[Perf][Diffusion] Add configurable H3 VAE stacked tiling

Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>

docs/user_guide/diffusion/parallelism/vae_parallelism.md

@@ -120,10 +120,18 @@ Additional requirements:
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
 | `vae_use_tiling` | bool | False | Must be set to `True` when using VAE patch parallelism. |
+| `vae_stack_tiling` | `"auto"` \| `"true"` \| `"false"` | `"false"` | For VAEs with native support (currently MiniMax H3), batch local decoder tiles into fewer model calls. `"auto"` enables it when the request has at least two tiles per VAE-parallel rank. |
 
 !!! note "Automatic VAE Tiling"
     When `vae_patch_parallel_size > 1` and the model has a distributed VAE (`DistributedVaeMixin`), the system automatically sets `vae_use_tiling=True` if not already enabled.
 
+!!! tip "MiniMax H3 stacked tiling"
+    MiniMax H3 supports lossless native stacked tiling. Enable it with
+    `--vae-stack-tiling auto` (or `true`) to batch each rank's spatial tiles
+    into one decoder call per temporal window. `auto` keeps the sequential
+    path for shapes that do not provide multiple tiles per rank, while
+    `false` remains the default to avoid increasing the VAE activation peak.
+
 ---
 
 ## Spatially-Sharded Decode (Wan)

examples/offline_inference/qwen2_5_omni/end2end.py

@@ -21,7 +21,6 @@ from vllm.multimodal.media.audio import load_audio
 from vllm.sampling_params import SamplingParams
 
 from vllm_omni.entrypoints.omni import Omni
-from vllm_omni.utils.profiler import add_profiler_config_arg
 from vllm_omni.utils.tracking_parser import TrackingArgumentParser
 
 SEED = 42
@@ -326,7 +325,6 @@ def main(args):
     else:
         query_result = query_func()
     args.quantization_config = quantization_config
-    # ``vars(args)`` below forwards profiler_config to Omni.
     omni_kwargs = vars(args).copy()
     # Override CLI --model with the derived model_name.
     omni_kwargs["model"] = model_name
@@ -380,11 +378,11 @@ def main(args):
         for i, prompt in enumerate(prompts):
             prompt["modalities"] = output_modalities
 
-    profiler_enabled = args.profiler_config is not None
+    profiler_enabled = bool(os.getenv("VLLM_TORCH_PROFILER_DIR"))
     if profiler_enabled and hasattr(omni, "start_profile"):
         omni.start_profile(stages=[0])
     elif profiler_enabled:
-        print("[Warn] --profiler-config was provided, but current engine does not support profiler controls.")
+        print("[Warn] VLLM_TORCH_PROFILER_DIR is set, but current engine does not support profiler controls.")
     omni_generator = omni.generate(prompts, sampling_params_list, py_generator=args.py_generator)
 
     # Determine output directory: prefer --output-dir; fallback to --output-wav
@@ -460,7 +458,6 @@ def parse_args():
         default=False,
         help="Enable writing detailed statistics (default: disabled)",
     )
-    add_profiler_config_arg(parser)
     parser.add_argument(
         "--stage-init-timeout",
         type=int,

examples/offline_inference/text_to_speech/cosyvoice3/end2end.py

@@ -14,7 +14,6 @@ from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.model_executor.models.cosyvoice3.tokenizer import get_qwen_tokenizer
 from vllm_omni.model_executor.models.cosyvoice3.utils import extract_text_token
 from vllm_omni.transformers_utils.configs.cosyvoice3 import CosyVoice3Config
-from vllm_omni.utils.profiler import add_profiler_config_arg
 
