3542 lines
124 KiB
Python
3542 lines
124 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
|
|
|
|
import json
|
|
import sys
|
|
from contextlib import contextmanager, nullcontext
|
|
from multiprocessing.reduction import ForkingPickler
|
|
from types import SimpleNamespace
|
|
from typing import Any, TypeVar
|
|
from unittest.mock import Mock, create_autospec, patch
|
|
|
|
import numpy as np
|
|
import pytest
|
|
import torch
|
|
import torch.nn as nn
|
|
|
|
from vllm_omni.diffusion.models.minimax_h3.lora import TurboSpec, parse_turbo_filename
|
|
|
|
pytestmark = [pytest.mark.core_model, pytest.mark.cpu, pytest.mark.diffusion]
|
|
|
|
_ItemT = TypeVar("_ItemT")
|
|
_ResultT = TypeVar("_ResultT")
|
|
|
|
|
|
def _append_and_return(items: list[_ItemT], item: _ItemT, result: _ResultT) -> _ResultT:
|
|
items.append(item)
|
|
return result
|
|
|
|
|
|
def _turbo_spec(filename: str) -> TurboSpec:
|
|
"""Build the spec a loaded artifact of this name would carry."""
|
|
spec = parse_turbo_filename(filename)
|
|
assert spec is not None
|
|
return spec
|
|
|
|
|
|
def _encoder_payload(
|
|
*,
|
|
task: str = "t2va",
|
|
height: int = 768,
|
|
width: int = 1344,
|
|
num_frames: int = 124,
|
|
) -> dict[str, Any]:
|
|
from vllm_omni.diffusion.models.minimax_h3.time_request import MINIMAX_H3_SHAPE_PLANNER
|
|
from vllm_omni.model_executor.models.minimax_h3.conditioning import MiniMaxH3EncoderConditioning
|
|
|
|
conditioning = MiniMaxH3EncoderConditioning(
|
|
hidden_states=torch.ones(1, 5120, dtype=torch.bfloat16),
|
|
token_tags=torch.ones(1, dtype=torch.int64),
|
|
task=task,
|
|
height=height,
|
|
width=width,
|
|
num_frames=num_frames,
|
|
latent_t=MINIMAX_H3_SHAPE_PLANNER.video_latent_t(num_frames),
|
|
audio_t=MINIMAX_H3_SHAPE_PLANNER.audio_latent_t(num_frames / 24),
|
|
)
|
|
return conditioning.to_omni_payload()
|
|
|
|
|
|
def _encoder_prompt(text: str = "test") -> dict[str, Any]:
|
|
return {
|
|
"prompt": text,
|
|
"additional_information": {"encoder_output": _encoder_payload()},
|
|
}
|
|
|
|
|
|
@pytest.mark.parametrize("field", ["hidden_states", "token_tags"])
|
|
@pytest.mark.parametrize("invalid", ["dtype", "strides", "sparse"])
|
|
def test_unified_encoder_preserves_main_text_validation(field, invalid):
|
|
from vllm_omni.model_executor.models.minimax_h3.conditioning import MiniMaxH3EncoderConditioning
|
|
|
|
payload = _encoder_payload()
|
|
target, key = (
|
|
(payload["hidden_states"], "output") if field == "hidden_states" else (payload["meta"], "token_role_ids")
|
|
)
|
|
shape = (2, 5120) if field == "hidden_states" else (2,)
|
|
dtype = torch.bfloat16 if field == "hidden_states" else torch.int64
|
|
value = torch.ones(shape, dtype=dtype)
|
|
if invalid == "dtype":
|
|
value = value.float()
|
|
elif invalid == "strides":
|
|
value = torch.ones((4, *shape[1:]), dtype=dtype)[::2]
|
|
else:
|
|
value = value.to_sparse()
|
|
payload["hidden_states"]["output"] = torch.ones(2, 5120, dtype=torch.bfloat16)
|
|
payload["meta"]["token_role_ids"] = torch.ones(2, dtype=torch.int64)
|
|
target[key] = value
|
|
with pytest.raises(ValueError, match="minimax_h3.text_conditioning/v1"):
|
|
MiniMaxH3EncoderConditioning.from_omni_payload(payload)
|
|
|
|
|
|
@pytest.mark.parametrize("column_tags", [False, True])
|
|
def test_unified_encoder_roundtrips_all_conditioning(column_tags):
|
|
from vllm_omni.model_executor.models.minimax_h3.conditioning import MiniMaxH3EncoderConditioning
|
|
|
|
conditioning = MiniMaxH3EncoderConditioning(
|
|
hidden_states=torch.ones(2, 5120, dtype=torch.bfloat16),
|
|
token_tags=torch.tensor([0, 1]),
|
|
task="ref2va",
|
|
height=768,
|
|
width=1344,
|
|
num_frames=124,
|
|
latent_t=37,
|
|
audio_t=207,
|
|
visual_condition=torch.arange(96, dtype=torch.float32).reshape(1, 96),
|
|
visual_condition_shapes=((1, 2, 2),),
|
|
audio_condition=torch.ones(160, 32),
|
|
audio_condition_lengths=(80,),
|
|
ref_blocks=({"kind": "video_audio", "ref_audio_t": 80, "latent_t": 1, "latent_h": 2, "latent_w": 2},),
|
|
keyframe_frame_indices=(0, -1),
|
|
)
|
|
payload = conditioning.to_omni_payload()
|
|
if column_tags:
|
|
payload["meta"]["token_role_ids"] = payload["meta"]["token_role_ids"].unsqueeze(-1)
|
|
restored = MiniMaxH3EncoderConditioning.from_omni_payload(payload)
|
|
for name in ("hidden_states", "token_tags", "visual_condition", "audio_condition"):
|
|
torch.testing.assert_close(getattr(restored, name), getattr(conditioning, name))
|
|
for name in ("visual_condition_shapes", "audio_condition_lengths", "ref_blocks", "keyframe_frame_indices"):
|
|
assert getattr(restored, name) == getattr(conditioning, name)
|
|
|
|
|
|
@pytest.mark.parametrize("task", ["t2va", "fl2va", "ref2va"])
|
|
@pytest.mark.parametrize("text_key", [None, "encoder_output", "text_encoder_output"])
|
|
def test_local_encoder_modes_preserve_validated_text_and_unified_handoff(task, text_key):
|
|
from PIL import Image
|
|
|
|
from vllm_omni.inputs.data import OmniDiffusionSamplingParams
|
|
from vllm_omni.model_executor.models.minimax_h3.conditioning import MiniMaxH3EncoderConditioning
|
|
|
|
pipeline = _distilled_pipeline([], {"fl2va": None, "ref2va": None})
|
|
pipeline.load_text_encoder = text_key is None
|
|
pipeline.load_vae_encoder = True
|
|
text = {"hidden_states": torch.ones(2, 5120, dtype=torch.bfloat16), "token_tags": torch.tensor([0, 1])}
|
|
pipeline.encode_prompt = Mock(return_value=(text["hidden_states"], text["token_tags"]))
|
|
pipeline.video_vae = Mock()
|
|
pipeline.video_vae.is_distributed_enabled.return_value = False
|
|
pipeline.video_vae.encode_image.side_effect = lambda image: torch.ones(
|
|
(image.height // 32) * (image.width // 32), 96
|
|
)
|
|
pipeline.audio_vae = Mock()
|
|
prompt = {"prompt": "A person waves."}
|
|
if task != "t2va":
|
|
prompt["multi_modal_data"] = {"image": Image.new("RGB", (256, 256))}
|
|
if text_key is not None:
|
|
prompt["additional_information"] = {text_key: text}
|
|
sampling = OmniDiffusionSamplingParams(
|
|
height=32,
|
|
width=32,
|
|
num_frames=96,
|
|
num_inference_steps=2,
|
|
extra_args={"task": task, "aspect_ratio": "1:1"},
|
|
)
|
|
conditioning, window_embeddings = pipeline._prepare_local_conditioning(
|
|
prompt, sampling, require_external_text=text_key is not None
|
|
)
|
|
assert window_embeddings is None
|
|
assert pipeline.encode_prompt.call_count == int(text_key is None)
|
|
assert pipeline.video_vae.encode_image.call_count == int(task != "t2va")
|
|
restored = MiniMaxH3EncoderConditioning.from_omni_payload(conditioning.to_omni_payload())
|
|
pipeline.load_text_encoder = pipeline.load_vae_encoder = False
|
|
context = pipeline._prepare_request_inputs(
|
|
{"additional_information": {"encoder_output": restored.to_omni_payload()}}, sampling
|
|
)
|
|
assert context["task"] == task
|
|
assert context["num_steps"] == 2
|
|
torch.testing.assert_close(context["text_embeddings"], text["hidden_states"])
|
|
assert pipeline.video_vae.encode_image.call_count == int(task != "t2va")
|
|
|
|
|
|
def test_decode_to_mp4_batches_consumer_transfers(monkeypatch):
|
|
from vllm_omni.diffusion.models.minimax_h3 import pipeline_minimax_h3 as mod
|
|
|
|
class FakeEncoder:
|
|
instances: list["FakeEncoder"] = []
|
|
|
|
def __init__(self, **kwargs):
|
|
self.pushes = []
|
|
self.kwargs = kwargs
|
|
self.__class__.instances.append(self)
|
|
|
|
def push(self, frames, *, on_consumed=None):
|
|
self.pushes.append(np.array(frames, copy=True))
|
|
if on_consumed is not None:
|
|
on_consumed()
|
|
|
|
def finish(self):
|
|
return b"mp4"
|
|
|
|
def abort(self):
|
|
raise AssertionError("unexpected abort")
|
|
|
|
class FakeAudioVAE:
|
|
def decode_latent(self, latent):
|
|
return torch.zeros(1, 1, 2)
|
|
|
|
class FakeVideoVAE:
|
|
chunk_value_range = (0.0, 1.0)
|
|
|
|
def decode_with_chunks(self, latent, *, on_chunk):
|
|
for value in (0.0, 0.25, 0.5):
|
|
on_chunk(torch.full((1, 3, 3, 2, 2), value))
|
|
|
|
pipeline = object.__new__(mod.MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
pipeline.audio_vae = FakeAudioVAE()
|
|
pipeline.video_vae = FakeVideoVAE()
|
|
pipeline.device = torch.device("cpu")
|
|
monkeypatch.setattr(mod.MiniMaxH3Pipeline, "_uses_manual_component_offload", lambda self, component: False)
|
|
monkeypatch.setattr(
|
|
"vllm_omni.diffusion.utils.chunked_video.ChunkedMP4Encoder",
|
|
FakeEncoder,
|
|
)
|
|
|
|
output = pipeline.decode_to_mp4(
|
|
torch.zeros(1),
|
|
torch.zeros(1),
|
|
height=2,
|
|
width=2,
|
|
batch_frames=4,
|
|
)
|
|
|
|
assert output == b"mp4"
|
|
encoder = FakeEncoder.instances[-1]
|
|
assert [frames.shape[0] for frames in encoder.pushes] == [6, 3]
|
|
assert np.array_equal(encoder.pushes[0][0], np.full((2, 2, 3), 0, dtype=np.uint8))
|
|
assert np.array_equal(encoder.pushes[0][3], np.full((2, 2, 3), 64, dtype=np.uint8))
|
|
assert encoder.pushes[1].shape == (3, 2, 2, 3)
|
|
|
|
|
|
def test_request_video_codec_options_reach_the_preencoded_mp4_encoder(monkeypatch):
|
|
"""A client's encoder options must survive the worker-side pre-encode path."""
