feat(sensenova): add vLLM-Omni rollout

.pre-commit-config.yaml

@@ -10,9 +10,11 @@ default_stages: [pre-commit, pre-push, manual]
 #     hooks must not rewrite it either.
 #   - unirl/models/boogu_image/vendor/ : pristine boogu-project/Boogu-Image model code
 #     (see VENDOR_COMMIT.txt); same re-vendor rationale as bagel/vendor/.
+#   - unirl/models/sensenova_u1/vendor/ : pristine OpenSenseNova/SenseNova-U1
+#     NEO-Unify model code (see VENDOR_COMMIT.txt).
 #   - benchmarks/*/*/data/ : vendored upstream benchmark prompt sets (GenEval,
 #     PartiPrompts, VBench, ...); kept byte-identical so upstream checksums verify.
-exclude: ^(unirl-reward-service/|unirl/models/bagel/vendor/|unirl/models/boogu_image/vendor/|benchmarks/[^/]+/[^/]+/data/)
+exclude: ^(unirl-reward-service/|unirl/models/bagel/vendor/|unirl/models/boogu_image/vendor/|unirl/models/sensenova_u1/vendor/|benchmarks/[^/]+/[^/]+/data/)
 
 repos:
   - repo: https://github.com/pre-commit/pre-commit-hooks

INSTALL.md

@@ -4,7 +4,7 @@ UniRL ships two mutually exclusive inference engines (`vllm` and `sglang`) — i
 
 | Engine | CUDA | glibc |
 |---|---|---|
-| **vllm-omni** | 12.9 | ≥ 2.28 |
+| **vllm-omni** | 12.9 | ≥ 2.31 |
 | **sglang** | 13.0 | ≥ 2.34 |
 
 ## vllm-omni
@@ -15,6 +15,11 @@ export VLLM_USE_PRECOMPILED=1   # else 30+ min CUDA build
 uv pip install -e ".[vllm,train,infer]"
 ```
 
+The vLLM 0.22 CUDA wheel itself is `manylinux_2_28`, but its required
+`llguidance` wheel is `manylinux_2_31`. On a glibc 2.28 image, use a newer
+runtime/container or build `llguidance` from source with Rust; otherwise the
+resolver falls back to an unsupported or stalled source build.
+
 ## sglang
 
 ```bash

README.md

@@ -90,6 +90,7 @@ dimension; all listed models are supported (✅).
 | Prompt-Enhancer | LLM + diffusion | Text → Text → Image | ✅ |
 | HunyuanImage3 | Unified AR + diffusion | Text → Image | ✅ |
 | Bagel | Unified AR + diffusion | Text / Text + Image → Image | ✅ |
+| SenseNova-U1.5 | Unified MoT pixel flow | Text → Image | ✅ |
 
 </div>
 

examples/README.md

@@ -15,7 +15,7 @@ entrypoint's built-in `config_name` — a safe place to start.
 
 | Domain | Entrypoint | Default recipe (start here) | Models |
 |---|---|---|---|
-| [`diffusion/`](diffusion/) | `python -m unirl.train_diffusion` | `diffusion/sd3/sd3_trainside` | `sd3`, `qwen_image`, `flux2_klein`, `wan21`, `wan22`, `hunyuan_video10`, `hunyuan_video15` |
+| [`diffusion/`](diffusion/) | `python -m unirl.train_diffusion` | `diffusion/sd3/sd3_trainside` | `sd3`, `qwen_image`, `flux2_klein`, `sensenova_u1_5`, `wan21`, `wan22`, `hunyuan_video10`, `hunyuan_video15` |
 | [`ar/`](ar/) | `python -m unirl.train_ar` | `ar/qwen_vl_grpo_geo3k_mc_4x8`, `ar/qwen3_drpo_4b_base_dapo_sglang` | `qwen_vl` (vision-language), `qwen3` (text-only) |
 | [`sft/`](sft/) | `python -m unirl.train_sft` | `sft/qwen3_sft` | `qwen3`, `qwen_vl`, `bagel`, `sd3`, `cosmos3`, `wan21` |
 | [`pe/`](pe/) | `python -m unirl.train_pe` | `pe/pe_trainside_pickscore` | `pe` (Qwen3 rewriter + SD3, PickScore/WISE reward) |
@@ -95,7 +95,7 @@ related recipes sort together.
 
