torch.randn(1, kv_heads, n, dim, device=dev, dtype=dt),
+                torch.randn(1, kv_heads, n, dim, device=dev, dtype=dt),
+            )
+        ]
+    )
+    paged = PagedDecodeCache.from_dynamic_cache(dyn, 1, dev, dt)
+    paged.set_length(n)
+    q = torch.randn(1, heads, 1, dim, device=dev, dtype=dt)
+    k = torch.zeros(1, kv_heads, 1, dim, device=dev, dtype=dt)
+    scale = dim**-0.5
+    # write the last slot back as itself so attend() is a pure read
+    k[0, :, 0] = dyn.layers[0].keys[0, :, n - 1]
+    v = torch.zeros_like(k)
+    v[0, :, 0] = dyn.layers[0].values[0, :, n - 1]
+    got = paged.attend(0, q, k, v, scale)
+    flat_k = paged.k[0].view(-1, kv_heads, dim)[:n]
+    flat_v = paged.v[0].view(-1, kv_heads, dim)[:n]
+    want = F.scaled_dot_product_attention(
+        q,
+        flat_k.unsqueeze(0).transpose(1, 2).contiguous(),
+        flat_v.unsqueeze(0).transpose(1, 2).contiguous(),
+        enable_gqa=True,
+        scale=scale,
+    ).transpose(1, 2)
+    torch.testing.assert_close(got.float(), want.float(), atol=2e-3, rtol=2e-3)
+
+
+@cuda_only
+@pytest.mark.cuda
+@pytest.mark.L4
+def test_a_captured_graph_follows_a_later_length():
+    """One capture has to serve every length in the bucket.
+
+    If the attended span were a Python value the graph would keep attending the
+    capture-time length, and the run would still look fast.
+    """
+    import torch.nn.functional as F
+
+    from vllm_omni.diffusion.models.sensenova_u1.paged_decode import PagedDecodeCache
+
+    heads, kv_heads, dim, n = 8, 2, 64, 200
+    dev, dt = torch.device("cuda"), torch.bfloat16
+    torch.manual_seed(0)
+    dyn = _Cache(
+        [
+            _Layer(
+                torch.randn(1, kv_heads, n, dim, device=dev, dtype=dt),
+                torch.randn(1, kv_heads, n, dim, device=dev, dtype=dt),
+            )
+        ]
+    )
+    paged = PagedDecodeCache.from_dynamic_cache(dyn, 1, dev, dt)
+    q = torch.randn(1, heads, 1, dim, device=dev, dtype=dt)
+    k = torch.zeros(1, kv_heads, 1, dim, device=dev, dtype=dt)
+    v = torch.zeros_like(k)
+    scale = dim**-0.5
+
+    paged.set_length(n)
+    k[0, :, 0] = dyn.layers[0].keys[0, :, n - 1]
+    v[0, :, 0] = dyn.layers[0].values[0, :, n - 1]
+
+    side = torch.cuda.Stream()
+    side.wait_stream(torch.cuda.current_stream())
+    with torch.cuda.stream(side):
+        for _ in range(3):
+            paged.attend(0, q, k, v, scale)
+    torch.cuda.current_stream().wait_stream(side)
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        out = paged.attend(0, q, k, v, scale)
+
+    shorter = n - 43
+    paged.set_length(shorter)
+    graph.replay()
+    torch.accelerator.synchronize()
+
+    flat_k = paged.k[0].view(-1, kv_heads, dim)[:shorter]
+    flat_v = paged.v[0].view(-1, kv_heads, dim)[:shorter]
+    want = F.scaled_dot_product_attention(
+        q,
+        flat_k.unsqueeze(0).transpose(1, 2).contiguous(),
+        flat_v.unsqueeze(0).transpose(1, 2).contiguous(),
+        enable_gqa=True,
+        scale=scale,
+    ).transpose(1, 2)
+    torch.testing.assert_close(out.float(), want.float(), atol=2e-3, rtol=2e-3)
+
+
+@cuda_only
+@pytest.mark.cuda
+@pytest.mark.L4
+def test_a_captured_graph_writes_each_step_to_its_own_slot():
+    """The write slot has to be a device tensor, not a Python index.
+
+    ``test_a_captured_graph_follows_a_later_length`` pins the attended span.
+    The slot written is the other half: a Python index is frozen at capture, so
+    every replay would overwrite the capture-time slot and the cache would stop
+    growing -- correct eagerly, silently wrong once captured.
+    """
+    from vllm_omni.diffusion.models.sensenova_u1.paged_decode import PagedDecodeCache
+
+    heads, kv_heads, dim, n = 8, 2, 64, 100
+    dev, dt = torch.device("cuda"), torch.bfloat16
+    torch.manual_seed(0)
+    dyn = _Cache(
+        [
+            _Layer(
+                torch.randn(1, kv_heads, n, dim, device=dev, dtype=dt),
+                torch.randn(1, kv_heads, n, dim, device=dev, dtype=dt),
+            )
+        ]
+    )
+    paged = PagedDecodeCache.from_dynamic_cache(dyn, 1, dev, dt)
+    q = torch.randn(1, heads, 1, dim, device=dev, dtype=dt)
+    k = torch.zeros(1, kv_heads, 1, dim, device=dev, dtype=dt)
+    v = torch.zeros_like(k)
+    scale = dim**-0.5
+
+    # Capture on the step that writes slot n-1, restoring what is already there
+    # so the capture itself changes nothing.
+    paged.set_length(n)
+    k[0, :, 0] = dyn.layers[0].keys[0, :, n - 1]
+    v[0, :, 0] = dyn.layers[0].values[0, :, n - 1]
+    side = torch.cuda.Stream()
+    side.wait_stream(torch.cuda.current_stream())
+    with torch.cuda.stream(side):
+        for _ in range(3):
+            paged.attend(0, q, k, v, scale)
+    torch.cuda.current_stream().wait_stream(side)
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        paged.attend(0, q, k, v, scale)
+
+    kept = paged.k[0].view(-1, kv_heads, dim)[n - 1].clone()
+
+    # The next decode step: one token further, different K/V.
+    paged.set_length(n + 1)
+    k[0, :, 0] = 1.5
+    v[0, :, 0] = -2.5
+    graph.replay()
+    current_omni_platform.synchronize()
+
+    flat_k = paged.k[0].view(-1, kv_heads, dim)
+    flat_v = paged.v[0].view(-1, kv_heads, dim)
+    torch.testing.assert_close(flat_k[n].float(), torch.full_like(flat_k[n], 1.5).float())
+    torch.testing.assert_close(flat_v[n].float(), torch.full_like(flat_v[n], -2.5).float())
+    torch.testing.assert_close(flat_k[n - 1].float(), kept.float())
+
+
+@cuda_only
+@pytest.mark.cuda
+@pytest.mark.L4
+def test_a_whole_decode_loop_matches_the_unpaged_path():
+    """Per-step kernel equivalence is not loop equivalence.
+
+    With the switch off a single-token step attends over a cache grown by
+    ``torch.cat`` and no mask; with it on the same step reads a bucket-sized
+    buffer through ``seqused_k``. Run the same sequence of steps through both,
+    crossing a bucket boundary on the way, and require the same argmax at every
+    step off a fixed projection -- which is the token id the pipeline would emit.
+    """
+    import torch.nn.functional as F
+
+    from vllm_omni.diffusion.models.sensenova_u1.paged_decode import PagedDecodeCache
+
+    heads, kv_heads, dim = 8, 2, 64
+    prefill, steps = 500, 40  # crosses the 512 bucket at step 13
+    dev, dt = torch.device("cuda"), torch.bfloat16
+    scale = dim**-0.5
+    torch.manual_seed(7)
+
+    keys = torch.randn(1, kv_heads, prefill, dim, device=dev, dtype=dt)
+    values = torch.randn(1, kv_heads, prefill, dim, device=dev, dtype=dt)
+    paged = PagedDecodeCache.from_dynamic_cache(_Cache([_Layer(keys, values)]), 1, dev, dt)
+    ref_k, ref_v = keys.clone(), values.clone()
+    proj = torch.randn(dim * heads, 128, device=dev, dtype=torch.float32)
+
+    paged_ids, ref_ids, worst = [], [], 0.0
+    for step in range(steps):
+        q = torch.randn(1, heads, 1, dim, device=dev, dtype=dt)
+        k = torch.randn(1, kv_heads, 1, dim, device=dev, dtype=dt)
+        v = torch.randn(1, kv_heads, 1, dim, device=dev, dtype=dt)
+
+        if paged.length + 1 > paged.bucket:
+            paged.grow(paged.length + 1)
+        paged.set_length(paged.length + 1)
+        got = paged.attend(0, q, k, v, scale)
+
+        ref_k = torch.cat([ref_k, k], dim=2)
+        ref_v = torch.cat([ref_v, v], dim=2)
+        want = F.scaled_dot_product_attention(q, ref_k, ref_v, enable_gqa=True, scale=scale).transpose(1, 2)
+
+        worst = max(worst, (got.float() - want.float()).abs().max().item())
+        paged_ids.append(int((got.float().reshape(1, -1) @ proj).argmax()))
+        ref_ids.append(int((want.float().reshape(1, -1) @ proj).argmax()))
+
+    assert paged.bucket == 1024, "the loop never crossed a bucket boundary"
+    assert paged_ids == ref_ids, (
+        f"argmax diverges at step {next(i for i, (a, b) in enumerate(zip(paged_ids, ref_ids)) if a != b)}"
+    )
+    assert worst < 5e-2, f"largest per-step deviation {worst:.3e}"
+
+
+class _StubLM(torch.nn.Module):
+    """Just enough CUDA work inside the capture for a pool to be allocated."""
+
+    def __init__(self, device):
+        super().__init__()
+        self.w = torch.randn(16, 64, device=device)
+
+    def forward(self, input_ids, indexes, past_key_values=None, use_cache=False, paged_cache=None):
+        x = input_ids.float().expand(1, 16) @ self.w
+        return SimpleNamespace(logits=x.unsqueeze(0))
+
+
+@cuda_only
+@pytest.mark.cuda
+@pytest.mark.L4
+def test_the_decode_graph_captures_into_the_shared_platform_pool(monkeypatch):
+    """``_decode_context`` builds a runner per request, so a private capture pool
+    hands every request its own arena and never gives it back. Measured on a
+    real model that was about 40 MiB of device memory per request, rising
+    linearly over twelve sequential think requests while the same twelve with
+    the switch off held flat.
+    """
+    from vllm.platforms import current_platform
+
+    from vllm_omni.diffusion.models.sensenova_u1.paged_decode import DecodeGraphRunner
+
+    pools = []
+    real_graph = torch.cuda.graph
+
+    def spy(graph, pool=None, **kwargs):
+        pools.append(pool)
+        return real_graph(graph, pool=pool, **kwargs)
+
+    monkeypatch.setattr(torch.cuda, "graph", spy)
+
+    dev = torch.device("cuda")
+    cache = PagedDecodeCache.from_dynamic_cache(_dyn_cache(8, dev), LAYERS, dev, torch.float32)
+    runner = DecodeGraphRunner(_StubLM(dev), cache, dev)
+    runner.step(0, 0)
+
+    assert pools, "no graph was captured"
+    assert pools[0] is not None, "captured into a fresh private pool"
+    assert pools[0] == current_platform.get_global_graph_pool()
+    assert not hasattr(runner, "_pool"), "the runner still keeps a pool of its own"
+
+
+# ---------------------------------------------------------------------------
+# The cache is only worth anything if production actually reaches it, and the
+# kwargs it needs are not in every wheel that exports the kernel. The tests
+# above would all still pass with the `forward_und` branch deleted, so these
+# drive the real dispatch instead of the cache object.
+# ---------------------------------------------------------------------------
+
+
+class _PagedProbe:
+    def __init__(self, out):
+        self.out = out
+        self.calls: list[tuple[int, torch.Size, float]] = []
+
+    def attend(self, layer_idx, query, key, value, scaling):
+        self.calls.append((layer_idx, query.shape, scaling))
+        return self.out
+
+
+class _AttnHost:
+    """Carries only what ``forward_und`` touches on the way to the branch."""
+
+    forward_und = SenseNovaU1Attention.forward_und
+    qkv_proj = q_norm = k_norm = q_norm_hw = k_norm_hw = None
+    layer_idx = 3
+    scaling = HEAD_DIM**-0.5
+
+    def __init__(self, seq, paged_out):
+        self._qkv = (
+            torch.randn(1, N_HEADS, seq, HEAD_DIM),
+            torch.randn(1, KV_HEADS, seq, HEAD_DIM),
+            torch.randn(1, KV_HEADS, seq, HEAD_DIM),
+        )
+        self.paged = _PagedProbe(paged_out)
+        self.sdpa_calls: list[torch.Tensor | None] = []
+
+    def _project_and_rope(self, *args, **kwargs):
+        return self._qkv
+
+    def _run_attn(self, query, key, value, mask):
+        self.sdpa_calls.append(mask)
+        return torch.zeros(1, N_HEADS, query.shape[2], HEAD_DIM)
+
+    def o_proj(self, hidden_states):
+        return hidden_states * 2, None
+
+
+def test_forward_und_sends_a_single_token_decode_to_the_paged_cache():
+    """Goes through the real dispatch, so deleting the branch turns this red."""
+    out = torch.randn(1, N_HEADS, 1, HEAD_DIM)
+    host = _AttnHost(seq=1, paged_out=out)
+    got = host.forward_und(torch.zeros(1, 1, 16), None, None, position_embeddings=None, paged_cache=host.paged)
+    assert len(host.paged.calls) == 1, "the paged hook was never reached"
+    assert host.sdpa_calls == [], "the same step also ran the SDPA path"
+    assert host.paged.calls[0][0] == host.layer_idx
+    torch.testing.assert_close(got, out.reshape(1, 1, -1).contiguous() * 2)
+
+
+@pytest.mark.parametrize(
+    "seq,mask",
+    [(2, None), (1, torch.zeros(1, 1, 1, 1))],
+    ids=["multi_token", "masked"],
+)
+def test_only_an_unmasked_single_token_step_takes_the_paged_path(seq, mask):
+    """Both other shapes are uncapturable, and the cache holds one row per step."""
+    host = _AttnHost(seq=seq, paged_out=torch.zeros(1, N_HEADS, seq, HEAD_DIM))