|
|
from vllm_omni.diffusion.models.minimax_h3 import pipeline_minimax_h3 as mod
|
|
|
|
class FakeEncoder:
|
|
instances: list["FakeEncoder"] = []
|
|
|
|
def __init__(self, **kwargs):
|
|
self.kwargs = kwargs
|
|
self.__class__.instances.append(self)
|
|
|
|
def push(self, frames, *, on_consumed=None):
|
|
del frames
|
|
if on_consumed is not None:
|
|
on_consumed()
|
|
|
|
def finish(self):
|
|
return b"mp4"
|
|
|
|
def abort(self):
|
|
raise AssertionError("unexpected abort")
|
|
|
|
class FakeAudioVAE:
|
|
def decode_latent(self, latent):
|
|
return torch.zeros(1, 1, 2)
|
|
|
|
class FakeVideoVAE:
|
|
chunk_value_range = (0.0, 1.0)
|
|
|
|
def decode_with_chunks(self, latent, *, on_chunk):
|
|
on_chunk(torch.zeros(1, 3, 1, 2, 2))
|
|
|
|
pipeline = object.__new__(mod.MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
pipeline.audio_vae = FakeAudioVAE()
|
|
pipeline.video_vae = FakeVideoVAE()
|
|
pipeline.device = torch.device("cpu")
|
|
monkeypatch.setattr(mod.MiniMaxH3Pipeline, "_uses_manual_component_offload", lambda self, component: False)
|
|
monkeypatch.setattr("vllm_omni.diffusion.utils.chunked_video.ChunkedMP4Encoder", FakeEncoder)
|
|
|
|
pipeline.decode_to_mp4(
|
|
torch.zeros(1),
|
|
torch.zeros(1),
|
|
height=2,
|
|
width=2,
|
|
video_codec_options={"preset": "ultrafast"},
|
|
)
|
|
|
|
assert FakeEncoder.instances[-1].kwargs["video_codec_options"] == {"preset": "ultrafast"}
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
("codec_extra", "expected"),
|
|
[
|
|
({}, {"preset": "ultrafast", "threads": "0"}),
|
|
({"video_codec_options": {"preset": "slow", "threads": 2}}, {"preset": "slow", "threads": "2"}),
|
|
({"video_codec_options": {}}, {}),
|
|
({"video_codec_options": None}, None),
|
|
],
|
|
)
|
|
@pytest.mark.parametrize("batch_extra, batch_frames", [({}, 17), ({"preencode_batch_frames": 5}, 5)])
|
|
def test_preencode_request_preserves_serving_codec_defaults(codec_extra, expected, batch_extra, batch_frames):
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
from vllm_omni.diffusion.request import OmniDiffusionRequest
|
|
from vllm_omni.diffusion.worker.request_batch import DiffusionRequestBatch
|
|
from vllm_omni.inputs.data import OmniDiffusionSamplingParams
|
|
|
|
pipeline = object.__new__(MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
pipeline.partition = "fl2va"
|
|
pipeline.supported_tasks = frozenset({"t2va"})
|
|
pipeline.default_video_shift = 12.0
|
|
pipeline.default_audio_shift = 3.0
|
|
pipeline.device = torch.device("cpu")
|
|
pipeline.od_config = SimpleNamespace()
|
|
pipeline.load_text_encoder = False
|
|
pipeline.load_vae_encoder = False
|
|
pipeline.text_encoder = object()
|
|
pipeline.encode_prompt = Mock(return_value=(torch.ones(1, 2), torch.ones(1, dtype=torch.long)))
|
|
pipeline._quality_policy = Mock()
|
|
pipeline._cache_dit_runtime = Mock()
|
|
pipeline.diffuse = Mock(return_value=(torch.zeros(1), torch.zeros(1)))
|
|
pipeline.decode_to_mp4 = Mock(return_value=b"mp4")
|
|
sampling = OmniDiffusionSamplingParams(
|
|
width=1344,
|
|
height=768,
|
|
fps=24,
|
|
num_frames=124,
|
|
num_inference_steps=50,
|
|
extra_args={"task": "t2va", "aspect_ratio": "16:9", "preencode_mp4": True, **codec_extra, **batch_extra},
|
|
)
|
|
batch = DiffusionRequestBatch(
|
|
[OmniDiffusionRequest(prompt=_encoder_prompt("test"), sampling_params=sampling, request_id="codec-defaults")]
|
|
)
|
|
|
|
pipeline.forward(batch)
|
|
|
|
assert pipeline.decode_to_mp4.call_args.kwargs["video_codec_options"] == expected
|
|
assert pipeline.decode_to_mp4.call_args.kwargs["batch_frames"] == batch_frames
|
|
|
|
|
|
def test_video_codec_options_are_normalized_for_the_encoder():
|
|
from vllm_omni.diffusion.utils.media_utils import normalize_video_codec_options
|
|
|
|
assert normalize_video_codec_options({"preset": "ultrafast", "threads": 0}) == {
|
|
"preset": "ultrafast",
|
|
"threads": "0",
|
|
}
|
|
assert normalize_video_codec_options(None) is None
|
|
with pytest.raises(ValueError, match="video_codec_options"):
|
|
normalize_video_codec_options("preset=ultrafast")
|
|
|
|
|
|
def test_h3_prepares_resolved_cache_state_immediately_before_denoise():
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
from vllm_omni.diffusion.request import OmniDiffusionRequest
|
|
from vllm_omni.diffusion.worker.request_batch import DiffusionRequestBatch
|
|
from vllm_omni.inputs.data import OmniDiffusionSamplingParams
|
|
|
|
pipeline = object.__new__(MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
pipeline.load_text_encoder = False
|
|
pipeline.load_vae_encoder = False
|
|
pipeline.partition = "fl2va"
|
|
pipeline.supported_tasks = frozenset({"t2va"})
|
|
pipeline.default_video_shift = 12.0
|
|
pipeline.default_audio_shift = 3.0
|
|
pipeline.device = torch.device("cpu")
|
|
pipeline.od_config = SimpleNamespace()
|
|
pipeline.load_text_encoder = False
|
|
pipeline.load_vae_encoder = False
|
|
cache_spec = object()
|
|
quality_plan = SimpleNamespace(cache_dit=cache_spec)
|
|
pipeline._quality_policy = Mock()
|
|
pipeline._quality_policy.resolve.return_value = quality_plan
|
|
events = []
|
|
pipeline._cache_dit_runtime = SimpleNamespace(prepare=lambda spec: events.append(("prepare", spec)))
|
|
|
|
def diffuse(**kwargs):
|
|
events.append(("diffuse", kwargs))
|
|
return torch.zeros(1), torch.zeros(1)
|
|
|
|
pipeline.diffuse = diffuse
|
|
decoded_audio = torch.zeros(1)
|
|
pipeline.decode = Mock(return_value=(torch.zeros(1, 3, 1, 1, 1), decoded_audio))
|
|
sampling = OmniDiffusionSamplingParams(
|
|
quality="high",
|
|
width=1344,
|
|
height=768,
|
|
fps=24,
|
|
num_frames=124,
|
|
num_inference_steps=50,
|
|
extra_args={"task": "t2va", "aspect_ratio": "16:9"},
|
|
)
|
|
batch = DiffusionRequestBatch(
|
|
[
|
|
OmniDiffusionRequest(
|
|
prompt=_encoder_prompt("quality boundary"),
|
|
sampling_params=sampling,
|
|
request_id="quality-boundary",
|
|
)
|
|
]
|
|
)
|
|
|
|
output = pipeline.forward(batch)
|
|
|
|
assert events[0] == ("prepare", cache_spec)
|
|
assert events[1][0] == "diffuse"
|
|
pipeline._quality_policy.resolve.assert_called_once_with(
|
|
quality="high",
|
|
num_inference_steps=50,
|
|
extra_args={"task": "t2va", "aspect_ratio": "16:9"},
|
|
)
|
|
assert output.output[0].shape == (1, 1, 1, 1, 3)
|
|
assert output.output[0].dtype == torch.uint8
|
|
assert output.output[1] is decoded_audio
|
|
|
|
|
|
def test_pipeline_import_registry_and_component_discovery():
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
from vllm_omni.diffusion.registry import (
|
|
_DIFFUSION_MODELS,
|
|
_DIFFUSION_POST_PROCESS_FUNCS,
|
|
)
|
|
|
|
assert _DIFFUSION_MODELS["MiniMaxH3Pipeline"] == (
|
|
"minimax_h3",
|
|
"pipeline_minimax_h3",
|
|
"MiniMaxH3Pipeline",
|
|
)
|
|
assert _DIFFUSION_MODELS["MiniMaxH3ModularPipeline"] == _DIFFUSION_MODELS["MiniMaxH3Pipeline"]
|
|
assert _DIFFUSION_POST_PROCESS_FUNCS["MiniMaxH3Pipeline"] == "get_minimax_h3_post_process_func"
|
|
assert (
|
|
_DIFFUSION_POST_PROCESS_FUNCS["MiniMaxH3ModularPipeline"] == _DIFFUSION_POST_PROCESS_FUNCS["MiniMaxH3Pipeline"]
|
|
)
|
|
assert MiniMaxH3Pipeline._dit_modules == ["transformer", "transformers_ref"]
|
|
assert MiniMaxH3Pipeline._encoder_modules == ["text_encoder"]
|
|
assert MiniMaxH3Pipeline._vae_modules == ["video_vae", "audio_vae"]
|
|
|
|
|
|
def test_diffusion_stage_has_only_dit_and_decoder_vaes_during_discovery():
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
from vllm_omni.diffusion.offloader.module_collector import ModuleDiscovery
|
|
|
|
pipeline = object.__new__(MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
pipeline.transformer = torch.nn.Linear(1, 1)
|
|
pipeline.video_vae = torch.nn.Linear(1, 1)
|
|
pipeline.audio_vae = torch.nn.Linear(1, 1)
|
|
pipeline._dit_modules = ["transformer"]
|
|
pipeline._encoder_modules = []
|
|
pipeline._vae_modules = ["video_vae", "audio_vae"]
|
|
|
|
discovered = ModuleDiscovery.discover(pipeline)
|
|
|
|
assert discovered.dit_names == ["transformer"]
|
|
assert discovered.encoder_names == []
|
|
assert discovered.vae_names == ["video_vae", "audio_vae"]
|
|
|
|
|
|
def test_diffusion_stage_requires_the_encoder_payload():