 # Upstream zero-shot reference clip
 ZERO_SHOT_PROMPT_URL = "https://raw.githubusercontent.com/FunAudioLLM/CosyVoice/main/asset/zero_shot_prompt.wav"
@@ -46,7 +45,6 @@ def run_e2e():
         help="Override the deploy config path. If unset, auto-loads "
         "vllm_omni/deploy/cosyvoice3.yaml based on the HF model_type.",
     )
-    add_profiler_config_arg(parser)
     parser.add_argument("--text", type=str, default="Hello, this is a test of the CosyVoice system capability.")
     parser.add_argument(
         "--prompt-text",
@@ -81,7 +79,6 @@ def run_e2e():
         deploy_config=args.deploy_config,
         tokenizer=args.tokenizer,
         log_stats=True,
-        profiler_config=args.profiler_config,
     )
 
     sampling_cfg = {"top_p": 0.8, "top_k": 25, "eos_token_id": 6561 + 1}
@@ -152,8 +149,8 @@ def run_e2e():
 
     sampling_params_list = [gpt_sampling, s2mel_sampling]
 
-    profiler_enabled = args.profiler_config is not None
-    if profiler_enabled:
+    # Start profiling (requires VLLM_TORCH_PROFILER_DIR env var)
+    if os.environ.get("VLLM_TORCH_PROFILER_DIR"):
         print("Starting profiler...")
         omni.start_profile()
 
@@ -161,7 +158,7 @@ def run_e2e():
     outputs = list(omni.generate(prompts, sampling_params_list=sampling_params_list[:2]))
 
     # Stop profiling and get results
-    if profiler_enabled:
+    if os.environ.get("VLLM_TORCH_PROFILER_DIR"):
         print("Stopping profiler...")
         profile_results = omni.stop_profile()
         print(f"Profile traces saved to: {profile_results}")

tests/diffusion/model_loader/test_diffusers_loader.py

@@ -614,6 +614,54 @@ def test_initialize_model_sets_current_diffusion_config_during_model_constructio
     assert get_current_diffusion_config_or_none() is None
 
 
+def test_configure_vae_stack_tiling_enables_tiling_and_native_capability():
+    from vllm_omni.diffusion.registry import _configure_vae_stack_tiling
+
+    calls = []
+    vae = SimpleNamespace(
+        supports_stack_tiling=True,
+        set_stack_tiling=lambda mode: calls.append(mode),
+    )
+    model = SimpleNamespace(vae=vae)
+    od_config = SimpleNamespace(
+        model_class_name="MiniMaxH3Pipeline",
+        vae_stack_tiling="auto",
+        vae_use_tiling=False,
+    )
+
+    _configure_vae_stack_tiling(model, od_config)
+
+    assert od_config.vae_use_tiling is True
+    assert calls == ["auto"]
+
+
+def test_configure_vae_stack_tiling_rejects_unsupported_vae():
+    from vllm_omni.diffusion.registry import _configure_vae_stack_tiling
+
+    od_config = SimpleNamespace(
+        model_class_name="UnsupportedPipeline",
+        vae_stack_tiling="true",
+        vae_use_tiling=False,
+    )
+
+    with pytest.raises(ValueError, match="does not declare stacked-tiling support"):
+        _configure_vae_stack_tiling(SimpleNamespace(vae=SimpleNamespace()), od_config)
+
+
+def test_configure_vae_stack_tiling_auto_falls_back_for_unsupported_vae():
+    from vllm_omni.diffusion.registry import _configure_vae_stack_tiling
+
+    od_config = SimpleNamespace(
+        model_class_name="UnsupportedPipeline",
+        vae_stack_tiling="auto",
+        vae_use_tiling=False,
+    )
+
+    _configure_vae_stack_tiling(SimpleNamespace(vae=SimpleNamespace()), od_config)
+
+    assert od_config.vae_use_tiling is False
+
+
 def test_load_model_custom_pipeline_sets_current_diffusion_config(monkeypatch):
     import vllm_omni.diffusion.model_loader.diffusers_loader as loader_mod
 

tests/diffusion/models/minimax_h3/test_minimax_h3_vae_tiling.py

@@ -30,6 +30,7 @@ class _FakeCheckpointModel:
     def __init__(self):
         self.vae_ratio = RATIO
         self.parallel_tiling = True
+        self.stack_tiling = False
         self.decoded_with = []
 
     def split_tiles(self, input_len, is_decoder=False):
@@ -57,6 +58,7 @@ def _vae(parallel_size, state):
     vae.model = _FakeCheckpointModel()
     vae.remote = remote
     vae.parallel_size = parallel_size
+    vae.stack_tiling_mode = "false"
     return vae
 
