[Feat] SenseNova-U1: CFG parallelism on two GPUs (1.95x, bit-identical)

docs/cookbook/diffusion/SenseNova/SenseNova-U1.5-8B-MoT.mdx

@@ -105,7 +105,29 @@ in one batch must use the same resolution, inference steps, guidance settings,
 and output options. Requests with multiple outputs or `think_mode=true` are
 executed sequentially.
 
-### 5.2 Ascend NPU optimizations
+### 5.2 CFG parallelism on two GPUs
+
+With classifier-free guidance enabled (the default `guidance_scale` is 4.0),
+every denoising step runs a conditional and an unconditional forward pass.
+`--enable-cfg-parallel` places one branch on each of two GPUs, so each GPU runs
+one forward pass per step and the two exchange the predicted velocity. The
+image is bit-identical to the single-GPU result.
+
+```bash Command
+sglang serve \
+  --model-path sensenova/SenseNova-U1.5-8B-MoT \
+  --port 30000 \
+  --num-gpus 2 \
+  --enable-cfg-parallel
+```
+
+Each GPU holds a full copy of the weights. Requests with `guidance_scale` 1 have
+a single branch, which both GPUs compute. SenseNova-U1 does not support tensor,
+sequence or data parallelism yet, so `--num-gpus 2` requires
+`--enable-cfg-parallel`. CFG parallelism works with `think_mode`, but not yet
+with dynamic request batching (`--batching-max-size` greater than 1).
+
+### 5.3 Ascend NPU optimizations
 
 On Ascend NPU, SenseNova-U1 automatically uses fused inference attention,
 RMSNorm, and SwiGLU MLP operators for supported inputs. FIA and SwiGLU fall back
@@ -172,6 +194,21 @@ first use.
     ------------------------------------------------------------
     ```
   </Tab>
+  <Tab title="NVIDIA H200 CFG parallel">
+
+    Environment: two NVIDIA H200 GPUs connected by NVLink; workload: the
+    Section 3.1 `sglang generate` command (2048 x 2048, 50 denoising steps,
+    CFG 4, seed 42) with `--warmup-mode request`; repetitions: 2 per row, both
+    within 0.03 s; latency: the warmed-up request time.
+
+    | GPUs | Extra flags | Latency (s) | Speedup | Peak reserved memory on GPU 0 (MB) |
+    |---:|---|---:|---:|---:|
+    | 1 | | 16.26 | 1.00x | 36446 |
+    | 2 | `--num-gpus 2 --enable-cfg-parallel` | 8.32 | 1.95x | 35606 |
+
+    Both configurations produce byte-identical PNG files.
+
+  </Tab>
   <Tab title="Ascend 910C">
 
     Environment: one Ascend 910C; workload: four concurrent 2048 x 2048

docs/src/snippets/configs/SenseNova/sensenova-u15.jsx

@@ -5,12 +5,18 @@ export const config = (() => {
     })[s.checkpoint || "default"];
   const plan = (s) => {
     const flags = ["--model-path " + modelOf(s), ...["--port {{PORT}}"]];
-    let gpus = 1;
+    const gpus = Number(s.gpus_per_node);
     let t = { tp_size: 1, ulysses_degree: 1, ring_degree: 1 };
     const env = [];
     const warnings = [];
     const errors = [];
-    if (s.batching === "on") flags.push("--batching-max-size 2", "--batching-delay-ms 100");
+    if (gpus === 2) {
+      flags.push("--num-gpus 2", "--enable-cfg-parallel");
+      warnings.push(
+        "CFG parallelism runs one guidance branch per GPU; requests with guidance_scale 1 compute the same branch on both GPUs.",
+      );
+      if (s.batching === "on") errors.push("Dynamic batching does not support CFG parallelism yet; use one GPU.");
+    } else if (s.batching === "on") flags.push("--batching-max-size 2", "--batching-delay-ms 100");
     return { flags, gpus, topology: t, env, warnings, errors };
   };
   const defaults = {
@@ -44,7 +50,7 @@ export const config = (() => {
     commandBuilder: {
       defaultSelection: defaults,
       resource: {
-        limits: { nodes: { min: 1, max: 1 }, gpus_per_node: { min: 1, max: 1 } },
+        limits: { nodes: { min: 1, max: 1 }, gpus_per_node: { min: 1, max: 2 } },
         verifiedRecipes: ["cuda"].map((hw) => {
           const s = { ...defaults, hw };
           const p = plan(s);
@@ -64,7 +70,7 @@ export const config = (() => {
           if (!config.supportedHardware.includes(s.hw)) errors.push("Choose a listed hardware platform.");
           if (Number(s.nodes) !== 1) errors.push("This recipe covers one node.");
           const count = Number(s.gpus_per_node);
-          if (!Number.isInteger(count) || count < 1 || count > 1)
+          if (!Number.isInteger(count) || count < 1 || count > 2)
             errors.push("Choose an integer GPU count within the documented limits.");
           if (Number(s.gpus_per_node) !== p.gpus) errors.push("This recipe requires " + p.gpus + " GPUs / node.");
           if (["tp_size", "ulysses_degree", "ring_degree"].some((key) => Number(t[key]) !== p.topology[key]))