+    host.forward_und(torch.zeros(1, seq, 16), None, mask, position_embeddings=None, paged_cache=host.paged)
+    assert host.paged.calls == [], "a step that cannot be captured reached the paged cache"
+    assert len(host.sdpa_calls) == 1
+
+
+def test_a_kernel_missing_any_kwarg_we_pass_is_not_supported(monkeypatch):
+    """The call names every argument, so the probe has to cover every name.
+
+    An older wheel exports ``flash_attn_varlen_func`` without the paged kwargs,
+    and a future one may rename the standard ones. Either way the probe is what
+    keeps that install on SDPA instead of raising TypeError at the first decode
+    step, with the feature on by default.
+    """
+
+    def no_paged_kwargs(q, k, v, max_seqlen_q, cu_seqlens_q, max_seqlen_k, cu_seqlens_k, causal=False):
+        raise AssertionError("must not be called")  # pragma: no cover
+
+    def renamed_standard_args(q, k, v, max_q, cu_q, max_k, seqused_k=None, block_table=None, causal=False):
+        raise AssertionError("must not be called")  # pragma: no cover
+
+    for kernel in (no_paged_kwargs, renamed_standard_args):
+        monkeypatch.setattr(paged_decode, "_flash_varlen", lambda kernel=kernel: kernel)
+        assert paged_decode.paged_decode_supported(torch.device("cuda"), HEAD_DIM) is False
+
+    def current_kernel(
+        q,
+        k,
+        v,
+        max_seqlen_q,
+        cu_seqlens_q,
+        max_seqlen_k,
+        seqused_k=None,
+        block_table=None,
+        softmax_scale=None,
+        causal=False,
+    ):
+        raise AssertionError("must not be called")  # pragma: no cover
+
+    monkeypatch.setattr(paged_decode, "_flash_varlen", lambda: current_kernel)
+    assert paged_decode.paged_decode_supported(torch.device("cuda"), HEAD_DIM) is True
+
+
+def test_only_q_k_and_v_reach_the_kernel_positionally(monkeypatch):
+    """A positional standard argument binds to whatever a future wheel puts in
+    that slot, and the probe above only checks that the names exist."""
+    seen = {}
+
+    def recorder(*args, **kwargs):
+        seen["args"], seen["kwargs"] = args, kwargs
+        return torch.zeros(1, N_HEADS, HEAD_DIM)
+
+    monkeypatch.setattr(paged_decode, "_flash_varlen", lambda: recorder)
+    cache = PagedDecodeCache(1, KV_HEADS, HEAD_DIM, 8, torch.device("cpu"), torch.float32)
+    cache.set_length(4)
+    cache.attend(
+        0,
+        torch.zeros(1, N_HEADS, 1, HEAD_DIM),
+        torch.zeros(1, KV_HEADS, 1, HEAD_DIM),
+        torch.zeros(1, KV_HEADS, 1, HEAD_DIM),
+        1.0,
+    )
+    assert len(seen["args"]) == 3, "only q, k and v may be positional"
+    assert set(seen["kwargs"]) == {
+        "max_seqlen_q",
+        "cu_seqlens_q",
+        "max_seqlen_k",
+        "seqused_k",
+        "block_table",
+        "softmax_scale",
+        "causal",
+    }
+
+
+def test_the_kill_switch_returns_no_cache(monkeypatch):
+    """The recipe documents `VLLM_OMNI_SENSENOVA_PAGED_DECODE=0` as the way back
+    to the ordinary cache, so the switch needs a test rather than only prose."""
+    from vllm_omni.diffusion.models.sensenova_u1.pipeline_sensenova_u1 import (
+        SenseNovaU1Pipeline,
+    )
+
+    monkeypatch.setenv("VLLM_OMNI_SENSENOVA_PAGED_DECODE", "0")
+    # The env check returns before anything else on the pipeline is touched.
+    host = object.__new__(SenseNovaU1Pipeline)
+    assert SenseNovaU1Pipeline._decode_context(host, object()) is None
+
+
+def _pipeline_host(monkeypatch, device="cpu", language_model=None):
+    """A bare pipeline carrying only what ``_decode_context`` touches."""
+    from vllm_omni.diffusion.models.sensenova_u1 import pipeline_sensenova_u1 as pipe_mod
+
+    monkeypatch.setenv("VLLM_OMNI_SENSENOVA_PAGED_DECODE", "1")
+    monkeypatch.setattr(pipe_mod, "paged_decode_supported", lambda dev, head_dim: True)
+    layer = SimpleNamespace(self_attn=SimpleNamespace(head_dim=HEAD_DIM))
+    host = object.__new__(pipe_mod.SenseNovaU1Pipeline)
+    # The pipeline is an nn.Module, and the language model assigned below is one
+    # too, so the bookkeeping has to exist before the assignment.
+    torch.nn.Module.__init__(host)
+    host.device = device
+    lm = language_model if language_model is not None else torch.nn.Module()
+    lm.model = SimpleNamespace(layers=[layer] * LAYERS)
+    host.language_model = lm
+    return pipe_mod, host
+
+
+def test_the_decode_context_is_reused_across_requests(monkeypatch):
+    """A capture binds the addresses it recorded, so a cache built per request
+    forces a capture per request. Two requests whose prefill fits the same
+    bucket have to come back with the same cache and the same runner, and with
+    the generation unchanged so the captures stay valid.
+    """
+    pipe_mod, host = _pipeline_host(monkeypatch)
+    monkeypatch.setattr(pipe_mod, "DecodeGraphRunner", lambda lm, cache, dev: SimpleNamespace(cache=cache))
+    ctx = pipe_mod.SenseNovaU1Pipeline._decode_context
+
+    first = ctx(host, _dyn_cache(10))
+    assert first is not None
+    second_input = _dyn_cache(20)
+    second = ctx(host, second_input)
+    assert second[0] is first[0], "the cache was rebuilt for a request that fits it"
+    assert second[1] is first[1], "the runner was rebuilt, so its captures are gone"
+    assert second[0].generation == first[0].generation, "reuse must not invalidate a capture"
+    assert second[0].length == 20
+
+    # The reused buffers must carry the new request's prefill, not the old one.
+    flat = second[0].k[0].view(-1, KV_HEADS, HEAD_DIM)[:20]
+    torch.testing.assert_close(flat, second_input.layers[0].keys[0].transpose(0, 1))
+
+    # A prefill past the bucket cannot share the buffers.
+    bigger = ctx(host, _dyn_cache(BUCKETS[0] + 1))
+    assert bigger[0] is not first[0], "a prefill past the bucket must get new buffers"
+
+
+def test_releasing_the_captures_forces_a_rebuild(monkeypatch):
+    """Sleep level 2 discards the memory a capture recorded, so what the
+    pipeline reuses across requests has to be droppable."""
+    pipe_mod, host = _pipeline_host(monkeypatch)
+    monkeypatch.setattr(pipe_mod, "DecodeGraphRunner", lambda lm, cache, dev: SimpleNamespace(cache=cache))
+    ctx = pipe_mod.SenseNovaU1Pipeline._decode_context
+
+    first = ctx(host, _dyn_cache(10))
+    assert ctx(host, _dyn_cache(10))[0] is first[0], "the fixture is not reusing to begin with"
+    pipe_mod.SenseNovaU1Pipeline.release_captured_graphs(host)
+    after = ctx(host, _dyn_cache(10))
+    assert after[0] is not first[0], "the cache survived the release"
+    assert after[1] is not first[1], "the runner survived the release"
+
+
+class _FakeLoRAWrapper(BaseLayerWithLoRA):
+    """Stands in for what ``replace_submodule`` leaves in the module tree."""
+
+    def __init__(self):
+        super().__init__()
+
+
+def test_no_capture_crosses_a_request_once_lora_wrappers_are_installed(monkeypatch):
+    """Base -> LoRA -> base. The reuse test cannot see an adapter change.
+
+    ``DiffusionLoRAManager`` replaces the decode path's linear layers with
+    ``BaseLayerWithLoRA`` and binds, rescales or resets slot 0 per request, and
+    the wrappers stay in the tree afterwards. A graph captured before an adapter
+    was bound would replay without its matmuls; one captured with it bound would
+    survive its removal. Neither moves the shape, dtype or bucket that decide
+    reuse, so the wrappers themselves have to end it.
+    """
+    pipe_mod, host = _pipeline_host(monkeypatch)
+    monkeypatch.setattr(pipe_mod, "DecodeGraphRunner", lambda lm, cache, dev: SimpleNamespace(cache=cache))
+    ctx = pipe_mod.SenseNovaU1Pipeline._decode_context
+
+    base = ctx(host, _dyn_cache(10))
+    assert ctx(host, _dyn_cache(10))[0] is base[0], "the fixture is not reusing to begin with"
+
+    host.language_model.add_module("q_proj", _FakeLoRAWrapper())
+    adapted = ctx(host, _dyn_cache(10))
+    assert adapted[0] is not base[0], "an adapted request reused the base capture"
+    assert adapted[1] is not base[1], "an adapted request reused the base runner"
+
+    back = ctx(host, _dyn_cache(10))
+    assert back[0] is not adapted[0], "a later request reused the adapted capture"
+    assert back[1] is not adapted[1], "a later request reused the adapted runner"
+    assert getattr(host, "_paged_decode", None) is None, "an adapted capture was left on the pipeline"
+
+
+@cuda_only
+@pytest.mark.cuda
+@pytest.mark.L4
+def test_a_second_request_replays_the_first_capture(monkeypatch):
+    """Reuse is only worth anything if the capture survives it."""
+    dev = torch.device("cuda")
+    pipe_mod, host = _pipeline_host(monkeypatch, device="cuda", language_model=_StubLM(dev))
+    ctx = pipe_mod.SenseNovaU1Pipeline._decode_context
+
+    cache, runner = ctx(host, _dyn_cache(10, dev))
+    cache.set_length(cache.length + 1)
+    runner.step(0, 0)
+    assert runner.captures == 1
+
+    cache2, runner2 = ctx(host, _dyn_cache(12, dev))
+    assert cache2 is cache and runner2 is runner
+    cache2.set_length(cache2.length + 1)
+    runner2.step(0, 0)
+    assert runner.captures == 1, "the second request captured the graph again"
+
+
+class _Tok:
+    """The two tokenizer calls ``_generate_text`` makes."""
+
+    EOS = 7
+
+    def convert_tokens_to_ids(self, token):
+        return self.EOS
+
+    def decode(self, ids, skip_special_tokens=False):
+        return " ".join(str(i) for i in ids)
+
+
+class _Out:
+    def __init__(self, logits, past_key_values):
+        self.logits = logits
+        self.past_key_values = past_key_values
+
+
+class _TextHost:
+    """Carries only what ``_generate_text`` touches."""
+
+    def __init__(self, context):
+        self._context = context
+        self.seen: list[object] = []
+        self.tokenizer = _Tok()
+
+    def _decode_context(self, past_key_values):
+        return self._context
+
+    def _ar_step(self, next_token, t_idx, past_key_values, decode=None):
+        self.seen.append(decode)
+        # Stop the loop: argmax of this row is the EOS id.
+        logits = torch.full((1, 1, 8), -1.0)
+        logits[0, 0, _Tok.EOS] = 1.0
+        return _Out(logits, past_key_values)
+
+
+def test_text_decoding_uses_the_paged_context_too():
+    """Image-to-text and text-to-text run their own decode loop. It used to call
+    ``_ar_step`` without the context, so the paged path was documented but never
+    reached outside think."""
+    from vllm_omni.diffusion.models.sensenova_u1.pipeline_sensenova_u1 import (
+        SenseNovaU1Pipeline,
+    )
+
+    context = object()
+    host = _TextHost(context)
+    prefix = torch.full((1, 1, 8), -1.0)
+    prefix[0, 0, 5] = 1.0  # not EOS, so one step runs
+    SenseNovaU1Pipeline._generate_text(host, prefix, past_key_values=None, t_idx=0)
+    assert host.seen == [context], f"text decoding ran with decode={host.seen}"
+
+
+class _WarmupReq:
+    """The engine's dummy request, which carries the reserved id."""
+
+    def __init__(self, request_id):
+        self.request_id = request_id
+        self.prompts = [{"prompt": "", "modalities": ["image"]}]
+
+
+def test_the_dummy_warmup_request_drives_one_decode_step(monkeypatch):
+    """The dummy request is a text-to-image run with think off, so it exercises
+    the prefill shape and never the decode one. Without this hook whatever the
+    deploy config compiles for decode lands on the first real request."""
+    from vllm_omni.diffusion.models.sensenova_u1.pipeline_sensenova_u1 import (
+        SenseNovaU1Pipeline,
+    )
+    from vllm_omni.diffusion.request import DUMMY_DIFFUSION_REQUEST_ID
+
+    calls: list[str] = []
+    host = object.__new__(SenseNovaU1Pipeline)
+    monkeypatch.setattr(host, "_warm_ar_decode", lambda: calls.append("warm"), raising=False)
+    monkeypatch.setattr(
+        SenseNovaU1Pipeline,
+        "_parse_request",
+        lambda self, req: (_ for _ in ()).throw(_StopForwardError()),
+        raising=False,
+    )
+
+    with pytest.raises(_StopForwardError):
+        SenseNovaU1Pipeline.forward(host, _WarmupReq(DUMMY_DIFFUSION_REQUEST_ID))
+    assert calls == ["warm"], "the dummy warmup request did not drive a decode step"
+
+    calls.clear()
+    with pytest.raises(_StopForwardError):
+        SenseNovaU1Pipeline.forward(host, _WarmupReq("a-real-request"))
+    assert calls == [], "a real request paid for the warmup"
+
+
+class _StopForwardError(Exception):
+    """Stops ``forward`` right after the warmup hook, so the test needs no model."""