 
@@ -89,6 +91,33 @@ def test_tile_count_uses_the_decoder_grid_not_the_encoder_one(h, w, decoder_tile
     assert as_encoder == encoder_tiles != decoder_tiles
 
 
+@pytest.mark.parametrize(
+    ("parallel_size", "h", "w", "expected"),
+    [
+        (1, 48, 84, True),
+        (2, 48, 84, True),
+        (4, 48, 84, True),
+        (8, 48, 84, True),
+        (16, 48, 84, False),
+        (4, 16, 16, False),
+    ],
+)
+def test_auto_stack_tiling_requires_multiple_tiles_per_rank(parallel_size, h, w, expected):
+    vae = _vae(parallel_size, {})
+    vae.set_stack_tiling("auto")
+
+    assert MiniMaxH3VideoVAE._should_stack_decode_tiles(vae, _latent(h, w)) is expected
+
+
+@pytest.mark.parametrize(("mode", "expected"), [(False, False), (True, True), ("false", False), ("true", True)])
+def test_stack_tiling_configuration_accepts_bool_and_string_modes(mode, expected):
+    vae = _vae(1, {})
+
+    vae.set_stack_tiling(mode)
+
+    assert vae.model.stack_tiling is expected
+
+
 def test_rank_local_tiling_restores_the_group_state():
     state = {"sp_size": 4, "sp_rank": 2, "sp_enabled": True, "sp_process_group": "pg"}
     vae = _vae(4, state)
@@ -156,6 +185,7 @@ def _dispatch_vae(parallel_size):
         seen["sp_enabled"] = state["sp_enabled"]
         seen["sp_process_group"] = state["sp_process_group"]
         seen["parallel_tiling"] = vae.model.parallel_tiling
+        seen["stack_tiling"] = vae.model.stack_tiling
         # the guard never inspects the decoded shape, so keep it small
         return torch.zeros(1, 3, 2, 4, 4)
 
@@ -212,3 +242,15 @@ def test_decode_latent_leaves_the_group_state_alone_when_the_grid_is_large_enoug
     assert seen["sp_enabled"] is True
     assert seen["sp_process_group"] is state["sp_process_group"]
     assert seen["parallel_tiling"] is True
+
+
+def test_decode_latent_restores_native_stack_tiling_state():
+    vae, _, seen = _dispatch_vae(4)
+    vae.set_stack_tiling("auto")
+    # The adapter must not leak its per-request decision into the remote VAE.
+    vae.model.stack_tiling = True
+
+    MiniMaxH3VideoVAE.decode_latent(vae, _latent(16, 16))
+
+    assert seen["stack_tiling"] is False
+    assert vae.model.stack_tiling is True

tests/docs/test_offline_profiler_config.py

@@ -1,23 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
-
-from pathlib import Path
-
-import pytest
-
-pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
-
-_REPO_ROOT = Path(__file__).resolve().parents[2]
-_EXAMPLES = (
-    "examples/offline_inference/text_to_speech/cosyvoice3/end2end.py",
-    "examples/offline_inference/qwen2_5_omni/end2end.py",
-)
-
-
-@pytest.mark.parametrize("relative_path", _EXAMPLES)
-def test_offline_examples_use_typed_profiler_config(relative_path: str):
-    source = (_REPO_ROOT / relative_path).read_text(encoding="utf-8")
-
-    assert "add_profiler_config_arg(parser)" in source
-    assert "profiler_enabled = args.profiler_config is not None" in source
-    assert "VLLM_TORCH_PROFILER_DIR" not in source

tests/entrypoints/test_async_omni_diffusion_config.py

@@ -395,6 +395,28 @@ def test_serve_cli_accepts_diffusion_pipeline_profiler_flag():
     assert stage_cfg["engine_args"]["enable_diffusion_pipeline_profiler"] is True
 