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
from vllm_omni.errors import OmniClientError
|
|
|
|
with pytest.raises(OmniClientError, match="requires encoder conditioning"):
|
|
MiniMaxH3Pipeline._extract_encoder_conditioning(None)
|
|
|
|
|
|
def test_encoder_free_stage_skips_missing_component_during_dlo_registration():
|
|
from vllm_omni.diffusion.models.minimax_h3 import pipeline_minimax_h3 as pipeline_module
|
|
|
|
cache = Mock()
|
|
video_vae = Mock()
|
|
audio_vae = Mock()
|
|
|
|
pipeline_module._register_dlo_component_cache(cache, None, video_vae, audio_vae)
|
|
|
|
video_vae.set_omni_component_cache.assert_called_once_with(cache)
|
|
audio_vae.set_omni_component_cache.assert_called_once_with(cache)
|
|
|
|
|
|
def _write_partition_index(path, *, partition, tasks):
|
|
path.mkdir(parents=True)
|
|
(path / "model_index.json").write_text(
|
|
json.dumps(
|
|
{
|
|
"_class_name": "MiniMaxH3Pipeline",
|
|
"_diffusers_version": "0.32.2",
|
|
"_minimax_h3": {
|
|
"partition": partition,
|
|
"tasks": tasks,
|
|
"sigma_shift_scales": {"video": 12.0, "audio": 3.0},
|
|
},
|
|
}
|
|
),
|
|
encoding="utf-8",
|
|
)
|
|
|
|
|
|
@pytest.mark.skip(reason="upstream utility was removed from this branch")
|
|
def test_modular_diffusers_index_is_resolved_generically(tmp_path):
|
|
from vllm_omni.config.resolver import resolve_omni_config
|
|
from vllm_omni.diffusion.data import OmniDiffusionConfig, resolve_model_class_name
|
|
from vllm_omni.diffusion.utils.hf_utils import is_diffusion_model
|
|
|
|
(tmp_path / "modular_model_index.json").write_text(
|
|
json.dumps(
|
|
{
|
|
"_class_name": "MiniMaxH3ModularPipeline",
|
|
"_diffusers_version": "0.36.0.dev0",
|
|
}
|
|
),
|
|
encoding="utf-8",
|
|
)
|
|
|
|
assert is_diffusion_model(str(tmp_path))
|
|
assert resolve_model_class_name(str(tmp_path)) == "MiniMaxH3ModularPipeline"
|
|
resolved = resolve_omni_config(
|
|
str(tmp_path),
|
|
trust_remote_code=False,
|
|
deploy_config_path=None,
|
|
cli_overrides=None,
|
|
stage_overrides=None,
|
|
strategy_config_path=None,
|
|
)
|
|
assert resolved.config_path is None
|
|
assert resolved.stage_by_id(0).engine_args.model_class_name == "MiniMaxH3ModularPipeline"
|
|
|
|
config = OmniDiffusionConfig(model=str(tmp_path))
|
|
config.enrich_config()
|
|
assert config.model_class_name == "MiniMaxH3ModularPipeline"
|
|
assert config.supports_multimodal_inputs
|
|
assert config.max_multimodal_image_inputs == 9
|
|
assert config.supports_mixed_reference_inputs
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
("task_type", "partition"),
|
|
[
|
|
(None, "combined"),
|
|
("auto", "combined"),
|
|
("t2va", "fl2va"),
|
|
("fl2va", "fl2va"),
|
|
("ref2va", "ref2va"),
|
|
],
|
|
)
|
|
def test_startup_task_selects_weight_partition(task_type, partition):
|
|
from vllm_omni.diffusion.models.minimax_h3.pipeline_minimax_h3 import _minimax_h3_partition_for_task
|
|
|
|
assert _minimax_h3_partition_for_task(task_type) == partition
|
|
|
|
|
|
def test_startup_auto_task_uses_explicit_local_partition(tmp_path):
|
|
from vllm_omni.diffusion.models.minimax_h3.pipeline_minimax_h3 import _minimax_h3_partition_for_task
|
|
|
|
for directory, partition, tasks in (
|
|
("FL2VA", "fl2va", ["t2va", "fl2va"]),
|
|
("Ref2VA", "ref2va", ["ref2va"]),
|
|
):
|
|
model_path = tmp_path / directory
|
|
_write_partition_index(model_path, partition=partition, tasks=tasks)
|
|
assert _minimax_h3_partition_for_task(None, str(model_path)) == partition
|
|
assert _minimax_h3_partition_for_task("auto", str(model_path)) == partition
|
|
assert _minimax_h3_partition_for_task("combined", str(model_path)) == "combined"
|
|
|
|
|
|
def test_startup_task_rejects_unsupported_value():
|
|
from vllm_omni.diffusion.models.minimax_h3.pipeline_minimax_h3 import _minimax_h3_partition_for_task
|
|
|
|
with pytest.raises(ValueError, match="task_type must be one of"):
|
|
_minimax_h3_partition_for_task("unsupported")
|
|
|
|
|
|
def test_encoder_task_validation_and_transformer_routing():
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
from vllm_omni.errors import OmniClientError
|
|
|
|
pipeline = object.__new__(MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
pipeline.partition = "combined"
|
|
pipeline.supported_tasks = frozenset({"t2va", "fl2va", "ref2va"})
|
|
pipeline.transformer = torch.nn.Identity()
|
|
pipeline.transformers_ref = torch.nn.Linear(1, 1)
|
|
assert pipeline._resolve_task(None, {}) == "t2va"
|
|
assert pipeline._resolve_task(None, {"image": object()}) == "fl2va"
|
|
assert pipeline._resolve_task(None, {"audio": object()}) == "ref2va"
|
|
assert pipeline._resolve_task(None, {"video": object()}) == "ref2va"
|
|
assert pipeline._resolve_task(None, {"audio": object()}, audio_mode="lock_source") == "t2va"
|
|
assert (
|
|
pipeline._resolve_task(
|
|
None,
|
|
{"image": object(), "audio": object()},
|
|
audio_mode="lock_source",
|
|
)
|
|
== "fl2va"
|
|
)
|
|
assert (
|
|
pipeline._resolve_task(
|
|
None,
|
|
{"video": object(), "audio": object()},
|
|
audio_mode="lock_source",
|
|
)
|
|
== "ref2va"
|
|
)
|
|
fl2v_spec = _turbo_spec("minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors")
|
|
ref2v_spec = _turbo_spec("minimax_h3_ref2v_turbo_8step_v1.0_768p_bf16.safetensors")
|
|
assert pipeline._resolve_task("fl2va", {"image": object()}, turbo_spec=fl2v_spec) == "fl2va"
|
|
with pytest.raises(OmniClientError, match="serves"):
|
|
pipeline._resolve_task("ref2va", {}, turbo_spec=fl2v_spec)
|
|
assert pipeline._resolve_task("ref2va", {"image": object()}, turbo_spec=ref2v_spec) == "ref2va"
|
|
assert pipeline._resolve_task("t2va") == "t2va"
|
|
assert pipeline._resolve_task("fl2va") == "fl2va"
|
|
assert pipeline._resolve_task("ref2va") == "ref2va"
|
|
pipeline.partition = "fl2va"
|
|
pipeline.supported_tasks = frozenset({"t2va", "fl2va"})
|
|
assert pipeline._resolve_task(None, {"audio": object()}, audio_mode="lock_source") == "t2va"
|
|
assert (
|
|
pipeline._resolve_task(
|
|
None,
|
|
{"image": object(), "audio": object()},
|
|
audio_mode="lock_source",
|
|
)
|
|
== "fl2va"
|
|
)
|
|
pipeline.partition = "ref2va"
|
|
pipeline.supported_tasks = frozenset({"ref2va"})
|
|
assert pipeline._resolve_task("ref2va") == "ref2va"
|
|
assert pipeline._resolve_task(None, {"audio": object()}, audio_mode="lock_source") == "ref2va"
|
|
|
|
pipeline.partition = "combined"
|
|
pipeline.supported_tasks = frozenset({"t2va", "fl2va", "ref2va"})
|
|
assert pipeline._transformer_for_task("fl2va") is pipeline.transformer
|
|
assert pipeline._transformer_for_task("ref2va") is pipeline.transformers_ref
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
(
|
|
"task_type",
|
|
"expected_partition",
|
|
"expected_dits",
|
|
"component_partition",
|
|
"source_partitions",
|
|
"expected_tasks",
|
|
),
|
|
[
|
|
(None, "combined", 2, "FL2VA", ["FL2VA", "Ref2VA"], {"t2va", "fl2va", "ref2va"}),
|
|
("fl2va", "fl2va", 1, "FL2VA", ["FL2VA"], {"t2va", "fl2va"}),
|
|
("ref2va", "ref2va", 1, "Ref2VA", ["Ref2VA"], {"ref2va"}),
|
|
],
|
|
)
|
|
@pytest.mark.parametrize("load_text_encoder", [None, True, False])
|
|
@pytest.mark.parametrize("trust_remote_code", [False, True])
|
|
def test_pipeline_loads_task_selected_components_with_encoder_ownership(
|
|
monkeypatch,
|
|
tmp_path,
|
|
task_type,
|
|
expected_partition,
|
|
expected_dits,
|
|
component_partition,
|
|
source_partitions,
|
|
expected_tasks,
|
|
load_text_encoder,
|
|
trust_remote_code,
|
|
):
|
|
from vllm_omni.diffusion.data import (
|
|
DiffusionParallelConfig,
|
|
OmniDiffusionConfig,
|
|
)
|
|
from vllm_omni.diffusion.models.minimax_h3 import (
|
|
pipeline_minimax_h3 as pipeline_module,
|
|
)
|
|
|
|
_write_partition_index(
|
|
tmp_path / "FL2VA",
|
|
partition="fl2va",
|
|
tasks=["t2va", "fl2va"],
|
|
)
|
|
_write_partition_index(
|
|
tmp_path / "Ref2VA",
|
|
partition="ref2va",
|
|
tasks=["ref2va"],
|
|
)
|
|
for partition_name in ("FL2VA", "Ref2VA"):
|
|
for component in (
|
|
"transformer",
|
|
"video_vae",
|
|
"audio_vae",
|
|
):
|
|
(tmp_path / partition_name / component).mkdir()
|
|
|
|
created: dict[str, list[Any]] = {"dit": [], "video_vae": [], "audio_vae": [], "text_encoder": []}