tests/diffusion/models/sensenova_u1/test_sensenova_u1_rmsnorm.py

@@ -0,0 +1,107 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""CPU tests for SenseNova-U1 RMSNorm and the U1.5 config surface.
+
+``Qwen3RMSNorm`` was a hand-rolled fp32 up/down-cast chain; it now calls the fused
+``F.rms_norm``. These pin the mathematics against an independent reference so a
+future kernel swap cannot silently change what the layer computes, and pin the two
+config fields that distinguish U1.5 from U1.
+"""
+
+import pytest
+import torch
+
+from vllm_omni.diffusion.models.sensenova_u1.sensenova_u1_transformer import Qwen3RMSNorm
+from vllm_omni.transformers_utils.configs.sensenova_u1 import SenseNovaU1Config
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
+
+
+def _reference(x: torch.Tensor, weight: torch.Tensor, eps: float) -> torch.Tensor:
+    """RMSNorm written out in float64 so it is independent of the implementation."""
+    x64 = x.to(torch.float64)
+    var = x64.pow(2).mean(-1, keepdim=True)
+    return (weight.to(torch.float64) * (x64 * torch.rsqrt(var + eps))).to(x.dtype)
+
+
+@pytest.mark.parametrize("hidden", [8, 128, 4096])
+@pytest.mark.parametrize(
+    "dtype,atol",
+    [(torch.float32, 1e-5), (torch.bfloat16, 3e-2), (torch.float16, 4e-3)],
+)
+def test_rmsnorm_matches_float64_reference(hidden: int, dtype: torch.dtype, atol: float):
+    torch.manual_seed(0)
+    norm = Qwen3RMSNorm(hidden, eps=1e-6).to(dtype)
+    with torch.no_grad():
+        norm.weight.copy_(torch.rand(hidden, dtype=dtype) + 0.5)
+    x = torch.randn(7, hidden, dtype=dtype)
+    got = norm(x)
+    want = _reference(x, norm.weight, norm.variance_epsilon)
+    assert got.dtype == dtype
+    assert got.shape == x.shape
+    torch.testing.assert_close(got.to(torch.float32), want.to(torch.float32), atol=atol, rtol=atol)
+
+
+def _cast_chain(x: torch.Tensor, weight: torch.Tensor, eps: float) -> torch.Tensor:
+    """The implementation this PR replaced, kept so its accuracy can be compared."""
+    var = x.to(torch.float32).pow(2).mean(-1, keepdim=True)
+    return weight * (x.to(torch.float32) * torch.rsqrt(var + eps)).to(weight.dtype)
+
+
+@pytest.mark.parametrize("hidden", [512, 4096])
+@pytest.mark.parametrize("dtype", [torch.bfloat16, torch.float16])
+def test_rmsnorm_is_closer_to_float64_than_the_cast_chain(hidden: int, dtype: torch.dtype):
+    """The cast chain rounds to ``dtype`` before the weight multiply and so rounds
+    twice, while ``F.rms_norm`` rounds once. Restoring the cast chain fails here."""
+    torch.manual_seed(0)
+    norm = Qwen3RMSNorm(hidden, eps=1e-6).to(dtype)
+    with torch.no_grad():
+        norm.weight.copy_(torch.rand(hidden, dtype=dtype) + 0.5)
+    x = torch.randn(64, hidden, dtype=dtype)
+
+    reference = _reference(x, norm.weight, norm.variance_epsilon).to(torch.float64)
+    scale = reference.abs().mean()
+
+    def relative_error(out: torch.Tensor) -> float:
+        return ((out.to(torch.float64) - reference).abs().mean() / scale).item()
+
+    fused = relative_error(norm(x))
+    legacy = relative_error(_cast_chain(x, norm.weight, norm.variance_epsilon))
+    assert fused < legacy, f"{dtype} hidden={hidden}: fused {fused:.3e} not better than {legacy:.3e}"
+
+
+def test_rmsnorm_is_scale_equivariant():
+    """RMSNorm divides by the RMS, so scaling the input must not change the output."""
+    torch.manual_seed(0)
+    norm = Qwen3RMSNorm(64, eps=1e-12)
+    x = torch.randn(4, 64)
+    torch.testing.assert_close(norm(x), norm(x * 7.5), atol=1e-4, rtol=1e-4)
+
+
+def test_rmsnorm_preserves_leading_dims():
+    norm = Qwen3RMSNorm(16, eps=1e-6)
+    for shape in [(16,), (3, 16), (2, 3, 16)]:
+        assert norm(torch.randn(*shape)).shape == shape
+
+
+def test_rmsnorm_weight_is_applied():
+    norm = Qwen3RMSNorm(4, eps=1e-6)
+    x = torch.randn(2, 4)
+    base = norm(x)
+    with torch.no_grad():
+        norm.weight.mul_(3.0)
+    torch.testing.assert_close(norm(x), base * 3.0, atol=1e-5, rtol=1e-5)
+
+
+def test_u15_config_fields_default_to_u1_values():
+    """U1.5 flips exactly these two; the defaults must stay U1's so old checkpoints
+    keep their behaviour when the keys are absent."""
+    cfg = SenseNovaU1Config()
+    assert cfg.use_pixel_head is False
+    assert cfg.noise_scale_max_value == 10.0
+
+
+def test_u15_config_fields_are_read_from_kwargs():
+    cfg = SenseNovaU1Config(use_pixel_head=True, noise_scale_max_value=16.0)
+    assert cfg.use_pixel_head is True
+    assert cfg.noise_scale_max_value == 16.0