 
+def test_serve_cli_forwards_vae_stack_tiling():
+    parser = TrackingArgumentParser()
+    subparsers = parser.add_subparsers(dest="command")
+    OmniServeCommand().subparser_init(subparsers)
+
+    args = parser.parse_args(
+        [
+            "serve",
+            "MiniMaxAI/MiniMax-H3",
+            "--omni",
+            "--vae-stack-tiling",
+            "auto",
+        ]
+    )
+
+    explicit_kwargs = args.get_explicit_kwargs_dict()
+    stage_cfg = AsyncOmniEngine._create_default_diffusion_stage_cfg(explicit_kwargs)[0]
+
+    assert args.vae_stack_tiling == "auto"
+    assert stage_cfg["engine_args"]["vae_stack_tiling"] == "auto"
+
+
 def test_serve_cli_forwards_distilled_lora_to_diffusion_stage():
     """Ensure startup distilled LoRA options reach the online diffusion stage."""
     parser = TrackingArgumentParser()

vllm_omni/config/omni_config.py

@@ -711,6 +711,7 @@ class _DiffusionConfigProjection:
     fa_deterministic: bool = False
     vae_use_slicing: bool = False
     vae_use_tiling: bool = False
+    vae_stack_tiling: Literal["auto", "true", "false"] = "false"
     mask_strategy_file_path: str | None = None
     skip_time_steps: int = 15
     VSA_sparsity: float = 0.0

vllm_omni/config/stage_config.py

@@ -11,7 +11,7 @@ from collections.abc import Callable, Mapping
 from dataclasses import asdict, dataclass, field, fields
 from enum import Enum
 from pathlib import Path
-from typing import Any, NamedTuple
+from typing import Any, Literal, NamedTuple
 
 from transformers import PretrainedConfig
 from vllm.logger import init_logger
@@ -462,6 +462,7 @@ class StageDeployConfig:
     step_execution: bool | None = None
     vae_use_slicing: bool | None = None
     vae_use_tiling: bool | None = None
+    vae_stack_tiling: Literal["auto", "true", "false"] | None = None
     boundary_ratio: float | None = None
     flow_shift: float | None = None
     diffusion_kv_cache_dtype: str | None = None

vllm_omni/diffusion/data.py

@@ -8,7 +8,7 @@ import random
 from collections.abc import Callable, Mapping
 from dataclasses import dataclass, field, fields
 from enum import Enum
-from typing import TYPE_CHECKING, Any
+from typing import TYPE_CHECKING, Any, Literal
 
 import diffusers
 import huggingface_hub
@@ -831,6 +831,7 @@ class OmniDiffusionConfig:
     # VAE memory optimization parameters
     vae_use_slicing: bool = False
     vae_use_tiling: bool = False
+    vae_stack_tiling: Literal["auto", "true", "false"] = "false"
 
     # STA (Sliding Tile Attention) parameters
     mask_strategy_file_path: str | None = None

vllm_omni/diffusion/models/minimax_h3/vae.py

@@ -9,7 +9,7 @@ import json
 from collections.abc import Iterator
 from contextlib import AbstractContextManager, contextmanager, nullcontext
 from pathlib import Path
-from typing import Any
+from typing import Any, ClassVar
 
 import torch
 import torch.distributed as dist
@@ -115,6 +115,8 @@ class _AudioVAEDeterminismContext(AbstractContextManager):
 class MiniMaxH3VideoVAE(nn.Module, DistributedVaeMixin):
     """Adapter around the checkpoint's native parallel-tiled video VAE."""
 