|
|
component_options = {}
|
|
|
|
class FakeModule(torch.nn.Module):
|
|
def __init__(self, *args, **kwargs):
|
|
super().__init__()
|
|
|
|
class FakeDiT(FakeModule):
|
|
def __init__(self, *args, **kwargs):
|
|
super().__init__()
|
|
created["dit"].append((args, kwargs))
|
|
|
|
def component_factory(name):
|
|
def create(path, *args, **kwargs):
|
|
created[name].append(str(path))
|
|
component_options[name] = kwargs
|
|
return FakeModule()
|
|
|
|
return create
|
|
|
|
monkeypatch.setattr(pipeline_module, "MiniMaxH3DiTModel", FakeDiT)
|
|
monkeypatch.setattr(pipeline_module, "MiniMaxH3Qwen3VLEncoder", component_factory("text_encoder"))
|
|
tokenizer_loader = Mock()
|
|
processor_loader = Mock()
|
|
monkeypatch.setattr(pipeline_module.Qwen2TokenizerFast, "from_pretrained", tokenizer_loader)
|
|
monkeypatch.setattr(pipeline_module.Qwen3VLProcessor, "from_pretrained", processor_loader)
|
|
monkeypatch.setattr(
|
|
pipeline_module,
|
|
"MiniMaxH3VideoVAE",
|
|
component_factory("video_vae"),
|
|
)
|
|
monkeypatch.setattr(
|
|
pipeline_module,
|
|
"MiniMaxH3AudioVAE",
|
|
component_factory("audio_vae"),
|
|
)
|
|
monkeypatch.setattr(
|
|
pipeline_module,
|
|
"get_local_device",
|
|
lambda: torch.device("cpu"),
|
|
)
|
|
download_calls = []
|
|
|
|
def fake_download(**kwargs):
|
|
download_calls.append(kwargs)
|
|
return str(tmp_path)
|
|
|
|
monkeypatch.setattr(
|
|
pipeline_module,
|
|
"download_weights_from_hf_specific",
|
|
fake_download,
|
|
)
|
|
|
|
od_config = OmniDiffusionConfig(
|
|
model="MiniMaxAI/MiniMax-H3",
|
|
trust_remote_code=trust_remote_code,
|
|
revision=None,
|
|
task_type=task_type,
|
|
quantization_config=None,
|
|
parallel_config=DiffusionParallelConfig(
|
|
cfg_parallel_size=1,
|
|
text_encoder_tp_size=1,
|
|
),
|
|
)
|
|
if load_text_encoder is not None:
|
|
od_config.model_loaded["text_encoder"] = load_text_encoder
|
|
od_config.model_loaded["vae_encoder"] = load_text_encoder
|
|
pipeline = pipeline_module.MiniMaxH3Pipeline(od_config=od_config)
|
|
expect_encoder = load_text_encoder is not False
|
|
|
|
assert pipeline.partition == expected_partition
|
|
assert pipeline.supported_tasks == expected_tasks
|
|
assert len(created["dit"]) == expected_dits
|
|
assert all(kwargs["diffusers_weights"] is False for _, kwargs in created["dit"])
|
|
component_path = tmp_path / component_partition
|
|
assert created["video_vae"] == [str(component_path / "video_vae")]
|
|
assert created["audio_vae"] == [str(component_path / "audio_vae")]
|
|
assert created["text_encoder"] == ([str(component_path / "text_encoder")] if expect_encoder else [])
|
|
assert tokenizer_loader.call_count == processor_loader.call_count == int(expect_encoder)
|
|
assert pipeline.load_text_encoder is expect_encoder
|
|
assert pipeline._encoder_modules == (["text_encoder"] if expect_encoder else [])
|
|
assert ("text_encoder" in pipeline._offload_plan.encoder_block_attrs) is expect_encoder
|
|
assert ("text_encoder" in pipeline._offload_plan.on_demand_component_paths) is expect_encoder
|
|
assert ("encode_prompt" in pipeline._PROFILER_TARGETS) is expect_encoder
|
|
assert ("_encode_local_media" in pipeline._PROFILER_TARGETS) is expect_encoder
|
|
assert component_options["video_vae"]["decode_only"] is not expect_encoder
|
|
assert component_options["audio_vae"]["decode_only"] is not expect_encoder
|
|
assert component_options["video_vae"]["trust_remote_code"] is trust_remote_code
|
|
assert component_options["audio_vae"]["trust_remote_code"] is trust_remote_code
|
|
from vllm_omni.diffusion.offloader.module_collector import ModuleDiscovery
|
|
|
|
assert ModuleDiscovery.discover(pipeline).encoder_names == (["text_encoder"] if expect_encoder else [])
|
|
expected_dit_modules = ["transformer", "transformers_ref"] if expected_dits == 2 else ["transformer"]
|
|
assert pipeline._dit_modules == expected_dit_modules
|
|
assert hasattr(pipeline, "transformers_ref") is (expected_dits == 2)
|
|
if expect_encoder:
|
|
assert created["text_encoder"] == [str(component_path / "text_encoder")]
|
|
assert tokenizer_loader.call_count == 1
|
|
assert processor_loader.call_count == 1
|
|
assert pipeline.text_encoder is not None
|
|
assert pipeline.tokenizer is not None
|
|
assert pipeline.processor is not None
|
|
assert pipeline._encoder_modules == ["text_encoder"]
|
|
assert pipeline.text_encoder_tp_size == 1
|
|
else:
|
|
assert created["text_encoder"] == []
|
|
tokenizer_loader.assert_not_called()
|
|
processor_loader.assert_not_called()
|
|
assert pipeline.text_encoder is None
|
|
assert pipeline.tokenizer is None
|
|
assert pipeline.processor is None
|
|
assert pipeline._encoder_modules == []
|
|
assert pipeline.text_encoder_tp_size == 0
|
|
assert pipeline.text_encoder_group is None
|
|
if expected_partition == "ref2va":
|
|
assert pipeline._transformer_for_task("ref2va") is pipeline.transformer
|
|
assert [source.model_or_path for source in pipeline.weights_sources] == [
|
|
*(str(tmp_path / partition_name) for partition_name in source_partitions),
|
|
*([str(component_path)] if expect_encoder else []),
|
|
]
|
|
assert [source.subfolder for source in pipeline.weights_sources] == [
|
|
*("transformer" for _ in source_partitions),
|
|
*(["text_encoder"] if expect_encoder else []),
|
|
]
|
|
expected_prefixes = ["transformer."]
|
|
if expected_dits == 2:
|
|
expected_prefixes.append("transformers_ref.")
|
|
if expect_encoder:
|
|
expected_prefixes.append("text_encoder.")
|
|
assert [source.prefix for source in pipeline.weights_sources] == expected_prefixes
|
|
expected_patterns = pipeline_module.MINIMAX_H3_DIFFUSION_DOWNLOAD_PATTERNS[expected_partition]
|
|
if expect_encoder:
|
|
expected_patterns = (
|
|
pipeline_module.MINIMAX_H3_DOWNLOAD_PATTERNS
|
|
if expected_partition == "combined"
|
|
else pipeline_module.MINIMAX_H3_TASK_DOWNLOAD_PATTERNS[expected_partition]
|
|
)
|
|
assert download_calls == [
|
|
{
|
|
"model_name_or_path": "MiniMaxAI/MiniMax-H3",
|
|
"cache_dir": None,
|
|
"allow_patterns": expected_patterns,
|
|
"revision": None,
|
|
"require_all": True,
|
|
}
|
|
]
|
|
|
|
|
|
def test_combined_weight_loader_routes_each_contiguous_partition():
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
|
|
class FakeTransformer(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.loaded = []
|
|
self.post_load_calls = 0
|
|
|
|
def load_weights(self, weights):
|
|
self.loaded = [name for name, _ in weights]
|
|
return set(self.loaded)
|
|
|
|
def post_load_weights(self):
|
|
self.post_load_calls += 1
|
|
|
|
pipeline = object.__new__(MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
pipeline.transformer = FakeTransformer()
|
|
pipeline.transformers_ref = FakeTransformer()
|
|
pipeline.video_vae = torch.nn.Identity()
|
|
pipeline.audio_vae = torch.nn.Identity()
|
|
loaded = pipeline.load_weights(
|
|
iter(
|
|
[
|
|
("transformer.a", torch.ones(1)),
|
|
("transformer.b", torch.ones(1)),
|
|
("transformers_ref.a", torch.ones(1)),
|
|
]
|
|
)
|
|
)
|
|
|
|
assert pipeline.transformer.loaded == ["a", "b"]
|
|
assert pipeline.transformers_ref.loaded == ["a"]
|
|
assert pipeline.transformer.post_load_calls == 1
|
|
assert pipeline.transformers_ref.post_load_calls == 1
|
|
assert loaded == {"transformer.a", "transformer.b", "transformers_ref.a"}
|
|
|
|
|
|
def test_dlo_offload_plan_includes_token_refiner():
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
|
|
assert MiniMaxH3Pipeline._offload_plan.offload_submodules == {"token_refiner": "blocks"}
|
|
assert MiniMaxH3Pipeline._offload_plan.resident_dit_paths == frozenset({"transformer"})
|
|
assert MiniMaxH3Pipeline._offload_plan.encoder_component_types == {"text_encoder": "text_encoder"}
|
|
assert MiniMaxH3Pipeline._offload_plan.encoder_block_attrs == {
|
|
"text_encoder": ("vision.blocks", "text_model.layers")
|
|
}
|
|
|
|
|
|
def test_joint_postprocess_is_multiprocessing_picklable():
|
|
from vllm_omni.diffusion.models.minimax_h3 import (
|
|
get_minimax_h3_post_process_func,
|
|
)
|
|
|
|
postprocess = get_minimax_h3_post_process_func(SimpleNamespace())
|
|
postprocess = ForkingPickler.loads(ForkingPickler.dumps(postprocess))