tests/diffusion/test_diffusion_model_runner.py

@@ -179,6 +179,30 @@ def test_request_scoped_cache_dit_lifecycle_is_pipeline_opt_in():
     assert events == [backend]
 
 
+def test_release_captured_graphs_clears_runners_and_delegates_to_the_pipeline():
+    """Both halves in one place, so a pipeline that keeps captures is collected
+    without the caller knowing which pipelines have one."""
+    released = []
+    runner = object.__new__(DiffusionModelRunner)
+    runner.graph_runners = {"decode": object()}
+    runner.pipeline = SimpleNamespace(release_captured_graphs=lambda: released.append(True))
+
+    runner.release_captured_graphs()
+
+    assert runner.graph_runners == {}
+    assert released == [True]
+
+
+def test_release_captured_graphs_tolerates_a_pipeline_without_captures():
+    """Most pipelines keep none, and the runner may not carry `graph_runners`."""
+    runner = object.__new__(DiffusionModelRunner)
+    runner.pipeline = object()
+
+    runner.release_captured_graphs()
+
+    assert not hasattr(runner, "graph_runners")
+
+
 def _make_runner(cache_backend, cache_backend_name: str, enable_cache_dit_summary: bool = True):
     runner = object.__new__(DiffusionModelRunner)
     runner.vllm_config = object()

tests/diffusion/test_diffusion_worker.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 """
 Unit tests for DiffusionWorker class.
@@ -143,6 +143,21 @@ class TestDiffusionWorkerSleep:
         assert "buffer1" in mock_gpu_worker._sleep_saved_buffers
         assert "buffer2" in mock_gpu_worker._sleep_saved_buffers
 
+    def test_sleep_level_2_releases_captures_through_the_model_runner(self, mocker: MockerFixture, mock_gpu_worker):
+        """Level 2 discards the memory captures were recorded against, and the
+        runner owns the release. The worker asks it once and does not reach past
+        it into the pipeline."""
+        mock_platform = mocker.patch("vllm_omni.diffusion.worker.diffusion_worker.current_omni_platform")
+        mock_platform.get_free_memory.side_effect = [10 * 1024**3, 12 * 1024**3]
+        mock_platform.get_device_total_memory.return_value = 80 * 1024**3
+        mocker.patch("vllm_omni.diffusion.worker.diffusion_worker.get_process_gpu_memory", side_effect=[0, 0])
+        mock_gpu_worker.model_runner.pipeline.named_buffers = mocker.Mock(return_value=[])
+
+        mock_gpu_worker.sleep(level=2)
+
+        mock_gpu_worker.model_runner.release_captured_graphs.assert_called_once_with()
+        mock_gpu_worker.model_runner.pipeline.release_captured_graphs.assert_not_called()
+
     def test_sleep_memory_freed_validation(self, mocker: MockerFixture, mock_gpu_worker):
         """Test that sleep validates memory was actually freed."""
         mock_allocator_class = patch_cumem_allocator(mocker)

vllm_omni/config/environment_variable_inventory.py

@@ -167,6 +167,7 @@ _MODEL_PROMOTE = (
     "VLLM_GEPARD_LOOKBACK_FRAMES",
     "VLLM_OMNI_FISH_KVCACHE_ATTN",
     "VLLM_OMNI_SANA_WM_STAGE1_TEXT_ENCODER",
+    "VLLM_OMNI_SENSENOVA_PAGED_DECODE",
 )
 
 _MODEL_REQUEST_SCOPE = (

vllm_omni/diffusion/lora/loader.py

@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 import os
 from collections import OrderedDict, defaultdict
 from collections.abc import Callable
@@ -41,7 +44,7 @@ def _prepare_lora_delta(
     lora_b_suffix: str = "lora_B.weight",
     lora_bias_suffix: str = "bias",
 ):
-    used_keys = set()
+    used_keys: set[str] = set()
     # stacked_params_mapping                       param_to_weight_names
     # [(".to_qkv", ".to_q.", "q")
     # (".to_qkv", ".to_k.", "k")  ========> {".to_qkv": [".to_q", ".to_k", ".to_v"]}
@@ -54,7 +57,10 @@ def _prepare_lora_delta(
     if param_to_weight_names is None:
         param_to_weight_names = defaultdict(list)
     for param_name, weight_names in param_to_weight_names.items():
-        if param_name not in base_key:
+        # Fused names can overlap by prefix (`.qkv_proj` is a substring of
+        # `.qkv_proj_mot_gen`), and a substring match stacks both mappings into
+        # one delta of twice the height.
+        if not base_key.endswith(param_name):
             continue
         is_stacked_param = True
         # handle lora_a_key and lora_b_key together
@@ -78,7 +84,7 @@ def _prepare_lora_delta(
                 stacked_deltas.append(delta)
                 used_keys.add(lora_a_key)
                 used_keys.add(lora_b_key)
-        continue
+        break
 
     if is_stacked_param:
         return torch.concat(stacked_deltas), used_keys
@@ -200,6 +206,10 @@ def _apply_diffusers_lora_alpha_scaling(state_dict: dict[str, torch.Tensor]) ->
 class LoraLoaderMixin:
     transformer_name = "transformer"
 
+    # Set by the pipeline this mixin is combined with; annotated, not assigned.
+    transformer: torch.nn.Module
+    _lora_loaded: dict[str | None, dict[str, torch.Tensor]]
+
     # Lazy initialization to avoid MRO issues: __init__ may not be called
     # when mixin with nn.Module
     @property
@@ -373,6 +383,7 @@ class QwenImageLoraLoaderMixin(LoraLoaderMixin):
 
 class WanLoraLoaderMixin(LoraLoaderMixin):
     transformer_2_name = "transformer_2"
+    has_transformer_2: bool
 