+    supports_stack_tiling: ClassVar[bool] = True
+
     def __init__(
         self,
         component_path: str,
@@ -153,6 +155,30 @@ class MiniMaxH3VideoVAE(nn.Module, DistributedVaeMixin):
         self.use_slicing = False
         self.parallel_size = 1
         self.device_module = torch.get_device_module()
+        self.stack_tiling_mode = "false"
+
+    def set_stack_tiling(self, mode: str | bool) -> None:
+        """Configure native decoder tile batching.
+
+        ``auto`` is resolved per decode because the useful batch size depends
+        on both the request's tile grid and the VAE parallel group size.
+        """
+        if not hasattr(self.model, "stack_tiling"):
+            raise RuntimeError("MiniMax H3 video VAE remote model does not expose stack_tiling")
+        if isinstance(mode, bool):
+            mode = "true" if mode else "false"
+        if mode not in ("auto", "true", "false"):
+            raise ValueError(f"invalid VAE stacked-tiling mode: {mode!r}")
+        self.stack_tiling_mode = mode
+        self.model.stack_tiling = mode == "true"
+
+    def _should_stack_decode_tiles(self, latent: torch.Tensor) -> bool:
+        if self.stack_tiling_mode != "auto":
+            return self.stack_tiling_mode == "true"
+        # Stacking is worthwhile when every rank has more than one tile in
+        # the common round-robin distribution. The tile grid is global and
+        # therefore identical on all ranks.
+        return self._decoder_tile_count(latent) >= 2 * self.parallel_size
 
     def load_to_device(self) -> None:
         if self._stager is not None:
@@ -388,7 +414,18 @@ class MiniMaxH3VideoVAE(nn.Module, DistributedVaeMixin):
             tiling_context = nullcontext()
 
         with tiling_context:
-            decoded = self.model.decode_base(latent * std + mean)
+            stack_tiling_mode = getattr(self, "stack_tiling_mode", "false")
+            if stack_tiling_mode == "false" and not hasattr(self.model, "stack_tiling"):
+                # Keep the default path compatible with older H3 snapshots
+                # that predate the native stacked-tiling attribute.
+                decoded = self.model.decode_base(latent * std + mean)
+            else:
+                previous_stack_tiling = self.model.stack_tiling
+                self.model.stack_tiling = self._should_stack_decode_tiles(latent)
+                try:
+                    decoded = self.model.decode_base(latent * std + mean)
+                finally:
+                    self.model.stack_tiling = previous_stack_tiling
         frames = self.model.processor.revert_tensor(decoded)
         if frames.ndim == 4:
             frames = frames.unsqueeze(0).transpose(1, 2)

vllm_omni/diffusion/registry.py

@@ -387,6 +387,38 @@ def _prepare_diffusion_quant_config(
     configure_quant_config(quant_config, model_class)
 
 
+def _configure_vae_stack_tiling(model: nn.Module, od_config: OmniDiffusionConfig) -> None:
+    """Configure optional native batching of VAE decoder tiles."""
+    mode = getattr(od_config, "vae_stack_tiling", "false")
+    if isinstance(mode, bool):
+        mode = "true" if mode else "false"
+    if mode not in ("auto", "true", "false"):
+        raise ValueError(f"invalid vae_stack_tiling mode: {mode!r}")
+    if mode == "false":
+        return
+
+    vae = getattr(model, "vae", None)
+    if vae is None or not getattr(vae, "supports_stack_tiling", False):
+        if mode == "auto":
+            logger.info(
+                "VAE stacked tiling auto mode is unavailable for %s; using sequential tiles.",
+                od_config.model_class_name,
+            )
+            return
+        raise ValueError(
+            f"vae_stack_tiling was requested, but {od_config.model_class_name} does not declare stacked-tiling support"
+        )
+    setter = getattr(vae, "set_stack_tiling", None)
+    if not callable(setter):
+        raise ValueError("vae_stack_tiling was requested, but the VAE does not provide set_stack_tiling()")
+
+    if not od_config.vae_use_tiling:
+        logger.info("vae_stack_tiling requires vae_use_tiling; automatically enabling it.")
+        od_config.vae_use_tiling = True
+    setter(mode)
+    logger.info("VAE stacked tiling mode set to %s for %s.", mode, od_config.model_class_name)
+
+
 def initialize_model(
     od_config: OmniDiffusionConfig,
 ) -> nn.Module:
@@ -428,6 +460,8 @@ def initialize_model(
             )
             od_config.vae_use_tiling = True
 