 | Segment | Position | Values (examples) | Omit when |
 |---|---|---|---|
-| `model` | required, first | `sd3`, `qwen_image`, `flux2_klein`, `wan21`, `wan22`, `hunyuan_video10`, `hunyuan_video15`, `qwen_vl`, `qwen3`, `hi3` | never |
+| `model` | required, first | `sd3`, `qwen_image`, `flux2_klein`, `sensenova_u1_5`, `wan21`, `wan22`, `hunyuan_video10`, `hunyuan_video15`, `qwen_vl`, `qwen3`, `hi3` | never |
 | `task` | after model | `t2v`, `i2v` | text-to-image (the implicit default) |
 | `size` | after task | `4b`, `14b` | only one size in the family |
 | `algorithm` | middle | `dancegrpo`, `mixgrpo`, `nft`, `flowdppo`, `grpo`, `drpo` | plain FlowGRPO (diffusion default); GRPO (AR default) |

examples/diffusion/sensenova_u1_5/sensenova_u1_5_trainside.yaml

@@ -0,0 +1,137 @@
+# @package _global_
+# SenseNova-U1.5-8B-MoT-Preview pixel-flow GRPO with trainside rollout.
+#
+# Launch:
+#   SENSENOVA_U1_PATH=/path/to/SenseNova-U1.5-8B-MoT-Preview \
+#   python -m unirl.train_diffusion --config-name diffusion/sensenova_u1_5/sensenova_u1_5_trainside
+
+num_devices: 8
+batch_size: 8
+adv_use_global_std: true
+num_rollouts: 10000
+
+logging:
+  report_to_wandb: false
+  project_name: unirl
+  run_name: sensenova-u1.5-grpo
+  entity: null
+  tags: [sensenova-u1.5, pixel-flow, trainside, fullft]
+
+bundle:
+  _target_: unirl.models.sensenova_u1.bundle.SenseNovaU1Bundle.from_config
+  config:
+    _target_: unirl.models.sensenova_u1.config.SenseNovaU1PipelineConfig
+    pretrained_model_ckpt_path: ${oc.env:SENSENOVA_U1_PATH,sensenova/SenseNova-U1.5-8B-MoT-Preview}
+    # Qwen3DecoderLayer interleaves frozen und and trainable gen parameters.
+    # FSDP2 requires one original dtype per shard group, so keep the complete
+    # model in fp32 masters and use bf16 only through mixed-precision compute.
+    model_precision: fp32
+    autocast_precision: bf16
+    trajectory_precision: bf16
+    logprob_precision: fp32
+    timestep_shift: 3.0
+    attention_backend: auto
+    # Matches the released U1.5 pretraining recipe: freeze the shared /
+    # understanding path, train fm_modules + every *_mot_gen branch.
+    full_finetune_generation: true
+
+pipeline:
+  _target_: unirl.models.sensenova_u1.pipeline.SenseNovaU1Pipeline
+  shift: ${bundle.config.timestep_shift}
+  autocast_precision: ${bundle.config.autocast_precision}
+  trajectory_precision: ${bundle.config.trajectory_precision}
+  logprob_precision: ${bundle.config.logprob_precision}
+  strategy:
+    _target_: unirl.sde.kernels.FlowSDEStrategy
+
+backend:
+  _target_: unirl.train.backend.fsdp.FSDPBackend
+  block_class_names: ["Qwen3DecoderLayer", "NEOVisionModel"]
+  trainable_attr: transformer
+  fsdp_cfg:
+    _target_: unirl.train.configs.FSDPConfig
+    param_dtype: bf16
+    master_dtype: fp32
+    cpu_offload: false
+    mixed_precision: true
+    fsdp_mode: full
+    reshard_after_forward: true
+    activation_checkpointing: true
+    use_torch_compile: false
+    # The trainable fm_modules tree is outside the decoder blocks.
+    root_wrap: true
+  optimizer_cfg:
+    _target_: unirl.train.backend.base.OptimizerConfig
+    learning_rate: 2.0e-5
+    adam_beta1: 0.9
+    adam_beta2: 0.95
+    adam_epsilon: 1.0e-8
+    weight_decay: 0.0
+  scheduler_cfg:
+    _target_: unirl.train.backend.base.LrSchedulerConfig
+    type: constant
+    warmup_steps: 0
+    total_steps: ${num_rollouts}
+
+rollout:
+  _target_: unirl.rollout.engine.trainside.engine.TrainsideRolloutEngine
+  stage_attrs: [diffusion]
+  # Prefix KV caches and pixel-flow trajectories are intentionally per-sample.
+  forward_batch_size: 1
+
+reward:
+  _target_: unirl.reward.service.RewardService
+  backend:
+    _target_: unirl.reward.local.pickscore.PickScoreRewardScorer
+    base_device: cuda
+    config:
+      _target_: unirl.reward.local.pickscore.PickScoreSpec
+      batch_size: 4
+      device: auto
+      processor_id: laion/CLIP-ViT-H-14-laion2B-s32B-b79K
+      model_id: yuvalkirstain/PickScore_v1
+
+algorithm:
+  _target_: unirl.algorithms.flowgrpo.FlowGRPO
+  stage_attr: diffusion
+  clip_range: 1.0e-4
+  clip_schedule: constant
+  conditions_cls:
+    _target_: hydra.utils.get_class
+    path: unirl.models.sensenova_u1.conditions.SenseNovaU1Conditions
+  params: ${sampling}
+
+stack:
+  _target_: unirl.train.stack.TrainStack
+  micro_batch_size: 1
+  max_grad_norm: 1.0
+  num_updates_per_batch: 2
+
+data_source:
+  _target_: unirl.data.data_source.MultimodalRLDataSource
+  args:
+    run:
+      data_path: datasets/pickscore/train.txt
+      eval_data_path: datasets/pickscore/test.txt
+      seed: 42
+    algorithm:
+      prompts_per_rollout: ${batch_size}
+
+sampling:
+  _target_: unirl.models.sensenova_u1.diffusion.SenseNovaU1DiffusionParams
+  num_inference_steps: 20
+  guidance_scale: 4.0
+  cfg_norm: global
+  cfg_interval: [0.0, 1.0]
+  t_eps: 0.02
+  height: 512
+  width: 512
+  eta: 0.7
+  samples_per_prompt: 4
+  seed: 42
+  init_same_noise: false
+  scheduler:
+    _target_: unirl.sde.index_schedule.AllSDEScheduler
+    num_timesteps: ${..num_inference_steps}
+    num_sde_steps: 3
+    timestep_fraction: [0.0, 0.5]