     def load_lora_weights(
         self,

vllm_omni/diffusion/models/sensenova_u1/paged_decode.py

@@ -0,0 +1,368 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""Paged KV cache for SenseNova-U1 autoregressive decode, and a CUDA graph over it.
+
+Autoregressive decode dominates a think request -- 80% of the wall clock, at
+about 20 ms per token for 42 layers -- and almost all of the GPU idle in that
+window is Python/ATen dispatch rather than any CUDA call. A CUDA graph retires
+those dispatches, but capture needs static shapes, and the default cache grows
+K/V with ``torch.cat`` on every step.
+
+Padding the cache to a bucket and masking the tail does give static shapes, but
+the mask is what costs: measured on an A800 at real decode shapes, a masked
+bucket attention runs 7.96-11.55 ms per step against 1.03 ms for an exact-length
+unmasked one, which is more than the graph saves.
+
+The way out is a paged cache. ``flash_attn_varlen_func`` takes the used length
+as a *tensor* (``seqused_k``) alongside a ``block_table``, so the buffers stay
+bucket-sized and capturable while the kernel reads only the valid prefix. One
+captured graph then serves every length in the bucket.
+
+Scope, and when to delete this. The cache is model-local on purpose:
+``DiffusionKVCacheManager`` reserves once per scheduler request, and
+``ARDiffusionModelRunner`` lives under ``vllm_omni/experimental``. SenseNova runs
+its whole autoregressive loop inside one pipeline ``forward``, so the decode
+steps never surface to the scheduler and it cannot admit, pool, evict, reuse a
+prefix for, or continuously batch them. What is here is therefore one set of
+buffers behind an identity block table, reused by whichever request fits them,
+not general paged-KV support: it holds while the pipeline serves one sequence
+per forward, and is released when sleep discards the memory it captured
+against. Once the decode loop is
+driven by a scheduler-visible runner, or the pipeline handles more than one
+sequence per forward, delete this file and use the manager instead.
+"""
+
+from __future__ import annotations
+
+import inspect
+
+import torch
+from vllm.logger import init_logger
+from vllm.platforms import current_platform
+
+logger = init_logger(__name__)
+
+BLOCK_SIZE = 16
+# Growth schedule. Each new bucket costs one capture, so keep the list short and
+# let the last entry cover the tail rather than doubling forever.
+BUCKETS = (512, 1024, 2048, 4096, 8192)
+# Past the last bucket the schedule grows by this much at a time. A request
+# decodes at most `max_think_tokens` (1024) or `max_tokens` (512) steps, so a
+# step this size costs at most one reallocation per request. Doubling the last
+# bucket instead would also cost at most one, but a 9,100-token request would
+# then reserve 766 MiB it never uses, against 118 MiB here.
+TAIL_STEP = 2048
+
+
+# Every argument to the kernel is passed by keyword, so the probe covers all of
+# them. An older wheel can export `flash_attn_varlen_func` without the paged
+# ones, or under different names, and a version string would not say which;
+# probe the signature and fall back to SDPA when any is missing.
+_REQUIRED_KWARGS = (
+    "max_seqlen_q",
+    "cu_seqlens_q",
+    "max_seqlen_k",
+    "seqused_k",
+    "block_table",
+    "softmax_scale",
+    "causal",
+)
+
+
+def _flash_varlen():
+    """Import lazily: the bundled kernel is CUDA-only and optional."""
+    from vllm.vllm_flash_attn import flash_attn_varlen_func
+
+    return flash_attn_varlen_func
+
+
+def _accepts_paged_kwargs(fn) -> bool:
+    try:
+        params = inspect.signature(fn).parameters
+    except (TypeError, ValueError):  # pragma: no cover - no introspectable signature
+        return False
+    return all(name in params for name in _REQUIRED_KWARGS)
+
+
+def paged_decode_supported(device: torch.device, head_dim: int) -> bool:
+    if device.type != "cuda":
+        return False
+    if head_dim % 8 or head_dim > 256:
+        return False
+    try:
+        fn = _flash_varlen()
+    except Exception as exc:  # pragma: no cover - depends on the wheel
+        logger.debug("Paged decode unavailable: %s", exc)
+        return False
+    if not _accepts_paged_kwargs(fn):
+        logger.debug("Paged decode unavailable: flash_attn_varlen_func lacks %s", _REQUIRED_KWARGS)
+        return False
+    return True
+
+
+def dynamic_lora_wrappers_present(module) -> bool:
+    """True once ``DiffusionLoRAManager`` has wrapped layers under ``module``.
+
+    The manager replaces linear layers with ``BaseLayerWithLoRA`` and then
+    binds, rescales or resets slot 0 per request; the wrappers stay in the tree
+    afterwards. Capture records the module tree and the branches that ran during
+    it, so a graph taken before an adapter was bound would replay without the
+    adapter's matmuls and one taken with it bound would survive its removal --
+    both silently, and neither visible in the shape, dtype and bucket that decide
+    reuse. Nothing captured may outlive a request once these exist.
+
+    The distilled LoRA is not affected: it is fused one-way into the weights and
+    leaves no wrapper behind.
+    """
+    try:
+        from vllm.lora.layers import BaseLayerWithLoRA
+    except ImportError:  # pragma: no cover - depends on the wheel
+        return False
+    return any(isinstance(m, BaseLayerWithLoRA) for m in module.modules())
+
+
+def _bucket_for(length: int) -> int:
+    for b in BUCKETS:
+        if length <= b:
+            return b
+    # Past the schedule, grow in TAIL_STEP-sized steps. Rounding to a whole
+    # number of blocks here would reallocate and re-capture every BLOCK_SIZE
+    # tokens: a think edit with two 2048x2048 inputs runs to about 9,100 tokens
+    # and paid 41 reallocations and 42 captures in one request.
+    over = length - BUCKETS[-1]
+    return BUCKETS[-1] + ((over + TAIL_STEP - 1) // TAIL_STEP) * TAIL_STEP
+
+
+class PagedDecodeCache:
+    """Per-layer paged K/V for single-token decode.
+
+    Layout follows what ``flash_attn_varlen_func`` accepts with a block table:
+    ``(num_blocks, BLOCK_SIZE, kv_heads, head_dim)``. ``seqused`` holds the
+    number of valid tokens and is the only thing that changes between decode
+    steps, which is what lets a captured graph be replayed unchanged.
+    """
+
+    def __init__(self, num_layers, kv_heads, head_dim, length, device, dtype):
+        self.kv_heads = kv_heads
+        self.head_dim = head_dim
+        self.device = device
+        self.dtype = dtype
+        self.bucket = _bucket_for(length)
+        nblocks = self.bucket // BLOCK_SIZE
+        shape = (nblocks, BLOCK_SIZE, kv_heads, head_dim)
+        self.k = [torch.zeros(shape, device=device, dtype=dtype) for _ in range(num_layers)]
+        self.v = [torch.zeros(shape, device=device, dtype=dtype) for _ in range(num_layers)]
+        self.block_table = torch.arange(nblocks, device=device, dtype=torch.int32).unsqueeze(0)
+        self.cu_seqlens_q = torch.tensor([0, 1], device=device, dtype=torch.int32)
+        self.seqused = torch.zeros(1, device=device, dtype=torch.int32)
+        # The write slot has to live in a tensor, not a Python int. Under graph
+        # capture a Python index is baked in at capture time, so every replay
+        # would overwrite the same slot -- correct eagerly, silently wrong once
+        # captured.
+        self.pos = torch.zeros(1, device=device, dtype=torch.int64)
+        self._length = 0
+        # Bumped whenever the buffers are reallocated, so a graph runner can
+        # tell that its capture no longer points at live memory.
+        self.generation = 0
+
+    @property
+    def length(self) -> int:
+        return self._length
+
+    def set_length(self, n: int) -> None:
+        self._length = n
+        self.seqused.fill_(n)
+        self.pos.fill_(n - 1)
+
+    def reusable_for(self, num_layers, kv_heads, head_dim, dtype, length) -> bool:
+        """Can this cache serve another request without reallocating?
+
+        Reusing the buffers is what lets captured graphs survive across
+        requests: a capture binds the addresses it recorded, so a fresh cache
+        forces a fresh capture.
+        """
+        return (
+            len(self.k) == num_layers
+            and self.kv_heads == kv_heads
+            and self.head_dim == head_dim
+            and self.dtype == dtype
+            and _bucket_for(length) <= self.bucket
+        )
+
+    def load_prefix(self, cache) -> bool:
+        """Copy a prefill cache into the paged buffers.
+
+        Only the first ``seqused`` rows are ever read, so the tail is left as
+        whatever the previous request wrote. Returns False for an empty cache.
+        """
+        layer0 = cache.layers[0]
+        if layer0.keys is None:
+            return False
+        prefix = layer0.keys.shape[2]
+        for i in range(len(self.k)):
+            keys, values = cache.layers[i].keys, cache.layers[i].values
+            # [B, H, S, D] -> [S, H, D], then into the paged blocks
+            self._write_prefix(i, keys[0].transpose(0, 1), values[0].transpose(0, 1))
+        self.set_length(prefix)
+        return True
+
+    @classmethod
+    def from_dynamic_cache(cls, cache, num_layers, device, dtype, min_length=0):
+        """Allocate buffers for a prefill cache and copy it in."""
+        layer0 = cache.layers[0]
+        if layer0.keys is None:
+            return None
+        _, kv_heads, prefix, head_dim = layer0.keys.shape
+        obj = cls(num_layers, kv_heads, head_dim, max(prefix + 1, min_length), device, dtype)
+        obj.load_prefix(cache)
+        return obj
+
+    def _write_prefix(self, layer_idx, keys_shd, values_shd):
+        n = keys_shd.shape[0]
+        flat_k = self.k[layer_idx].view(-1, self.kv_heads, self.head_dim)
+        flat_v = self.v[layer_idx].view(-1, self.kv_heads, self.head_dim)
+        flat_k[:n].copy_(keys_shd)
+        flat_v[:n].copy_(values_shd)
+
+    def grow(self, length: int) -> bool:
+        """Move to the next bucket, keeping the tokens already stored.
+
+        Everything is built into locals first and published in one step. A
+        resize that failed part way -- an allocation for a later layer, say --
+        would otherwise leave new K beside old V under the ``generation`` a
+        captured graph is keyed on, so the next request would keep selecting
+        that graph and replay it over storage that had just been freed.
+        """
+        new_bucket = _bucket_for(length)
+        if new_bucket == self.bucket:
+            return False
+        nblocks = new_bucket // BLOCK_SIZE
+        shape = (nblocks, BLOCK_SIZE, self.kv_heads, self.head_dim)
+        old_len = self._length
+        new_k, new_v = [], []
+        for old_k, old_v in zip(self.k, self.v):
+            k = torch.zeros(shape, device=self.device, dtype=self.dtype)
+            v = torch.zeros(shape, device=self.device, dtype=self.dtype)
+            k.view(-1, self.kv_heads, self.head_dim)[:old_len].copy_(
+                old_k.view(-1, self.kv_heads, self.head_dim)[:old_len]
+            )
+            v.view(-1, self.kv_heads, self.head_dim)[:old_len].copy_(
+                old_v.view(-1, self.kv_heads, self.head_dim)[:old_len]
+            )
+            new_k.append(k)
+            new_v.append(v)
+        block_table = torch.arange(nblocks, device=self.device, dtype=torch.int32).unsqueeze(0)
+
+        # Publish. New buffers mean any graph captured against the old ones is
+        # stale, so the generation moves in the same step that hands them over.
+        self.k, self.v, self.block_table, self.bucket = new_k, new_v, block_table, new_bucket
+        self.generation += 1
+        self.set_length(old_len)
+        logger.debug("Paged decode grew to bucket %d at length %d (generation %d)", new_bucket, length, self.generation)
+        return True
+
+    def attend(self, layer_idx, query_bhsd, key_bhsd, value_bhsd, softmax_scale):
+        """Append this step's K/V and attend over the whole valid prefix.
+
+        Writes at ``length - 1`` because the caller has already advanced the
+        length for this step. Both the slot and the attended length come from
+        device tensors, so a captured graph reads whatever they hold at replay
+        rather than whatever they held at capture.
+        """
+        flat_k = self.k[layer_idx].view(-1, self.kv_heads, self.head_dim)
+        flat_v = self.v[layer_idx].view(-1, self.kv_heads, self.head_dim)
+        flat_k.index_copy_(0, self.pos, key_bhsd[0, :, 0].unsqueeze(0))
+        flat_v.index_copy_(0, self.pos, value_bhsd[0, :, 0].unsqueeze(0))
+        heads, head_dim = query_bhsd.shape[1], query_bhsd.shape[3]
+        q = query_bhsd[0].transpose(0, 1).reshape(1, heads, head_dim)
+        out = _flash_varlen()(
+            q,
+            self.k[layer_idx],
+            self.v[layer_idx],
+            max_seqlen_q=1,
+            cu_seqlens_q=self.cu_seqlens_q,
+            max_seqlen_k=self.bucket,
+            seqused_k=self.seqused,
+            block_table=self.block_table,
+            softmax_scale=softmax_scale,
+            causal=False,
+        )
+        return out.reshape(1, 1, heads, head_dim)
+
+    def to_dynamic_cache(self, cache) -> None:
+        """Write the decoded K/V back so the generation stage sees one cache.
+
+        Only decode is paged; the DiT stage that follows reads the ordinary
+        cache, so the two are reconciled once at the hand-off rather than kept
+        in sync on every step.
+        """
+        n = self._length
+        for i, layer in enumerate(cache.layers):
+            keys = self.k[i].view(-1, self.kv_heads, self.head_dim)[:n]
+            values = self.v[i].view(-1, self.kv_heads, self.head_dim)[:n]
+            layer.keys = keys.transpose(0, 1).unsqueeze(0).contiguous()
+            layer.values = values.transpose(0, 1).unsqueeze(0).contiguous()
+
+
+class DecodeGraphRunner:
+    """One captured decode step per KV bucket, replayed for every token.
+
+    The whole point of the paged cache is that everything the step reads which
+    varies -- the token, its position, and how much of the cache is live -- sits
+    in device tensors. So a single capture serves every step in a bucket: fill
+    the tensors, replay, read the logits out of the static output.
+
+    A capture is invalidated when the cache reallocates (tracked by
+    ``PagedDecodeCache.generation``), which happens once per bucket boundary.
+    """
+
+    def __init__(self, language_model, cache, device):
+        self.lm = language_model
+        self.cache = cache
+        self.device = device
+        self.input_ids = torch.zeros(1, 1, dtype=torch.long, device=device)
+        self.indexes = torch.zeros(3, 1, dtype=torch.long, device=device)
+        self._graphs: dict[tuple[int, int], tuple[torch.cuda.CUDAGraph, torch.Tensor]] = {}
+        self.captures = 0
+
+    def _forward(self):
+        return self.lm(
+            input_ids=self.input_ids,
+            indexes=self.indexes,
+            past_key_values=None,
+            use_cache=False,
+            paged_cache=self.cache,
+        ).logits
+
+    def _capture(self, key):
+        # Warm up on a side stream first so cuBLAS workspaces and any lazy
+        # allocation happen outside the capture. The warm-up recomputes this
+        # same step, writing the same K/V to the same slot, so it is idempotent.
+        side = torch.cuda.Stream()
+        side.wait_stream(torch.cuda.current_stream())
+        with torch.cuda.stream(side):
+            for _ in range(3):
+                self._forward()
+        torch.cuda.current_stream().wait_stream(side)
+
+        graph = torch.cuda.CUDAGraph()
+        # A private pool would hand every runner its own arena and never give it
+        # back, and a runner is rebuilt whenever the cache reallocates or a LoRA
+        # wrapper appears: measured, that grew device memory by about 40 MiB per
+        # request and never levelled off. The platform pool is shared, which is
+        # what every other captured path in the tree uses.
+        with torch.cuda.graph(graph, pool=current_platform.get_global_graph_pool()):
+            logits = self._forward()
+        self.captures += 1
+        logger.debug("Captured decode graph for bucket=%d generation=%d", self.cache.bucket, self.cache.generation)
+        self._graphs[key] = (graph, logits)
+        return self._graphs[key]
+
+    def step(self, token, t_index):
+        """Run one decode step. Returns the static logits tensor."""
+        self.input_ids[0, 0] = token
+        self.indexes[0, 0] = t_index
+        key = (self.cache.bucket, self.cache.generation)
+        entry = self._graphs.get(key) or self._capture(key)
+        entry[0].replay()
+        return entry[1]