+        _configure_vae_stack_tiling(model, od_config)
+
         # Configure VAE memory optimization settings from config
         if hasattr(model, "vae") and hasattr(model.vae, "use_slicing"):
             model.vae.use_slicing = od_config.vae_use_slicing

vllm_omni/engine/arg_utils.py

@@ -7,7 +7,7 @@ import os
 import tempfile
 from collections.abc import Callable
 from dataclasses import dataclass, field, fields
-from typing import Any, cast
+from typing import Any, Literal, cast
 
 from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
 from vllm.logger import init_logger
@@ -566,6 +566,7 @@ class OrchestratorArgs:
     step_execution: bool = False
     vae_use_slicing: bool = False
     vae_use_tiling: bool = False
+    vae_stack_tiling: Literal["auto", "true", "false"] = "false"
     enable_multithread_weight_load: bool = True
     num_weight_load_threads: int = 4
     enable_cpu_offload: bool = False

vllm_omni/engine/async_omni_engine.py

@@ -1132,6 +1132,7 @@ class AsyncOmniEngine:
             "request_batch_max_wait_ms": kwargs.get("request_batch_max_wait_ms", 0.0),
             "vae_use_slicing": kwargs.get("vae_use_slicing", False),
             "vae_use_tiling": kwargs.get("vae_use_tiling", False),
+            "vae_stack_tiling": kwargs.get("vae_stack_tiling", "false"),
             "cache_backend": cache_backend,
             "cache_config": cache_config,
             "enable_cache_dit_summary": kwargs.get("enable_cache_dit_summary", False),

vllm_omni/entrypoints/cli/serve.py

@@ -686,6 +686,15 @@ class OmniServeCommand(CLISubcommand):
             action="store_true",
             help="Enable VAE tiling for memory optimization (useful for mitigating OOM issues).",
         )
+        omni_config_group.add_argument(
+            "--vae-stack-tiling",
+            choices=("auto", "true", "false"),
+            default="false",
+            help=(
+                "Batch native VAE tiles into fewer model calls. 'auto' enables it "
+                "when each VAE-parallel rank has multiple tiles (default: false)."
+            ),
+        )
 
         # Parallel weight loading (faster diffusion startup)
         omni_config_group.add_argument(

vllm_omni/utils/profiler.py

@@ -1,37 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
-
-from __future__ import annotations
-
-import argparse
-import json
-from typing import Any
-
-
-def parse_json_object(value: str, flag_name: str = "argument") -> dict[str, Any]:
-    """Parse a CLI value as a JSON object, attributing errors to ``flag_name``."""
-    try:
-        config = json.loads(value)
-    except json.JSONDecodeError as e:
-        raise argparse.ArgumentTypeError(f"{flag_name} must be valid JSON: {e}") from e
-    if not isinstance(config, dict):
-        raise argparse.ArgumentTypeError(f"{flag_name} must be a JSON object")
-    return config
-
-
-def parse_profiler_config(value: str) -> dict[str, Any]:
-    """Parse the JSON object passed to ``--profiler-config``."""
-    return parse_json_object(value, flag_name="--profiler-config")
-
-
-def add_profiler_config_arg(parser: argparse.ArgumentParser) -> None:
-    """Add the shared JSON profiler configuration option to ``parser``."""
-    parser.add_argument(
-        "--profiler-config",
-        type=parse_profiler_config,
-        default=None,
-        help=(
-            "JSON profiler config for torch/cuda profiling, e.g. "
-            '\'{"profiler":"torch","torch_profiler_dir":"./perf"}\'.'
-        ),
-    )