|
|
# decode() hands over quantised frames already laid out as (B, T, H, W, C).
|
|
video = torch.arange(2 * 2 * 4 * 5 * 3, dtype=torch.uint8).reshape(2, 2, 4, 5, 3)
|
|
audio = torch.arange(12, dtype=torch.float32).reshape(1, 2, 6)
|
|
|
|
result = postprocess((video, audio), output_type="np")
|
|
|
|
assert isinstance(result["video"], list)
|
|
assert result["video"][0].dtype == np.uint8
|
|
assert result["video"][0].shape == (2, 4, 5, 3)
|
|
np.testing.assert_array_equal(result["video"][0], video[0].numpy())
|
|
np.testing.assert_array_equal(result["audio"], audio.numpy())
|
|
assert result["audio_sample_rate"] == 32000
|
|
assert result["fps"] == 24
|
|
|
|
|
|
def test_joint_video_output_is_quantized_before_transfer():
|
|
from vllm_omni.diffusion.models.minimax_h3.pipeline_minimax_h3 import (
|
|
_minimax_h3_post_process,
|
|
_prepare_minimax_h3_video_output,
|
|
)
|
|
|
|
video = torch.linspace(-0.1, 1.1, 2 * 3 * 2 * 4 * 5, dtype=torch.float64).reshape(2, 3, 2, 4, 5)
|
|
expected = torch.round(video.float().clamp(0, 1) * 255).to(torch.uint8).permute(0, 2, 3, 4, 1)
|
|
prepared = _prepare_minimax_h3_video_output(video)
|
|
|
|
assert prepared.dtype == torch.uint8
|
|
assert prepared.is_contiguous()
|
|
assert prepared.shape == (2, 2, 4, 5, 3)
|
|
torch.testing.assert_close(prepared, expected, rtol=0, atol=0)
|
|
|
|
result = _minimax_h3_post_process((prepared, torch.zeros(1, 2, 6)))
|
|
assert result["video"][0].dtype == np.uint8
|
|
np.testing.assert_array_equal(result["video"][0], expected[0].numpy())
|
|
|
|
|
|
def test_cfg_parallel_is_rejected_for_distilled_checkpoint():
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
|
|
od_config = SimpleNamespace(
|
|
parallel_config=SimpleNamespace(cfg_parallel_size=2),
|
|
)
|
|
with pytest.raises(ValueError, match="CFG-distilled"):
|
|
MiniMaxH3Pipeline(od_config=od_config)
|
|
|
|
|
|
def test_shape_contract_matches_three_reference_tasks():
|
|
from vllm_omni.diffusion.models.minimax_h3.time_request import (
|
|
MINIMAX_H3_SHAPE_PLANNER,
|
|
minimax_h3_align_frame_count,
|
|
)
|
|
|
|
assert minimax_h3_align_frame_count(round(8.7 * 24)) == 209
|
|
assert MINIMAX_H3_SHAPE_PLANNER.video_latent_t(209) == 62
|
|
assert MINIMAX_H3_SHAPE_PLANNER.audio_latent_t(8.7) == 348
|
|
assert minimax_h3_align_frame_count(round(5.0 * 24)) == 124
|
|
assert MINIMAX_H3_SHAPE_PLANNER.video_latent_t(124) == 37
|
|
assert MINIMAX_H3_SHAPE_PLANNER.audio_latent_t(5.0) == 200
|
|
assert MINIMAX_H3_SHAPE_PLANNER.frame_count_from_video_latent_t(62) == 209
|
|
|
|
|
|
def test_shifted_sigma_schedule_matches_reference_values():
|
|
from vllm_omni.diffusion.models.minimax_h3.time_request import (
|
|
minimax_h3_time_shift_sigmas,
|
|
)
|
|
|
|
sigmas = minimax_h3_time_shift_sigmas(num_steps=4, shift_scale=12.0)
|
|
|
|
assert sigmas == pytest.approx(
|
|
[1.0, 0.9729729891, 0.9230769277, 0.8000000119, 0.0],
|
|
abs=1e-7,
|
|
)
|
|
|
|
|
|
@pytest.mark.parametrize("num_steps", [1, 8, 50])
|
|
@pytest.mark.parametrize("shift_scale", [3.0, 12.0])
|
|
def test_uniform_sigma_schedule_has_one_interval_per_requested_step(num_steps, shift_scale):
|
|
from vllm_omni.diffusion.models.minimax_h3.time_request import minimax_h3_time_shift_sigmas
|
|
|
|
sigmas = minimax_h3_time_shift_sigmas(num_steps=num_steps, shift_scale=shift_scale)
|
|
|
|
assert len(sigmas) == num_steps + 1
|
|
assert sigmas[0] == 1.0
|
|
assert sigmas[-1] == 0.0
|
|
assert all(current > following for current, following in zip(sigmas, sigmas[1:]))
|
|
|
|
|
|
def test_base_schedule_overrides_the_uniform_sigma_positions():
|
|
from vllm_omni.diffusion.models.minimax_h3.time_request import (
|
|
minimax_h3_time_shift_sigmas,
|
|
)
|
|
|
|
base_schedule = [1.0, 0.7, 0.4, 0.15, 0.0]
|
|
|
|
video = minimax_h3_time_shift_sigmas(
|
|
num_steps=len(base_schedule),
|
|
shift_scale=12.0,
|
|
base_schedule=base_schedule,
|
|
)
|
|
audio = minimax_h3_time_shift_sigmas(
|
|
num_steps=len(base_schedule),
|
|
shift_scale=3.0,
|
|
base_schedule=base_schedule,
|
|
)
|
|
|
|
assert video == pytest.approx([1.0, 0.9655172, 0.8888889, 0.6792453, 0.0], abs=1e-6)
|
|
assert audio == pytest.approx([1.0, 0.875, 0.6666667, 0.3461539, 0.0], abs=1e-6)
|
|
assert video != pytest.approx(
|
|
minimax_h3_time_shift_sigmas(num_steps=len(base_schedule), shift_scale=12.0),
|
|
abs=1e-6,
|
|
)
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
"base_schedule",
|
|
[
|
|
[1.5, 0.5, 0.0],
|
|
[1.0, 0.5, 0.1],
|
|
[1.0, 0.5, 0.5, 0.0],
|
|
[1.0],
|
|
[],
|
|
[1.0, float("nan"), 0.4, 0.0],
|
|
[1.0, float("inf"), 0.4, 0.0],
|
|
],
|
|
)
|
|
def test_base_schedule_rejects_malformed_positions(base_schedule):
|
|
from vllm_omni.diffusion.models.minimax_h3.time_request import (
|
|
minimax_h3_time_shift_sigmas,
|
|
)
|
|
|
|
with pytest.raises(ValueError):
|
|
minimax_h3_time_shift_sigmas(
|
|
num_steps=max(len(base_schedule), 1),
|
|
shift_scale=12.0,
|
|
base_schedule=base_schedule,
|
|
)
|
|
|
|
|
|
def _distilled_pipeline(diffuse_calls, base_schedule_by_partition):
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
from vllm_omni.diffusion.sched import DMD2SigmaSchedule
|
|
|
|
schedules = {
|
|
partition: None if positions is None else DMD2SigmaSchedule.from_positions(positions)
|
|
for partition, positions in base_schedule_by_partition.items()
|
|
}
|
|
pipeline = object.__new__(MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
pipeline.partition = "combined"
|
|
pipeline.supported_tasks = frozenset({"t2va", "fl2va", "ref2va"})
|
|
pipeline.default_video_shift = 12.0
|
|
pipeline.default_audio_shift = 3.0
|
|
pipeline.device = torch.device("cpu")
|
|
pipeline.od_config = SimpleNamespace()
|
|
pipeline._base_schedule_by_partition = schedules
|
|
pipeline.load_text_encoder = False
|
|
pipeline.load_vae_encoder = False
|
|
pipeline._quality_policy = Mock()
|
|
pipeline._quality_policy.resolve.return_value = SimpleNamespace(cache_dit=None)
|
|
pipeline._cache_dit_runtime = SimpleNamespace(prepare=lambda spec: None)
|
|
|
|
def diffuse(**kwargs):
|
|
diffuse_calls.append(kwargs)
|
|
return torch.zeros(1), torch.zeros(1)
|
|
|
|
pipeline.diffuse = diffuse
|
|
pipeline.decode = Mock(return_value=(torch.zeros(1, 3, 1, 1, 1), torch.zeros(1)))
|
|
return pipeline
|
|
|
|
|
|
def _t2va_batch(num_inference_steps=None):
|
|
from vllm_omni.diffusion.request import OmniDiffusionRequest
|
|
from vllm_omni.diffusion.worker.request_batch import DiffusionRequestBatch
|
|
from vllm_omni.inputs.data import OmniDiffusionSamplingParams
|
|
|
|
sampling = OmniDiffusionSamplingParams(
|
|
quality="lossless",
|
|
width=1344,
|
|
height=768,
|
|
fps=24,
|
|
num_frames=124,
|
|
num_inference_steps=num_inference_steps,
|
|
extra_args={"task": "t2va", "aspect_ratio": "16:9"},
|
|
)
|
|
return DiffusionRequestBatch(
|
|
[
|
|
OmniDiffusionRequest(
|
|
prompt=_encoder_prompt("distilled schedule"),
|
|
sampling_params=sampling,
|
|
request_id="distilled",
|
|
)
|
|
]
|
|
)
|
|
|
|
|
|
def test_base_schedule_is_scoped_to_the_serving_partition():
|
|
"""A distilled FL2VA must not drag a regular Ref2VA onto its step count."""
|
|
distilled = [1.0, 0.7, 0.4, 0.15, 0.0]
|
|
pipeline = _distilled_pipeline([], {"fl2va": distilled, "ref2va": None})
|
|
|
|
assert pipeline._base_schedule_for_task("t2va").base_schedule == tuple(distilled)
|
|
assert pipeline._base_schedule_for_task("fl2va").base_schedule == tuple(distilled)
|
|
assert pipeline._base_schedule_for_task("ref2va") is None
|
|
|
|
|
|
def test_partially_constructed_pipeline_falls_back_to_the_uniform_schedule():
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
|
|
pipeline = object.__new__(MiniMaxH3Pipeline)
|
|
|
|
assert pipeline._base_schedule_for_task("t2va") is None
|
|
|
|
|
|
def test_distilled_forward_reports_denoising_steps_not_sigma_boundaries():
|
|
diffuse_calls: list[dict[str, Any]] = []
|
|
base_schedule = [1.0, 0.7, 0.4, 0.15, 0.0]
|
|
pipeline = _distilled_pipeline(diffuse_calls, {"fl2va": base_schedule, "ref2va": None})
|
|
|
|
pipeline.forward(_t2va_batch())
|
|
|
|
assert diffuse_calls[0]["base_schedule"] == tuple(base_schedule)