python/sglang/multimodal_gen/configs/pipeline_configs/sensenova_u1.py

@@ -46,6 +46,32 @@ def _component_residency_requests_offload(value) -> bool:
     return False
 
 
+def validate_sensenova_u1_parallelism(server_args) -> None:
+    # ServerArgs rejects CFG parallel on one GPU, so one GPU needs no more checks.
+    if server_args.num_gpus == 1:
+        return
+    if not (
+        server_args.num_gpus == 2
+        and server_args.enable_cfg_parallel
+        and server_args.cfg_parallel_degree == 2
+        and server_args.tp_size == 1
+        and server_args.sp_degree == 1
+        and server_args.dp_size == 1
+    ):
+        raise ValueError(
+            "SenseNovaU1Pipeline supports num_gpus=1, or num_gpus=2 with "
+            "--enable-cfg-parallel (one CFG branch per GPU). Native tensor, "
+            "sequence and data parallelism are not implemented yet."
+        )
+    # Each rank times the batching delay on its own clock, so the two CFG
+    # ranks can dispatch different batches and deadlock in a collective.
+    if server_args.batching_max_size > 1:
+        raise ValueError(
+            "SenseNovaU1Pipeline does not support dynamic batching "
+            "(--batching-max-size > 1) with --enable-cfg-parallel yet."
+        )
+
+
 def _set_compatible_runtime_defaults(server_args) -> None:
     compatible_defaults = {
         "component_residency": None,
@@ -80,7 +106,6 @@ class SenseNovaU1PipelineConfig(PipelineConfig):
     task_type: ModelTaskType = ModelTaskType.T2I
     model_precision: str = "bf16"
     should_use_guidance: bool = True
-    supports_cfg_parallel: bool = False
 
     def supports_dynamic_batching(self):
         return True
@@ -108,11 +133,7 @@ def supports_sequential_multi_output_inference(self):
         return True
 
     def validate_server_args(self, server_args) -> None:
-        if server_args.num_gpus != 1:
-            raise ValueError(
-                "SenseNovaU1Pipeline currently supports num_gpus=1. "
-                "Native tensor/pipeline parallelism is not implemented yet."
-            )
+        validate_sensenova_u1_parallelism(server_args)
         if getattr(server_args, "enable_torch_compile", False):
             raise ValueError(
                 "SenseNovaU1Pipeline does not support torch.compile yet. "
@@ -193,7 +214,6 @@ def get_model_deployment_config(self) -> ModelDeploymentConfig:
             speed_mode_enable_torch_compile_by_default=False,
             keep_resident_min_available_gb=80,
             auto_enable_cfg_parallel=False,
-            supports_cfg_parallel=False,
         )
 
 

python/sglang/multimodal_gen/configs/sample/sensenova_u1.py

@@ -68,6 +68,11 @@ def get_cli_args(cls, args):
             if key in _PUBLIC_OVERRIDE_FIELDS
         }
 
+    def should_use_classifier_free_guidance(self) -> bool:
+        # The model builds its own empty-prompt unconditional branch, so CFG
+        # depends on the guidance scale alone, not on negative_prompt.
+        return self.guidance_scale > 1.0
+
     def __post_init__(self) -> None:
         if isinstance(self.cfg_interval, list):
             self.cfg_interval = tuple(float(x) for x in self.cfg_interval)