vllm_omni/diffusion/models/sensenova_u1/pipeline_sensenova_u1.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """SenseNova-U1 Pipeline for vLLM-Omni.
 
 SenseNova-U1 is a unified Qwen3-based model that uses Mixture-of-Tokenizers
@@ -35,6 +35,13 @@ from vllm.logger import init_logger
 from vllm_omni.diffusion.data import DiffusionOutput, OmniDiffusionConfig
 from vllm_omni.diffusion.distributed.cfg_parallel import CFGParallelMixin
 from vllm_omni.diffusion.distributed.utils import get_local_device
+from vllm_omni.diffusion.lora.loader import (
+    LoraLoaderMixin,
+    _apply_diffusers_lora_alpha_scaling,
+    _load_lora_state_dict,
+    _prepare_lora_delta,
+    _remap_state_dict_keys,
+)
 from vllm_omni.diffusion.model_loader.diffusers_loader import DiffusersPipelineLoader
 from vllm_omni.diffusion.models.interface import SupportsComponentDiscovery
 from vllm_omni.diffusion.profiler.diffusion_pipeline_profiler import DiffusionPipelineProfilerMixin
@@ -44,7 +51,14 @@ from vllm_omni.transformers_utils.configs.sensenova_u1 import (
     SenseNovaU1Config,
 )
 
+from .paged_decode import (
+    DecodeGraphRunner,
+    PagedDecodeCache,
+    dynamic_lora_wrappers_present,
+    paged_decode_supported,
+)
 from .sensenova_u1_transformer import (
+    SenseNovaU1CausalLMOutput,
     SenseNovaU1ForCausalLM,
     clear_flash_kv_cache,
     create_block_causal_mask,
@@ -446,7 +460,13 @@ class SenseNovaU1DenoisingAdapter(nn.Module):
         return self.language_model(*args, **kwargs)
 
 
-class SenseNovaU1Pipeline(nn.Module, SupportsComponentDiscovery, DiffusionPipelineProfilerMixin, CFGParallelMixin):
+class SenseNovaU1Pipeline(
+    nn.Module,
+    SupportsComponentDiscovery,
+    DiffusionPipelineProfilerMixin,
+    CFGParallelMixin,
+    LoraLoaderMixin,
+):
     """SenseNova-U1 text-to-image and image-to-image pipeline for vllm-omni.
 
     Builds the full model graph internally:
@@ -709,8 +729,71 @@ class SenseNovaU1Pipeline(nn.Module, SupportsComponentDiscovery, DiffusionPipeli
         sigma = shift * sigma / (1 + (shift - 1) * sigma)
         return 1 - sigma
 
-    def _ar_step(self, next_token, t_idx, past_key_values):
+    def _decode_context(self, past_key_values):
+        """Paged cache plus a captured graph for the AR loop, or ``None``.
+
+        Decode is where a think request spends most of its time, and every step
+        issues the same work at the same shapes, so it is worth capturing. The
+        cache has to be paged for that: it keeps the buffers a fixed size while
+        the kernel reads only ``seqused_k`` of them, which is what lets one
+        capture serve every step until the next bucket.
+
+        Both live on the pipeline and are reused across requests unless a LoRA
+        adapter is being served dynamically. A capture binds the addresses it
+        recorded, so a cache built per request forces a capture per request:
+        measured, that left about 40 MiB of device memory behind on every
+        request and re-captured every time. Reuse rests on the same thing the
+        paged path already rests on -- one sequence in flight per pipeline
+        forward.
+        """
+        if os.environ.get("VLLM_OMNI_SENSENOVA_PAGED_DECODE", "1") != "1":
+            return None
+        head_dim = self.language_model.model.layers[0].self_attn.head_dim
+        if not paged_decode_supported(torch.device(self.device), head_dim):
+            return None
+        if past_key_values is None or not past_key_values.layers:
+            return None
+        layer0 = past_key_values.layers[0]
+        if layer0.keys is None:
+            return None
+        _, kv_heads, prefix, kv_head_dim = layer0.keys.shape
+        device = torch.device(self.device)
+        num_layers = len(self.language_model.model.layers)
+        # Reuse is off once the LoRA manager holds the decode path: the test
+        # below sees shape, dtype and bucket, none of which move when an adapter
+        # is bound, rescaled or reset between requests.
+        reuse = not dynamic_lora_wrappers_present(self.language_model)
+        decode = getattr(self, "_paged_decode", None) if reuse else None
+        if decode is None or not decode[0].reusable_for(
+            num_layers, kv_heads, kv_head_dim, layer0.keys.dtype, prefix + 1
+        ):
+            cache = PagedDecodeCache(num_layers, kv_heads, kv_head_dim, prefix + 1, device, layer0.keys.dtype)
+            decode = (cache, DecodeGraphRunner(self.language_model, cache, device))
+            self._paged_decode = decode if reuse else None
+        if not decode[0].load_prefix(past_key_values):
+            return None
+        return decode
+
+    def release_captured_graphs(self) -> None:
+        """Drop the reused paged cache and the graphs captured against it.
+
+        Sleep level 2 discards the memory a capture recorded, so anything held
+        across requests has to go with it. The next request rebuilds both.
+        """
+        self._paged_decode = None
+
+    def _ar_step(self, next_token, t_idx, past_key_values, decode=None):
         """One autoregressive token step. Constructs single-token `indexes` explicitly."""
+        if decode is not None:
+            cache, runner = decode
+            if cache.length + 1 > cache.bucket:
+                cache.grow(cache.length + 1)
+            cache.set_length(cache.length + 1)
+            logits = runner.step(int(next_token), t_idx + 1)
+            # The caller reassigns its cache handle from this field; the paged
+            # tokens are merged into it once, at the hand-off below.
+            return SenseNovaU1CausalLMOutput(logits=logits, past_key_values=past_key_values)
+
         indexes = torch.tensor([[t_idx + 1], [0], [0]], dtype=torch.long, device=self.device)
         outputs = self.language_model(
             input_ids=next_token.unsqueeze(0),
@@ -725,24 +808,28 @@ class SenseNovaU1Pipeline(nn.Module, SupportsComponentDiscovery, DiffusionPipeli
         think_end_token_id = self.tokenizer.convert_tokens_to_ids("</think>")
         think_token_ids = []
         next_token = torch.argmax(prefix_outputs.logits[:, -1, :], dim=-1)
+        decode = self._decode_context(past_key_values)
 
         for _ in range(max_think_tokens):
             token_item = next_token.item()
             if token_item == eos_token_id:
                 break
             if token_item == think_end_token_id:
-                outputs = self._ar_step(next_token, t_idx, past_key_values)
+                outputs = self._ar_step(next_token, t_idx, past_key_values, decode)
                 past_key_values = outputs.past_key_values
                 t_idx += 1
                 think_token_ids.append(token_item)
                 break
 
             think_token_ids.append(token_item)
-            outputs = self._ar_step(next_token, t_idx, past_key_values)
+            outputs = self._ar_step(next_token, t_idx, past_key_values, decode)
             past_key_values = outputs.past_key_values
             t_idx += 1
             next_token = torch.argmax(outputs.logits[:, -1, :], dim=-1)
 