|
|
# Five boundaries describe four denoising steps.
|
|
assert diffuse_calls[0]["num_steps"] == 4
|
|
pipeline._quality_policy.resolve.assert_called_once_with(
|
|
quality="lossless",
|
|
num_inference_steps=4,
|
|
extra_args={"task": "t2va", "aspect_ratio": "16:9"},
|
|
)
|
|
|
|
|
|
def test_distilled_forward_accepts_the_matching_explicit_step_count():
|
|
diffuse_calls: list[dict[str, Any]] = []
|
|
pipeline = _distilled_pipeline(diffuse_calls, {"fl2va": [1.0, 0.7, 0.4, 0.15, 0.0], "ref2va": None})
|
|
|
|
pipeline.forward(_t2va_batch(num_inference_steps=4))
|
|
|
|
assert diffuse_calls[0]["num_steps"] == 4
|
|
|
|
|
|
def test_distilled_forward_rejects_a_mismatched_explicit_step_count():
|
|
from vllm_omni.errors import OmniClientError
|
|
|
|
pipeline = _distilled_pipeline([], {"fl2va": [1.0, 0.7, 0.4, 0.15, 0.0], "ref2va": None})
|
|
|
|
with pytest.raises(OmniClientError, match="must be 4 or omitted"):
|
|
pipeline.forward(_t2va_batch(num_inference_steps=50))
|
|
|
|
|
|
def test_absent_base_schedule_key_differs_from_an_empty_list():
|
|
from vllm_omni.diffusion.models.minimax_h3.pipeline_minimax_h3 import (
|
|
_read_base_schedule,
|
|
)
|
|
|
|
assert _read_base_schedule({}) is None
|
|
assert _read_base_schedule({"base_schedule": [1.0, 0.5, 0.0]}).base_schedule == (1.0, 0.5, 0.0)
|
|
with pytest.raises(ValueError):
|
|
_read_base_schedule({"base_schedule": []})
|
|
|
|
|
|
def test_cudnn_packed_attention_uses_python_length_without_padding_mask():
|
|
from vllm_omni.diffusion.models.minimax_h3.minimax_h3_transformer import (
|
|
MiniMaxH3Attention,
|
|
)
|
|
|
|
class FakeBackend:
|
|
supports_prefix_kv_slicing = True
|
|
|
|
@classmethod
|
|
def supports_packed_mask_free(cls) -> bool:
|
|
return False
|
|
|
|
class FakeAttention(torch.nn.Module):
|
|
attn_backend = FakeBackend
|
|
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.metadata = None
|
|
|
|
def forward(self, query, key, value, metadata):
|
|
self.metadata = metadata
|
|
return query
|
|
|
|
attention = object.__new__(MiniMaxH3Attention)
|
|
torch.nn.Module.__init__(attention)
|
|
attention.attention = FakeAttention()
|
|
# Model-side Ulysses shards rows before Attention gathers them again.
|
|
# The Python packed_total must therefore remain global even though q is
|
|
# local here.
|
|
q = torch.randn(4, 2, 4)
|
|
cu_seqlens = torch.tensor([0, 5, 8], dtype=torch.int32)
|
|
|
|
output = attention._run_packed_attention(
|
|
q,
|
|
q,
|
|
q,
|
|
cu_seqlens=cu_seqlens,
|
|
max_seqlen=5,
|
|
packed_total=8,
|
|
)
|
|
|
|
assert output.shape == q.shape
|
|
metadata = attention.attention.metadata
|
|
assert metadata is not None
|
|
assert metadata.attn_mask is None
|
|
assert metadata.extra["valid_kv_length"] == 5
|
|
|
|
|
|
def test_packed_attention_skips_mask_for_packed_mask_free_backend():
|
|
from vllm_omni.diffusion.attention.backends.abstract import (
|
|
VideoTokenLayout,
|
|
VideoTokenSpan,
|
|
)
|
|
from vllm_omni.diffusion.models.minimax_h3.minimax_h3_transformer import (
|
|
MiniMaxH3Attention,
|
|
)
|
|
|
|
class FakeBackend:
|
|
supports_prefix_kv_slicing = False
|
|
|
|
@classmethod
|
|
def supports_packed_mask_free(cls) -> bool:
|
|
return True
|
|
|
|
class FakeAttention(torch.nn.Module):
|
|
attn_backend = FakeBackend
|
|
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.metadata = None
|
|
|
|
def forward(self, query, key, value, metadata):
|
|
self.metadata = metadata
|
|
return query
|
|
|
|
attention = object.__new__(MiniMaxH3Attention)
|
|
torch.nn.Module.__init__(attention)
|
|
attention.attention = FakeAttention()
|
|
q = torch.randn(8, 2, 4)
|
|
video_layout = VideoTokenLayout(
|
|
used_len=5,
|
|
video_spans=(VideoTokenSpan(start=3, latent_grid=(1, 1, 2), role="target"),),
|
|
)
|
|
|
|
attention._run_packed_attention(
|
|
q,
|
|
q,
|
|
q,
|
|
cu_seqlens=torch.tensor([0, 5, 8], dtype=torch.int32),
|
|
max_seqlen=5,
|
|
packed_total=8,
|
|
video_layout=video_layout,
|
|
vsa_prefix_segments=(1, 2),
|
|
gate_compress=torch.ones_like(q),
|
|
)
|
|
|
|
metadata = attention.attention.metadata
|
|
assert metadata is not None
|
|
assert metadata.attn_mask is None
|
|
assert metadata.extra["valid_kv_length"] == 5
|
|
assert metadata.extra["npu_attn_varlen"] is True
|
|
assert metadata.video_layout is video_layout
|
|
assert metadata.extra["vsa_h3_prefix_segments"] == (1, 2)
|
|
assert "vsa_h3_video_shape" not in metadata.extra
|
|
assert "vsa_h3_target_start" not in metadata.extra
|
|
assert metadata.extra["gate_compress"].shape == (1, *q.shape)
|
|
packed_padding = metadata.packed_padding
|
|
assert packed_padding is not None
|
|
assert packed_padding.q_length == 5
|
|
assert packed_padding.kv_length == 5
|
|
assert packed_padding.cu_seqlens_q.tolist() == [0, 5]
|
|
assert packed_padding.cu_seqlens_k.tolist() == [0, 5]
|
|
|
|
|
|
def test_packed_attention_keeps_padding_mask_for_other_backends():
|
|
from vllm_omni.diffusion.models.minimax_h3.minimax_h3_transformer import (
|
|
MiniMaxH3Attention,
|
|
)
|
|
|
|
class FakeBackend:
|
|
supports_prefix_kv_slicing = False
|
|
|
|
@classmethod
|
|
def supports_packed_mask_free(cls) -> bool:
|
|
return False
|
|
|
|
class FakeAttention(torch.nn.Module):
|
|
attn_backend = FakeBackend
|
|
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.metadata = None
|
|
|
|
def forward(self, query, key, value, metadata):
|
|
self.metadata = metadata
|
|
return query
|
|
|
|
attention = object.__new__(MiniMaxH3Attention)
|
|
torch.nn.Module.__init__(attention)
|
|
attention.attention = FakeAttention()
|
|
q = torch.randn(8, 2, 4)
|
|
|
|
attention._run_packed_attention(
|
|
q,
|
|
q,
|
|
q,
|
|
cu_seqlens=torch.tensor([0, 5, 8], dtype=torch.int32),
|
|
max_seqlen=5,
|
|
packed_total=8,
|
|
)
|
|
|
|
metadata = attention.attention.metadata
|
|
assert metadata is not None
|
|
assert torch.equal(
|
|
metadata.attn_mask,
|
|
torch.tensor([[True, True, True, True, True, False, False, False]]),
|
|
)
|
|
|
|
|
|
def test_ring_packed_attention_publishes_prefix_length_without_global_mask():
|
|
attention = _fake_packed_attention("CUDNN_ATTN")
|
|
attention.attention.use_ring = True
|
|
q = torch.randn(4, 2, 4)
|
|
|
|
attention._run_packed_attention(
|
|
q,
|
|
q,
|
|
q,
|
|
cu_seqlens=torch.tensor([0, 5, 8], dtype=torch.int32),
|
|
max_seqlen=5,
|
|
packed_total=8,
|
|
)
|
|
|
|
metadata = attention.attention.metadata
|
|
assert metadata.attn_mask is None
|
|
assert metadata.extra["valid_kv_length"] == 5
|
|
|
|
|
|
def test_ring_packed_attention_keeps_mask_when_runtime_sp_is_inactive(monkeypatch):
|
|
from vllm_omni.diffusion.models.minimax_h3 import minimax_h3_transformer
|
|
|
|
class InactiveContext:
|
|
sp_active = False
|
|
|
|
attention = _fake_packed_attention("CUDNN_ATTN")
|
|
attention.attention.use_ring = True
|
|
monkeypatch.setattr(
|
|
minimax_h3_transformer,
|
|
"is_forward_context_available",
|
|
lambda: True,
|
|
raising=False,
|
|
)
|
|
monkeypatch.setattr(
|
|
minimax_h3_transformer,
|
|
"get_forward_context",
|
|
InactiveContext,
|
|
raising=False,
|
|
)
|
|
q = torch.randn(8, 2, 4)
|
|
|
|
attention._run_packed_attention(
|
|
q,
|
|
q,
|
|
q,
|
|
cu_seqlens=torch.tensor([0, 5, 8], dtype=torch.int32),
|
|
max_seqlen=5,
|
|
packed_total=8,
|
|
)
|
|
|
|
metadata = attention.attention.metadata
|
|
assert torch.equal(
|
|
metadata.attn_mask,
|
|
torch.tensor([[True, True, True, True, True, False, False, False]]),
|
|
)
|
|
assert metadata.extra["npu_attn_varlen"] is True
|
|
|
|
|
|
def _fake_packed_attention(
|
|
backend_name: str,
|
|
*,
|
|
prefix_kv_slicing: bool = False,
|
|
supports_multi_doc: bool = False,
|
|
):
|
|
from vllm_omni.diffusion.models.minimax_h3.minimax_h3_transformer import (
|
|
MiniMaxH3Attention,
|
|
)
|
|
|
|
class FakeBackend:
|
|
supports_prefix_kv_slicing = prefix_kv_slicing
|
|
|
|
@staticmethod
|
|
def get_name() -> str:
|
|
return backend_name
|
|
|
|
@classmethod
|
|
def supports_multi_doc_packed_varlen(cls) -> bool:
|
|
return supports_multi_doc
|
|
|
|
@classmethod
|
|
def supports_packed_mask_free(cls) -> bool:
|
|
return False
|
|
|
|
class FakeAttention(torch.nn.Module):
|
|
attn_backend = FakeBackend
|
|
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.metadata = None
|
|
|
|
def forward(self, query, key, value, metadata):
|
|
self.metadata = metadata
|
|
return query
|
|
|
|
attention = object.__new__(MiniMaxH3Attention)
|
|
torch.nn.Module.__init__(attention)
|
|
attention.attention = FakeAttention()
|
|
return attention
|
|
|
|
|
|
def test_packed_attention_drops_prefix_semantics_for_a_co_batched_layout():
|
|
"""A batch's valid rows are block-diagonal, so no prefix can describe them."""
|
|
attention = _fake_packed_attention("FLASH_ATTN", supports_multi_doc=True)
|
|
q = torch.randn(12, 2, 4)
|
|
|
|
attention._run_packed_attention(
|
|
q,
|
|
q,
|
|
q,
|
|
# Two requests: rows + padding tail each.
|
|
cu_seqlens=torch.tensor([0, 5, 6, 11, 12], dtype=torch.int32),
|
|
max_seqlen=5,
|
|
packed_total=12,
|
|
num_requests=2,
|
|
)
|
|
|
|
assert attention.attention.metadata.attn_mask is None
|
|
assert attention.attention.metadata.extra["valid_kv_length"] == 12
|
|
|
|
|
|
@pytest.mark.parametrize("backend_name", ["TORCH_SDPA", "FLASH_ATTN"])
|
|
def test_packed_attention_rejects_backends_without_multi_doc_capability(backend_name):
|
|
"""Backend names alone must not allow cross-request attention."""