examples/diffusion/sensenova_u1_5/sensenova_u1_5_vllmomni_full_ipc.yaml

@@ -0,0 +1,148 @@
+# @package _global_
+# SenseNova-U1.5 pixel-flow GRPO with colocated vLLM-Omni rollout and
+# bucketed CUDA-IPC full-weight synchronization.
+#
+# Launch:
+#   SENSENOVA_U1_PATH=/path/to/SenseNova-U1.5-8B-MoT-Preview \
+#   python -m unirl.train_diffusion \
+#     --config-name diffusion/sensenova_u1_5/sensenova_u1_5_vllmomni_full_ipc
+
+num_devices: 8
+batch_size: 8
+adv_use_global_std: true
+num_rollouts: 10000
+weight_sync_interval: 1
+
+logging:
+  report_to_wandb: false
+  project_name: unirl
+  run_name: sensenova-u1.5-grpo-vllmomni
+  entity: null
+  tags: [sensenova-u1.5, pixel-flow, vllm-omni, fullft, ipc]
+
+bundle:
+  _target_: unirl.models.sensenova_u1.bundle.SenseNovaU1Bundle.from_config
+  config:
+    _target_: unirl.models.sensenova_u1.config.SenseNovaU1PipelineConfig
+    pretrained_model_ckpt_path: ${oc.env:SENSENOVA_U1_PATH,sensenova/SenseNova-U1.5-8B-MoT-Preview}
+    model_precision: fp32
+    autocast_precision: bf16
+    trajectory_precision: bf16
+    logprob_precision: fp32
+    timestep_shift: 3.0
+    attention_backend: auto
+    full_finetune_generation: true
+
+pipeline:
+  _target_: unirl.models.sensenova_u1.pipeline.SenseNovaU1Pipeline
+  shift: ${bundle.config.timestep_shift}
+  autocast_precision: ${bundle.config.autocast_precision}
+  trajectory_precision: ${bundle.config.trajectory_precision}
+  logprob_precision: ${bundle.config.logprob_precision}
+  strategy:
+    _target_: unirl.sde.kernels.FlowSDEStrategy
+
+backend:
+  _target_: unirl.train.backend.fsdp.FSDPBackend
+  block_class_names: ["Qwen3DecoderLayer", "NEOVisionModel"]
+  trainable_attr: transformer
+  fsdp_cfg:
+    _target_: unirl.train.configs.FSDPConfig
+    param_dtype: bf16
+    master_dtype: fp32
+    cpu_offload: false
+    mixed_precision: true
+    fsdp_mode: full
+    reshard_after_forward: true
+    activation_checkpointing: true
+    use_torch_compile: false
+    root_wrap: true
+  optimizer_cfg:
+    _target_: unirl.train.backend.base.OptimizerConfig
+    learning_rate: 2.0e-5
+    adam_beta1: 0.9
+    adam_beta2: 0.95
+    adam_epsilon: 1.0e-8
+    weight_decay: 0.0
+  scheduler_cfg:
+    _target_: unirl.train.backend.base.LrSchedulerConfig
+    type: constant
+    warmup_steps: 0
+    total_steps: ${num_rollouts}
+
+rollout:
+  _target_: unirl.rollout.engine.vllm_omni.engine.VLLMOmniRolloutEngine
+  model_config: ${bundle.config}
+  config:
+    _target_: unirl.rollout.engine.vllm_omni.config.VLLMOmniEngineConfig
+    model_path: ${bundle.config.pretrained_model_ckpt_path}
+    modality: sensenova_u1_t2i
+    enable_sleep_mode: true
+
+reward:
+  _target_: unirl.reward.service.RewardService
+  backend:
+    _target_: unirl.reward.local.pickscore.PickScoreRewardScorer
+    base_device: cuda
+    config:
+      _target_: unirl.reward.local.pickscore.PickScoreSpec
+      batch_size: 4
+      device: auto
+      processor_id: laion/CLIP-ViT-H-14-laion2B-s32B-b79K
+      model_id: yuvalkirstain/PickScore_v1
+
+algorithm:
+  _target_: unirl.algorithms.flowgrpo.FlowGRPO
+  stage_attr: diffusion
+  old_logp_source: replay
+  clip_range: 1.0e-4
+  clip_schedule: constant
+  conditions_cls:
+    _target_: hydra.utils.get_class
+    path: unirl.models.sensenova_u1.conditions.SenseNovaU1Conditions
+  params: ${sampling}
+
+stack:
+  _target_: unirl.train.stack.TrainStack
+  micro_batch_size: 1
+  max_grad_norm: 1.0
+  num_updates_per_batch: 2
+
+data_source:
+  _target_: unirl.data.data_source.MultimodalRLDataSource
+  args:
+    run:
+      data_path: datasets/pickscore/train.txt
+      eval_data_path: datasets/pickscore/test.txt
+      seed: 42
+    algorithm:
+      prompts_per_rollout: ${batch_size}
+
+sampling:
+  _target_: unirl.models.sensenova_u1.diffusion.SenseNovaU1DiffusionParams
+  num_inference_steps: 20
+  guidance_scale: 4.0
+  cfg_norm: global
+  cfg_interval: [0.0, 1.0]
+  t_eps: 0.02
+  height: 512
+  width: 512
+  eta: 0.7
+  samples_per_prompt: 4
+  seed: 42
+  init_same_noise: false
+  scheduler:
+    _target_: unirl.sde.index_schedule.AllSDEScheduler
+    num_timesteps: ${..num_inference_steps}
+    num_sde_steps: 3
+    timestep_fraction: [0.0, 0.5]
+
+sync:
+  _target_: unirl.distributed.weight_sync.full.ipc.IPCWeightSync
+  bucket_size_mb: 2048
+  use_shm: false
+  flush_cache: true
+  wire_dtype: bf16
+  # Strip SenseNovaU1TrainableModel's facade prefix. vLLM-Omni exposes the
+  # same children directly as language_model.*, vision_model.*, fm_modules.*.
+  name_remap: {"model.*": "*"}

examples/run_reward_curve_verification_taiji.sh

@@ -224,8 +224,8 @@ else
     require_torch_flavor "2.11.0+cu129"
 fi
 if [ "${PROFILE}" = "sd3-vllm-omni" ] || [ "${PROFILE}" = "qwen-omni" ]; then
-    require_dist_version "vllm" "0.20.0"
-    require_dist_version "vllm-omni" "0.20.0"
+    require_dist_version "vllm" "0.22.0+cu129"
+    require_dist_version "vllm-omni" "0.22.0"
 fi
 
 WANDB_OVERRIDES=(