+        if decode is not None:
+            decode[0].to_dynamic_cache(past_key_values)
+
         # Append "\n\n<img>" tokens to cache
         append_ids = self.tokenizer(
             "\n\n<img>",
@@ -797,6 +884,10 @@ class SenseNovaU1Pipeline(nn.Module, SupportsComponentDiscovery, DiffusionPipeli
         """Autoregressive text decoding seeded from prefix logits."""
         eos_token_id = self.tokenizer.convert_tokens_to_ids("<|im_end|>")
         generated_ids: list[int] = []
+        # Same paged decode as the think loop. No write-back at the end: the
+        # caller returns the decoded string and never reads the cache again,
+        # unlike the think path, which hands it to the generation stage.
+        decode = self._decode_context(past_key_values)
 
         logits = prefix_logits[:, -1, :]
         for _ in range(max_tokens):
@@ -810,7 +901,7 @@ class SenseNovaU1Pipeline(nn.Module, SupportsComponentDiscovery, DiffusionPipeli
             if token_id == eos_token_id:
                 break
             generated_ids.append(token_id)
-            outputs = self._ar_step(next_token, t_idx, past_key_values)
+            outputs = self._ar_step(next_token, t_idx, past_key_values, decode)
             past_key_values = outputs.past_key_values
             logits = outputs.logits[:, -1, :]
             t_idx += 1
@@ -1174,8 +1265,37 @@ class SenseNovaU1Pipeline(nn.Module, SupportsComponentDiscovery, DiffusionPipeli
             layer.values = layer.values.expand(batch_size, *layer.values.shape[1:])
         prepare_flash_kv_cache(kv, current_len=token_hw, batch_size=batch_size)
 
+    @staticmethod
+    def _is_warmup_request(req) -> bool:
+        is_dummy_run = getattr(req, "is_dummy_run", None)
+        if callable(is_dummy_run):
+            return bool(is_dummy_run())
+        return OmniDiffusionRequest.is_dummy_run_request_id(getattr(req, "request_id", None))
+
+    def _warm_ar_decode(self) -> None:
+        """Run one single-token decode at startup so its compiled region is built.
+
+        The engine's dummy request is a text-to-image run with think off. It
+        exercises the prefill shape but never the decode one, so whatever the
+        deploy config compiles for decode lands on the first real request
+        instead of on readiness. Warmup is best effort and must not fail startup.
+        """
+        try:
+            lm = self.language_model
+            device = torch.device(self.device)
+            ids = torch.zeros(1, 2, dtype=torch.long, device=device)
+            idx = torch.zeros(3, 2, dtype=torch.long, device=device)
+            prefill = lm(input_ids=ids, indexes=idx, use_cache=True)
+            nxt = torch.zeros(1, 1, dtype=torch.long, device=device)
+            nidx = torch.tensor([[2], [0], [0]], dtype=torch.long, device=device)
+            lm(input_ids=nxt, indexes=nidx, past_key_values=prefill.past_key_values, use_cache=True)
+        except Exception as exc:  # pragma: no cover - warmup is best effort
+            logger.warning("Autoregressive decode warmup skipped: %s", exc)
+
     @torch.inference_mode()
     def forward(self, req: DiffusionRequestBatch) -> DiffusionOutput:
+        if self._is_warmup_request(req):
+            self._warm_ar_decode()
         p = self._parse_request(req)
 
         input_images = self._extract_input_images(p.first_prompt)
@@ -1408,21 +1528,129 @@ class SenseNovaU1Pipeline(nn.Module, SupportsComponentDiscovery, DiffusionPipeli
     # Weight loading
     # -----------------------------------------------------------------------
 
+    # Shard name -> fused parameter, used by weight loading and LoRA loading.
+    # More specific _mot_gen patterns FIRST to avoid substring ambiguity
+    # (e.g. `.q_proj` is a substring of `.q_proj_mot_gen`).
+    stacked_params_mapping: ClassVar[list[tuple[str, str, str | int]]] = [
+        (".qkv_proj_mot_gen", ".q_proj_mot_gen", "q"),
+        (".qkv_proj_mot_gen", ".k_proj_mot_gen", "k"),
+        (".qkv_proj_mot_gen", ".v_proj_mot_gen", "v"),
+        (".qkv_proj", ".q_proj", "q"),
+        (".qkv_proj", ".k_proj", "k"),
+        (".qkv_proj", ".v_proj", "v"),
+        # MLP gate/up fused into MergedColumnParallelLinear
+        (".gate_up_proj", ".gate_proj", 0),
+        (".gate_up_proj", ".up_proj", 1),
+    ]
+
+    def load_lora_weights(
+        self,
+        pretrained_model_name_or_path: str | list[str],
+        adapter_name: str | None = None,
+    ) -> None:
+        """Fuse a distilled few-step LoRA into the weights.
+
+        The checkpoints ship kohya `lora_down`/`lora_up`/`alpha` names, renamed
+        here to the Diffusers names before `load_lora_into_module` routes each
+        delta into its slice of the fused projections.
+        """
+        if adapter_name in self.lora_loaded:
+            return
+
+        paths = (
+            [pretrained_model_name_or_path]
+            if isinstance(pretrained_model_name_or_path, str)
+            else list(pretrained_model_name_or_path)
+        )
+        if len(paths) != 1:
+            raise ValueError(f"Expected exactly one distilled LoRA file, got {len(paths)}: {paths}")
+
+        state_dict = _load_lora_state_dict(paths[0])
+        state_dict = _remap_state_dict_keys(
+            state_dict,
+            [(".lora_down.weight", ".lora_A.weight"), (".lora_up.weight", ".lora_B.weight")],
+        )
+        # Fold alpha and the low-rank product in fp32; scaling B in bf16 would
+        # round once before the matmul and once after.
+        state_dict = {k: v.to(torch.float32) for k, v in state_dict.items()}
+        state_dict = _apply_diffusers_lora_alpha_scaling(state_dict)
+
+        applied = self._add_lora_deltas(state_dict)
+        if not applied:
+            raise ValueError(f"Distilled LoRA {paths[0]} matched no parameter of this pipeline.")
+        logger.info("Fused distilled LoRA %s into %d parameters", paths[0], applied)
+        # Fusion is one-way and there is no unload path, so keep only a sentinel;
+        # the fp32 state dict is ~1.5 GiB and would stay resident for the process.
+        self.lora_loaded[adapter_name] = paths[0]
+
+    def _fused_shards(self, base_key: str) -> list[tuple[str, str | int]] | None:
+        """Unfused source names and shard ids for a fused parameter, else None."""
+        for param_name, weight_name, shard_id in self.stacked_params_mapping:
+            if base_key.endswith(param_name):
+                return [
+                    (base_key[: -len(param_name)] + w, sid)
+                    for p, w, sid in self.stacked_params_mapping
+                    if p == param_name
+                ]
+        return None
+
+    def _add_lora_deltas(self, state_dict: dict[str, torch.Tensor]) -> int:
+        """Add every matching delta into its parameter, sharding through vLLM.
+
+        The delta read from the checkpoint spans the whole layer. A parameter of
+        a parallel layer only holds this rank's slice of it, so the delta goes
+        through the layer's own weight loader -- into a zeroed copy of the
+        parameter, which is then added -- instead of being added directly.
+
+        Each parameter is fused in place as it is reached, so a failure part way
+        through leaves the earlier ones fused and the pipeline has to be
+        reloaded. Deferring the writes would mean holding every delta at once,
+        and the deltas are full rank: 15.09 GiB for the 8B checkpoint, against
+        the 192 MiB one parameter costs here. Fusion runs at load time, before
+        the pipeline serves anything, so a partially fused model is never
+        reachable from a request.
+        """
+        applied = 0
+        for name, param in self.named_parameters():
+            if not name.endswith(".weight"):
+                continue
+            base_key = name[: -len(".weight")]
+            shards = self._fused_shards(base_key)
+
+            deltas: list[tuple[torch.Tensor, str | int | None]] = []
+            for src, shard_id in shards or [(base_key, None)]:
+                delta, _ = _prepare_lora_delta(state_dict, src, None)
+                if delta is None:
+                    deltas = []
+                    break
+                deltas.append((delta, shard_id))
+            if not deltas:
+                continue
+
+            weight_loader = getattr(param, "weight_loader", None)
+            buffer = torch.zeros_like(param.data)
+            if weight_loader is None:
+                # Not a parallel layer: the delta is already rank-local.
+                buffer = torch.cat([d for d, _ in deltas]).to(device=param.device, dtype=param.dtype)
+            else:
+                saved = param.data
+                try:
+                    param.data = buffer
+                    for delta, shard_id in deltas:
+                        delta = delta.to(device=param.device, dtype=param.dtype)
+                        if shard_id is None:
+                            weight_loader(param, delta)
+                        else:
+                            weight_loader(param, delta, shard_id)
+                finally:
+                    param.data = saved
+            with torch.no_grad():
+                param.data.add_(buffer)
+            applied += 1
+        return applied
+
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
-        stacked_params_mapping = [
-            # More specific _mot_gen patterns FIRST to avoid substring
-            # ambiguity (e.g. `.q_proj` is a substring of `.q_proj_mot_gen`).
-            (".qkv_proj_mot_gen", ".q_proj_mot_gen", "q"),
-            (".qkv_proj_mot_gen", ".k_proj_mot_gen", "k"),
-            (".qkv_proj_mot_gen", ".v_proj_mot_gen", "v"),
-            (".qkv_proj", ".q_proj", "q"),
-            (".qkv_proj", ".k_proj", "k"),
-            (".qkv_proj", ".v_proj", "v"),
-            # MLP gate/up fused into MergedColumnParallelLinear
-            (".gate_up_proj", ".gate_proj", 0),
-            (".gate_up_proj", ".up_proj", 1),
-        ]
-        self.stacked_params_mapping = stacked_params_mapping
+        stacked_params_mapping = self.stacked_params_mapping
 
         params_dict = dict(self.named_parameters())
         loaded_params: set[str] = set()

vllm_omni/diffusion/models/sensenova_u1/sensenova_u1_transformer.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Qwen3 LLM with Mixture-of-Tokenizers (MoT) for SenseNova-U1.
 