|
|
attention = _fake_packed_attention(backend_name, supports_multi_doc=False)
|
|
q = torch.randn(12, 2, 4)
|
|
|
|
with pytest.raises(ValueError, match=backend_name):
|
|
attention._run_packed_attention(
|
|
q,
|
|
q,
|
|
q,
|
|
cu_seqlens=torch.tensor([0, 5, 6, 11, 12], dtype=torch.int32),
|
|
max_seqlen=5,
|
|
packed_total=12,
|
|
num_requests=2,
|
|
)
|
|
|
|
|
|
def test_rainfusion_packed_padding_stays_mask_free_on_unaligned_lengths():
|
|
"""Regression for the #5543 follow-up crash fixed in #7235.
|
|
|
|
MiniMax-H3 pads every packed request to a multiple of 64 rows, so any
|
|
non-64-aligned length reaches this path with used < packed_total. Before
|
|
RAINFUSION_ATTN advertised supports_prefix_kv_slicing, the model built a
|
|
padding mask here that _assert_metadata_compatible then rejected at the
|
|
first sparse layer. This drives the real backend class (not a fake flag)
|
|
through the packed path; both the sparse dispatch and the dense fallback
|
|
share this metadata construction and neither ever reads attn_mask.
|
|
"""
|
|
from vllm_omni.diffusion.attention.backends.rainfusion_attn import (
|
|
RainFusionAttentionBackend,
|
|
)
|
|
from vllm_omni.diffusion.models.minimax_h3.minimax_h3_transformer import (
|
|
MiniMaxH3Attention,
|
|
)
|
|
|
|
class FakeAttention(torch.nn.Module):
|
|
attn_backend = RainFusionAttentionBackend
|
|
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.metadata = None
|
|
|
|
def forward(self, query, key, value, metadata):
|
|
self.metadata = metadata
|
|
return query
|
|
|
|
attention = object.__new__(MiniMaxH3Attention)
|
|
torch.nn.Module.__init__(attention)
|
|
attention.attention = FakeAttention()
|
|
q = torch.randn(8, 2, 4)
|
|
|
|
attention._run_packed_attention(
|
|
q,
|
|
q,
|
|
q,
|
|
# One request: 5 valid rows padded to the 8-row alignment boundary.
|
|
cu_seqlens=torch.tensor([0, 5, 8], dtype=torch.int32),
|
|
max_seqlen=5,
|
|
packed_total=8,
|
|
)
|
|
|
|
metadata = attention.attention.metadata
|
|
assert metadata is not None
|
|
assert metadata.attn_mask is None
|
|
assert metadata.extra["valid_kv_length"] == 5
|
|
|
|
|
|
def test_reference_image_resize_contract():
|
|
from PIL import Image
|
|
|
|
from vllm_omni.model_executor.models.minimax_h3.preprocessing import (
|
|
resolve_minimax_h3_reference_image_shape,
|
|
)
|
|
|
|
assert resolve_minimax_h3_reference_image_shape(Image.new("RGB", (1080, 1440))) == (
|
|
2048,
|
|
2720,
|
|
)
|
|
with pytest.raises(ValueError, match="aspect ratio"):
|
|
resolve_minimax_h3_reference_image_shape(Image.new("RGB", (100, 501)))
|
|
|
|
|
|
def test_fl2va_supports_first_last_and_explicit_frame_index_contracts():
|
|
from vllm_omni.model_executor.models.minimax_h3.encoder_processing import (
|
|
_resolve_fl2va_keyframe_indices,
|
|
)
|
|
|
|
assert _resolve_fl2va_keyframe_indices({}, 1) == [0]
|
|
assert _resolve_fl2va_keyframe_indices({}, 2) == [0, -1]
|
|
assert _resolve_fl2va_keyframe_indices({"frame_index": -1}, 1) == [-1]
|
|
assert _resolve_fl2va_keyframe_indices({"frame_indices": [0, -1]}, 2) == [0, -1]
|
|
with pytest.raises(ValueError, match="frame_indices"):
|
|
_resolve_fl2va_keyframe_indices({"frame_indices": [0, 1]}, 2)
|
|
|
|
|
|
def test_minimax_h3_uses_the_official_output_canvas_policy():
|
|
from vllm_omni.model_executor.models.minimax_h3.preprocessing import (
|
|
resolve_minimax_h3_output_canvas,
|
|
)
|
|
|
|
assert resolve_minimax_h3_output_canvas(21 / 9, 768) == (672, 1536)
|
|
assert resolve_minimax_h3_output_canvas(16 / 9, 768) == (768, 1344)
|
|
assert resolve_minimax_h3_output_canvas(9 / 16, 768) == (1344, 768)
|
|
with pytest.raises(ValueError, match="short_edge"):
|
|
resolve_minimax_h3_output_canvas(16 / 9, 720)
|
|
|
|
|
|
def test_minimax_h3_accepts_sglang_auto_aspect_ratio_alias():
|
|
from vllm_omni.model_executor.models.minimax_h3.encoder_processing import resolve_minimax_h3_shape
|
|
|
|
sampling = SimpleNamespace(
|
|
fps=24,
|
|
num_frames=1,
|
|
height=None,
|
|
width=None,
|
|
extra_args={"duration": 5.0, "aspect_ratio": "auto"},
|
|
)
|
|
height, width, *_ = resolve_minimax_h3_shape("ref2va", sampling, None)
|
|
assert (height, width) == (768, 1344)
|
|
|
|
|
|
def test_minimax_h3_advertises_the_official_ref2va_image_limit():
|
|
from vllm_omni.diffusion.model_metadata import get_diffusion_model_metadata
|
|
|
|
assert get_diffusion_model_metadata("MiniMaxH3Pipeline").max_multimodal_image_inputs == 9
|
|
assert get_diffusion_model_metadata("MiniMaxH3Pipeline").supports_latent_mask_editing
|
|
|
|
|
|
def test_text_attention_routes_local_gqa_heads_through_sdpa_helper(monkeypatch):
|
|
from vllm_omni.diffusion.models.minimax_h3 import encoder as encoder_module
|
|
|
|
class FakeEncoderGroup:
|
|
rank_in_group = 0
|
|
world_size = 2
|
|
|
|
def all_reduce(self, tensor):
|
|
del tensor
|
|
|
|
config = SimpleNamespace(
|
|
hidden_size=8,
|
|
num_attention_heads=4,
|
|
num_key_value_heads=2,
|
|
head_dim=2,
|
|
rms_norm_eps=1e-6,
|
|
)
|
|
with (
|
|
patch("vllm.model_executor.parameter.get_tensor_model_parallel_rank", return_value=0),
|
|
patch("vllm.model_executor.parameter.get_tensor_model_parallel_world_size", return_value=1),
|
|
):
|
|
attention = encoder_module.MiniMaxH3Qwen3VLTextAttention(FakeEncoderGroup(), config, torch.float32)
|
|
# This CPU contract test only needs projection shapes and head reshaping.
|
|
# Avoid dispatching platform-specific norm and linear kernels (for example,
|
|
# rocm_unquantized_gemm when the test suite is collected on ROCm).
|
|
attention.q_norm = nn.Identity()
|
|
attention.k_norm = nn.Identity()
|
|
with torch.no_grad():
|
|
attention.qkv_proj.weight.zero_()
|
|
attention.o_proj = nn.Linear(
|
|
attention.qkv_proj.local_num_heads * attention.head_dim,
|
|
config.hidden_size,
|
|
bias=False,
|
|
)
|
|
attn_call: dict[str, object] = {}
|
|
|
|
def fake_attention(query, key, value):
|
|
attn_call.update(query_shape=query.shape, key_shape=key.shape, value_shape=value.shape)
|
|
return query
|
|
|
|
monkeypatch.setattr(encoder_module, "_scaled_dot_product_attention", fake_attention)
|
|
hidden_states = torch.randn(1, 3, config.hidden_size)
|
|
cos = torch.ones(1, 3, config.head_dim)
|
|
sin = torch.zeros_like(cos)
|
|
|
|
output = attention(hidden_states, (cos, sin))
|
|
|
|
assert attn_call["query_shape"] == (1, config.num_attention_heads // 2, 3, config.head_dim)
|
|
assert attn_call["key_shape"] == (1, config.num_key_value_heads // 2, 3, config.head_dim)
|
|
assert attn_call["value_shape"] == (1, config.num_key_value_heads // 2, 3, config.head_dim)
|
|
assert output.shape == hidden_states.shape
|
|
|
|
|
|
def test_text_attention_sdpa_helper_preserves_expanded_kv_fallback(monkeypatch):
|
|
from vllm_omni.diffusion.models.minimax_h3 import encoder as encoder_module
|
|
|
|
sdpa_call: dict[str, object] = {}
|
|
|
|
def fake_sdpa(query, key, value, **kwargs):
|
|
sdpa_call.update(query_shape=query.shape, key_shape=key.shape, value_shape=value.shape, kwargs=kwargs)
|
|
return query
|
|
|
|
monkeypatch.setattr(torch.nn.functional, "scaled_dot_product_attention", fake_sdpa)
|
|
output = encoder_module._scaled_dot_product_attention(
|
|
torch.randn(1, 4, 3, 8),
|
|
torch.randn(1, 2, 3, 8),
|
|
torch.randn(1, 2, 3, 8),
|
|
)
|
|
|
|
assert sdpa_call["query_shape"] == sdpa_call["key_shape"] == sdpa_call["value_shape"] == (1, 4, 3, 8)
|
|
assert sdpa_call["kwargs"] == {"dropout_p": 0.0, "is_causal": True}
|
|
assert output.shape == (1, 4, 3, 8)
|
|
|
|
|
|
def test_encoder_forward_uses_hook_compatible_encode_entrypoint():
|
|
from vllm_omni.diffusion.models.minimax_h3.encoder import (
|
|
MiniMaxH3Qwen3VLEncoder,
|
|
)
|
|
|
|
encoder = object.__new__(MiniMaxH3Qwen3VLEncoder)
|
|
torch.nn.Module.__init__(encoder)
|
|
expected = torch.ones(2, 3)
|
|
encoder.encode_ids = Mock(return_value=expected)
|
|
input_ids = torch.tensor([1, 2])
|
|
pixel_values = torch.ones(1, 4)
|
|
image_grid_thw = torch.tensor([[1, 1, 1]])
|
|
|
|
actual = encoder(
|
|
input_ids,
|
|
pixel_values=pixel_values,
|
|
image_grid_thw=image_grid_thw,
|
|
)
|
|
|
|
assert actual is expected
|
|
encoder.encode_ids.assert_called_once_with(
|
|
input_ids,
|
|
pixel_values=pixel_values,
|
|
image_grid_thw=image_grid_thw,
|
|
)
|
|
|
|
|
|
def test_encoder_forward_forwards_video_inputs():
|
|
from vllm_omni.diffusion.models.minimax_h3.encoder import (
|
|
MiniMaxH3Qwen3VLEncoder,
|
|
)
|
|
|
|
encoder = object.__new__(MiniMaxH3Qwen3VLEncoder)
|
|
torch.nn.Module.__init__(encoder)
|
|
expected = torch.ones(2, 3)
|
|
encoder.encode_ids = Mock(return_value=expected)
|
|
input_ids = torch.tensor([1, 2])
|
|
pixel_values_videos = torch.ones(1, 4)
|
|
video_grid_thw = torch.tensor([[2, 1, 1]])
|
|
|
|
actual = encoder(
|
|
input_ids,
|
|
pixel_values_videos=pixel_values_videos,
|
|
video_grid_thw=video_grid_thw,
|
|
)
|
|
|
|
assert actual is expected
|
|
encoder.encode_ids.assert_called_once_with(
|
|
input_ids,
|
|
pixel_values=None,
|
|
image_grid_thw=None,
|
|
pixel_values_videos=pixel_values_videos,
|
|
video_grid_thw=video_grid_thw,
|
|
)
|
|
|
|
|
|
def test_reference_video_shape_uses_h3_adapt_shape_policy():
|
|
from vllm_omni.model_executor.models.minimax_h3.reference_video import (
|
|
_reference_video_shape,
|
|
)
|
|
|
|
assert _reference_video_shape(1280, 720) == (1344, 768)
|
|
assert _reference_video_shape(3844, 2160) == (1344, 768)
|
|
|
|
|
|
def test_text_encoder_stub_constructs_without_group_or_weights():
|
|
from vllm_omni.diffusion.models.minimax_h3.encoder import (
|
|
MiniMaxH3Qwen3VLEncoder,
|
|
)
|
|
from vllm_omni.diffusion.models.minimax_h3.pipeline_minimax_h3 import (
|
|
MiniMaxH3Pipeline,
|
|
)
|
|
from vllm_omni.diffusion.offloader.component_utils import (
|
|
get_encoder_block_groups,
|
|
)
|
|
|
|
encoder = MiniMaxH3Qwen3VLEncoder(
|
|
"/nonexistent/text_encoder",
|
|
device=torch.device("cpu"),
|
|
load_model=False,
|
|
encoder_group=None,
|
|
)
