feat(api): expose sampler_mode and 4 DiT params on /release_task HTTP API

Five GenerationParams fields that were only accessible via Gradio's
direct Python path are now accepted by the /release_task HTTP endpoint:

- sampler_mode: "euler" (default) or "heun" sampler selection
- velocity_norm_threshold: velocity prediction norm clamping (0=off)
- velocity_ema_factor: velocity EMA smoothing (0=off)
- latent_shift: additive shift on DiT latents before VAE decode
- latent_rescale: multiplicative rescale on DiT latents before VAE decode

All five already exist on the internal GenerationParams dataclass and
are wired through the diffusion loop. This change adds them to:
- GenerateMusicRequest (Pydantic request model)
- PARAM_ALIASES (camelCase alias support)
- build_generate_music_request (request builder)
- build_generation_setup (GenerationParams wiring)

Default values match the existing GenerationParams defaults, so
omitting them preserves current behavior. Non-target code paths
(Gradio UI, training, CLI) are unchanged.

Author: FlexOr2

acestep/api/http/release_task_models.py

@@ -97,6 +97,26 @@ class GenerateMusicRequest(BaseModel):
         default=None,
         description="Custom timesteps (comma-separated, e.g., '0.97,0.76,0.615,0.5,0.395,0.28,0.18,0.085,0'). Overrides inference_steps and shift.",
     )
+    sampler_mode: str = Field(
+        default="euler",
+        description="Diffusion sampler mode: 'euler' (first-order) or 'heun' (second-order predictor-corrector).",
+    )
+    velocity_norm_threshold: float = Field(
+        default=0.0,
+        description="Clamp velocity prediction norms during diffusion (0 = disabled, try 2.0).",
+    )
+    velocity_ema_factor: float = Field(
+        default=0.0,
+        description="Velocity EMA smoothing during diffusion (0 = disabled, try 0.1).",
+    )
+    latent_shift: float = Field(
+        default=0.0,
+        description="Additive shift applied to DiT latents before VAE decode (0 = no shift).",
+    )
+    latent_rescale: float = Field(
+        default=1.0,
+        description="Multiplicative rescale applied to DiT latents before VAE decode (1.0 = no rescale).",
+    )
 
     audio_format: str = Field(
         default="mp3",

acestep/api/http/release_task_param_parser.py

@@ -43,6 +43,11 @@
     "instruction": ["instruction"],
     "track_name": ["track_name", "trackName"],
     "track_classes": ["track_classes", "trackClasses", "instruments"],
+    "sampler_mode": ["sampler_mode", "samplerMode"],
+    "velocity_norm_threshold": ["velocity_norm_threshold", "velocityNormThreshold"],
+    "velocity_ema_factor": ["velocity_ema_factor", "velocityEmaFactor"],
+    "latent_shift": ["latent_shift", "latentShift"],
+    "latent_rescale": ["latent_rescale", "latentRescale"],
 }
 
 

acestep/api/http/release_task_request_builder.py

@@ -81,6 +81,11 @@ def build_generate_music_request(
         cfg_interval_end=parser.float("cfg_interval_end", 1.0),
         infer_method=parser.str("infer_method", "ode"),
         shift=parser.float("shift", 3.0),
+        sampler_mode=parser.str("sampler_mode", "euler"),
+        velocity_norm_threshold=parser.float("velocity_norm_threshold", 0.0),
+        velocity_ema_factor=parser.float("velocity_ema_factor", 0.0),
+        latent_shift=parser.float("latent_shift", 0.0),
+        latent_rescale=parser.float("latent_rescale", 1.0),
         audio_format=parser.str("audio_format", "mp3"),
         use_tiled_decode=parser.bool("use_tiled_decode", True),
         lm_model_path=parser.str("lm_model_path") or None,

acestep/api/job_generation_setup.py

@@ -174,6 +174,11 @@ def build_generation_setup(
         cfg_interval_end=req.cfg_interval_end,
         shift=req.shift,
         infer_method=req.infer_method,
+        sampler_mode=getattr(req, "sampler_mode", "euler"),
+        velocity_norm_threshold=getattr(req, "velocity_norm_threshold", 0.0),
+        velocity_ema_factor=getattr(req, "velocity_ema_factor", 0.0),
+        latent_shift=getattr(req, "latent_shift", 0.0),
+        latent_rescale=getattr(req, "latent_rescale", 1.0),
         timesteps=parsed_timesteps,
         repainting_start=req.repainting_start,
         repainting_end=req.repainting_end if req.repainting_end else -1,