 Ported from the sensenova_u1 package with vllm tensor-parallel support:
@@ -16,6 +16,7 @@ from dataclasses import dataclass
 
 import torch
 import torch.nn as nn
+import torch.nn.functional as F
 from cache_dit import ForwardPattern
 from transformers.cache_utils import DynamicCache
 from vllm.logger import init_logger
@@ -163,6 +164,18 @@ class Qwen3RotaryEmbedding(nn.Module):
         return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
 
 
+def _build_3d_rope(config) -> tuple[Qwen3RotaryEmbedding, Qwen3RotaryEmbedding]:
+    """The t and h/w rotary modules: t uses half head_dim, h/w a quarter each."""
+    t_config = copy.deepcopy(config)
+    t_config.head_dim = config.head_dim // 2
+
+    hw_config = copy.deepcopy(config)
+    hw_config.head_dim = config.head_dim // 4
+    hw_config.rope_theta = config.rope_theta_hw
+    hw_config.max_position_embeddings = config.max_position_embeddings_hw
+    return Qwen3RotaryEmbedding(config=t_config), Qwen3RotaryEmbedding(config=hw_config)
+
+
 # ---------------------------------------------------------------------------
 # Norms
 # ---------------------------------------------------------------------------
@@ -175,11 +188,9 @@ class Qwen3RMSNorm(nn.Module):
         self.variance_epsilon = eps
 
     def forward(self, hidden_states):
-        input_dtype = hidden_states.dtype
-        hidden_states = hidden_states.to(torch.float32)
-        variance = hidden_states.pow(2).mean(-1, keepdim=True)
-        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
-        return self.weight * hidden_states.to(input_dtype)
+        # F.rms_norm keeps the fp32 accumulation internal and rounds once; the
+        # cast chain it replaces rounded to bf16 before the weight multiply.
+        return F.rms_norm(hidden_states, self.weight.shape, self.weight, self.variance_epsilon)
 
 
 # ---------------------------------------------------------------------------
@@ -303,23 +314,16 @@ class SenseNovaU1Attention(nn.Module):
         self.q_norm_hw_mot_gen = Qwen3RMSNorm(self.head_dim // 2, eps=config.rms_norm_eps)
         self.k_norm_hw_mot_gen = Qwen3RMSNorm(self.head_dim // 2, eps=config.rms_norm_eps)
 
-        # 3D RoPE: t uses half head_dim, h/w each use quarter head_dim
-        t_config = copy.deepcopy(config)
-        t_config.head_dim = config.head_dim // 2
-        self.rotary_emb = Qwen3RotaryEmbedding(config=t_config)
-
-        hw_config = copy.deepcopy(config)
-        hw_config.head_dim = config.head_dim // 4
-        hw_config.rope_theta = config.rope_theta_hw
-        hw_config.max_position_embeddings = config.max_position_embeddings_hw
-        self.rotary_emb_hw = Qwen3RotaryEmbedding(config=hw_config)
-
         self.attn = Attention(
             num_heads=self.num_heads,
             head_size=self.head_dim,
             causal=False,
             softmax_scale=self.scaling,
-            num_kv_heads=self.num_heads,
+            # Keep K/V at their real head count. The SDPA backend already picks
+            # a fused GQA kernel when the runtime shape allows one and expands
+            # K/V itself when it does not; expanding here hides the GQA shape
+            # from that check, so it never fires.
+            num_kv_heads=self.num_kv_heads,
             prefix=f"{prefix}.attn",
         )
         self.attn.attention = self.attn.sdpa_fallback
@@ -339,10 +343,6 @@ class SenseNovaU1Attention(nn.Module):
         attention_mask: torch.Tensor | None,
     ) -> torch.Tensor:
         """Run unified attention with [B, H, S, D] inputs. Returns [B, S, H, D]."""
-        if self.num_kv_groups > 1:
-            n = self.num_kv_groups
-            key_bhsd = key_bhsd.repeat_interleave(n, dim=1)
-            value_bhsd = value_bhsd.repeat_interleave(n, dim=1)
         q = query_bhsd.transpose(1, 2).contiguous()
         k = key_bhsd.transpose(1, 2).contiguous()
         v = value_bhsd.transpose(1, 2).contiguous()
@@ -358,16 +358,16 @@ class SenseNovaU1Attention(nn.Module):
         attention_mask: torch.Tensor | None,
     ) -> torch.Tensor:
         """Run unified attention with [B, S, H, D] inputs. Returns [B, S, H, D]."""
-        if self.num_kv_groups > 1:
-            n = self.num_kv_groups
-            key_bshd = key_bshd.repeat_interleave(n, dim=2)
-            value_bshd = value_bshd.repeat_interleave(n, dim=2)
         attention_mask = self._align_mask_dtype(attention_mask, query_bshd)
         attn_metadata = AttentionMetadata(attn_mask=attention_mask) if attention_mask is not None else None
         return self.attn(query_bshd, key_bshd, value_bshd, attn_metadata)
 
-    def _project_and_rope(self, hidden_states, indexes, qkv_proj, q_norm, k_norm, q_norm_hw, k_norm_hw):
-        """Project Q/K/V via the given QKVParallelLinear and apply 3D RoPE."""
+    def _project_and_rope(self, hidden_states, position_embeddings, qkv_proj, q_norm, k_norm, q_norm_hw, k_norm_hw):
+        """Project Q/K/V via the given QKVParallelLinear and apply 3D RoPE.
+
+        The three tables come from the model, which builds them once per forward
+        because every layer is handed the same `indexes`.
+        """
         input_shape = hidden_states.shape[:-1]  # (B, S)
         qkv, _ = qkv_proj(hidden_states)
         q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
@@ -376,9 +376,7 @@ class SenseNovaU1Attention(nn.Module):
         k = k.view(*input_shape, self.num_kv_heads, self.head_dim)
         v = v.view(*input_shape, self.num_kv_heads, self.head_dim)
 
-        cos_t, sin_t = self.rotary_emb(hidden_states, indexes[0].unsqueeze(0))
-        cos_h, sin_h = self.rotary_emb_hw(hidden_states, indexes[1].unsqueeze(0))
-        cos_w, sin_w = self.rotary_emb_hw(hidden_states, indexes[2].unsqueeze(0))
+        (cos_t, sin_t), (cos_h, sin_h), (cos_w, sin_w) = position_embeddings
 
         value_states = v.transpose(1, 2)  # [B, H, S, D]
         try:
@@ -433,13 +431,23 @@ class SenseNovaU1Attention(nn.Module):
         input_shape = hidden_states.shape[:-1]
         query_states, key_states, value_states = self._project_and_rope(
             hidden_states,
-            indexes,
+            kwargs["position_embeddings"],
             self.qkv_proj,
             self.q_norm,
             self.k_norm,
             self.q_norm_hw,
             self.k_norm_hw,
         )
+        # Single-token decode with a paged cache: append this step's K/V and let
+        # the kernel read the valid prefix via `seqused_k`. Shapes stay fixed,
+        # which is what makes the step capturable as a CUDA graph.
+        paged = kwargs.get("paged_cache")
+        if paged is not None and attention_mask is None and query_states.shape[2] == 1:
+            attn_output = paged.attend(self.layer_idx, query_states, key_states, value_states, self.scaling)
+            attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+            attn_output, _ = self.o_proj(attn_output)
+            return attn_output
+
         update_cache = kwargs.get("update_cache", True)
         if past_key_values is not None:
             if update_cache:
@@ -460,7 +468,7 @@ class SenseNovaU1Attention(nn.Module):
         input_shape = hidden_states.shape[:-1]
         query_states, key_states, value_states = self._project_and_rope(
             hidden_states,
-            indexes,
+            kwargs["position_embeddings"],
             self.qkv_proj_mot_gen,
             self.q_norm_mot_gen,
             self.k_norm_mot_gen,
@@ -659,6 +667,10 @@ class SenseNovaU1Model(nn.Module):
         )
         self.norm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
         self.norm_mot_gen = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        # Every layer receives the same `indexes`, so the tables are the same for
+        # all of them: build them once per forward here and thread them down,
+        # rather than giving every layer its own pair.
+        self.rotary_emb, self.rotary_emb_hw = _build_3d_rope(config)
 
     def forward(
         self,
@@ -690,15 +702,25 @@ class SenseNovaU1Model(nn.Module):
             past_len = past_key_values.get_seq_length() if past_key_values else 0
             seq_len = inputs_embeds.shape[1]
             total_len = past_len + seq_len
-            mask = torch.zeros(1, 1, seq_len, total_len, device=inputs_embeds.device)
             if seq_len > 1:
+                mask = torch.zeros(1, 1, seq_len, total_len, device=inputs_embeds.device)
                 causal = torch.tril(torch.ones(seq_len, seq_len, device=inputs_embeds.device))
                 mask[:, :, :, past_len:] = torch.where(causal == 1, 0.0, float("-inf"))
+            else:
+                # A single query token attends to every cached key, so the mask
+                # this branch used to build was all zeros. Passing it changed no
+                # result, only which SDPA kernel was reachable.
+                mask = None
             causal_mask_mapping = {"full_attention": mask}
         else:
             causal_mask_mapping = attention_mask
 
         hidden_states = inputs_embeds
+        position_embeddings = (
+            self.rotary_emb(hidden_states, indexes[0].unsqueeze(0)),
+            self.rotary_emb_hw(hidden_states, indexes[1].unsqueeze(0)),
+            self.rotary_emb_hw(hidden_states, indexes[2].unsqueeze(0)),
+        )
         for layer in self.layers:
             hidden_states = layer(
                 hidden_states,
@@ -706,6 +728,7 @@ class SenseNovaU1Model(nn.Module):
                 exist_und=exist_und,
                 exist_gen=exist_gen,
                 indexes=indexes,
+                position_embeddings=position_embeddings,
                 attention_mask=causal_mask_mapping,
                 past_key_values=past_key_values,
                 use_cache=use_cache,

vllm_omni/diffusion/worker/diffusion_model_runner.py

@@ -519,6 +519,23 @@ class DiffusionModelRunner(OmniConnectorModelRunnerMixin):
         if self.prompt_embed_cache is not None:
             self.prompt_embed_cache.clear()
 
+    def release_captured_graphs(self) -> None:
+        """Drop every CUDA graph held for this model, wherever it is kept.
+
+        Sleep level 2 discards the memory a capture was recorded against, so a
+        graph that outlives it replays over freed storage. The runner owns
+        compilation and execution resources, so it owns the release: a pipeline
+        that captures graphs of its own implements ``release_captured_graphs``
+        and is collected here, instead of every caller having to know which
+        pipelines have one.
+        """
+        runners = getattr(self, "graph_runners", None)
+        if runners is not None:
+            runners.clear()
+        release = getattr(self.pipeline, "release_captured_graphs", None)
+        if callable(release):
+            release()
+
     def get_prompt_embed_cache_stats(self) -> dict | None:
         """Return hit/miss statistics for the prompt-embedding cache, if enabled.
 

vllm_omni/diffusion/worker/diffusion_worker.py

@@ -786,9 +786,8 @@ class DiffusionWorker:
         usage_before = allocator.get_current_usage()
 
         if level == 2 and self.model_runner is not None:
-            if hasattr(self.model_runner, "graph_runners"):
-                self.model_runner.graph_runners.clear()
-                logger.info(f"[Worker {self.rank}] CUDA Graphs cleared.")
+            self.model_runner.release_captured_graphs()
+            logger.info(f"[Worker {self.rank}] CUDA Graphs cleared.")
             model = self.model_runner.pipeline
             self._sleep_saved_buffers = {name: buffer.cpu().clone() for name, buffer in model.named_buffers()}