|
|
assert not encoder.is_loaded
|
|
assert encoder.tp_size == 1
|
|
# The stub has no parameters, so it never contributes to the runner's
|
|
# strict missing-parameter check on non-encoder ranks.
|
|
assert list(encoder.named_parameters()) == []
|
|
assert (
|
|
get_encoder_block_groups(
|
|
encoder,
|
|
"text_encoder",
|
|
MiniMaxH3Pipeline._offload_plan,
|
|
strict=True,
|
|
)
|
|
== []
|
|
)
|
|
|
|
# An arbitrary empty module is not silently treated as a distributed
|
|
# stub; explicit plans still validate their declared paths.
|
|
with pytest.raises(ValueError, match=r"text_encoder\.vision\.blocks was not found"):
|
|
get_encoder_block_groups(
|
|
nn.Module(),
|
|
"text_encoder",
|
|
MiniMaxH3Pipeline._offload_plan,
|
|
strict=True,
|
|
)
|
|
|
|
|
|
def test_global_quant_config_is_shared_by_dit_and_encoder():
|
|
from vllm_omni.quantization import build_quant_config, resolve_component_quant_config
|
|
|
|
global_config = build_quant_config("fp8")
|
|
|
|
assert resolve_component_quant_config(global_config, "transformer") is global_config
|
|
assert resolve_component_quant_config(global_config, "text_encoder") is global_config
|
|
|
|
|
|
def test_minimax_h3_fp8_component_selection():
|
|
from vllm_omni.quantization import build_quant_config, resolve_component_quant_config
|
|
|
|
dit_only = build_quant_config({"transformer": {"method": "fp8"}})
|
|
encoder_only = build_quant_config({"text_encoder": {"method": "fp8"}})
|
|
|
|
assert resolve_component_quant_config(dit_only, "transformer").get_name() == "fp8"
|
|
assert resolve_component_quant_config(dit_only, "text_encoder") is None
|
|
assert resolve_component_quant_config(encoder_only, "transformer") is None
|
|
assert resolve_component_quant_config(encoder_only, "text_encoder").get_name() == "fp8"
|
|
|
|
|
|
def test_text_encoder_rejects_serialized_fp8():
|
|
from vllm_omni.diffusion.models.minimax_h3.encoder import _validate_encoder_quant_config
|
|
|
|
config = SimpleNamespace(
|
|
get_name=lambda: "fp8",
|
|
is_checkpoint_fp8_serialized=True,
|
|
)
|
|
with pytest.raises(ValueError, match="online FP8 only"):
|
|
_validate_encoder_quant_config(config)
|
|
|
|
|
|
def test_text_encoder_linear_delegates_quantization_to_vllm_factory():
|
|
from vllm.model_executor.layers.linear import UnquantizedLinearMethod
|
|
from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
|
|
|
|
import vllm_omni.diffusion.models.minimax_h3.encoder as encoder_module
|
|
|
|
group = SimpleNamespace(rank_in_group=0, world_size=1)
|
|
method = UnquantizedLinearMethod()
|
|
quant_config = create_autospec(QuantizationConfig, instance=True)
|
|
quant_config.online_quantization_config = None
|
|
quant_config.get_quant_method.return_value = method
|
|
# vLLM 0.30: no online quantization for this checkpoint-quantized layer
|
|
quant_config.online_quantization_config = None
|
|
prefix = "text_encoder.text_model.layers.0.mlp.gate_up_proj"
|
|
|
|
with (
|
|
patch("vllm.model_executor.parameter.get_tensor_model_parallel_rank", return_value=0),
|
|
patch("vllm.model_executor.parameter.get_tensor_model_parallel_world_size", return_value=1),
|
|
):
|
|
linear = encoder_module.MiniMaxH3Qwen3VLMergedColumnParallelLinear(
|
|
group,
|
|
input_size=4,
|
|
intermediate_size=8,
|
|
dtype=torch.bfloat16,
|
|
quant_config=quant_config,
|
|
prefix=prefix,
|
|
)
|
|
|
|
assert linear.quant_method is method
|
|
# vLLM 0.30 resolves the method through resolve_quant_method(), which calls
|
|
# get_quant_method(layer, prefix) POSITIONALLY; 0.29 passed prefix= by keyword.
|
|
quant_config.get_quant_method.assert_called_once_with(linear, prefix)
|
|
|
|
|
|
def test_no_offload_keeps_text_encoder_resident():
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
|
|
pipeline = object.__new__(MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
pipeline.od_config = SimpleNamespace(
|
|
enable_cpu_offload=False,
|
|
enable_layerwise_offload=False,
|
|
)
|
|
pipeline.text_encoder = Mock()
|
|
expected = torch.ones(2, 3)
|
|
pipeline.text_encoder.encode_ids.return_value = expected
|
|
input_ids = torch.tensor([1, 2])
|
|
|
|
actual = pipeline._encode_text_hidden(input_ids, {})
|
|
|
|
assert actual is expected
|
|
pipeline.text_encoder.load_to_device.assert_called_once_with()
|
|
pipeline.text_encoder.offload_to_cpu.assert_not_called()
|
|
|
|
|
|
def test_model_offload_uses_hooked_text_encoder_call():
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
|
|
pipeline = object.__new__(MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
pipeline.od_config = SimpleNamespace(
|
|
enable_cpu_offload=True,
|
|
enable_layerwise_offload=False,
|
|
)
|
|
expected = torch.ones(2, 3)
|
|
pipeline.text_encoder = Mock(return_value=expected)
|
|
pipeline._model_cpu_offload_modules = [pipeline.text_encoder]
|
|
input_ids = torch.tensor([1, 2])
|
|
vision_kwargs = {"pixel_values": torch.ones(1, 4)}
|
|
|
|
actual = pipeline._encode_text_hidden(input_ids, vision_kwargs)
|
|
|
|
assert actual is expected
|
|
pipeline.text_encoder.assert_called_once_with(input_ids, **vision_kwargs)
|
|
pipeline.text_encoder.load_to_device.assert_not_called()
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
("enable_layerwise", "enable_distributed"),
|
|
[(True, False), (False, True)],
|
|
)
|
|
def test_legacy_layer_offload_preserves_minimax_stage_lifecycle(enable_layerwise, enable_distributed):
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
|
|
pipeline = object.__new__(MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
pipeline.od_config = SimpleNamespace(
|
|
diffusion_offload_config=None,
|
|
enable_cpu_offload=False,
|
|
enable_layerwise_offload=enable_layerwise,
|
|
enable_distributed_layerwise_offload=enable_distributed,
|
|
)
|
|
pipeline.text_encoder = Mock()
|
|
expected = torch.ones(2, 3)
|
|
pipeline.text_encoder.encode_ids.return_value = expected
|
|
|
|
actual = pipeline._encode_text_hidden(torch.tensor([1, 2]), {})
|
|
|
|
assert actual is expected
|
|
pipeline.text_encoder.load_to_device.assert_called_once_with()
|
|
pipeline.text_encoder.offload_to_cpu.assert_called_once_with()
|
|
|
|
vae = Mock()
|
|
with pipeline._component_on_device(vae):
|
|
pass
|
|
vae.load_to_device.assert_called_once_with()
|
|
vae.offload_to_cpu.assert_called_once_with()
|
|
|
|
|
|
def test_layerwise_encoder_selection_releases_text_encoder():
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
|
|
pipeline = object.__new__(MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
pipeline.od_config = SimpleNamespace(
|
|
enable_cpu_offload=False,
|
|
enable_layerwise_offload=False,
|
|
enable_distributed_layerwise_offload=False,
|
|
diffusion_offload_config={
|
|
"mode": "layer",
|
|
"components": ["text_encoder"],
|
|
},
|
|
)
|
|
pipeline.text_encoder = Mock()
|
|
expected = torch.ones(2, 3)
|
|
pipeline.text_encoder.encode_ids.return_value = expected
|
|
|
|
actual = pipeline._encode_text_hidden(torch.tensor([1, 2]), {})
|
|
|
|
assert actual is expected
|
|
pipeline.text_encoder.load_to_device.assert_called_once_with()
|
|
pipeline.text_encoder.offload_to_cpu.assert_called_once_with()
|
|
|
|
|
|
def test_layerwise_dit_only_keeps_text_encoder_resident():
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
|
|
pipeline = object.__new__(MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
pipeline.od_config = SimpleNamespace(
|
|
enable_cpu_offload=False,
|
|
enable_layerwise_offload=False,
|
|
enable_distributed_layerwise_offload=False,
|
|
diffusion_offload_config={"mode": "layer", "components": ["dit"]},
|
|
)
|
|
pipeline.text_encoder = Mock()
|
|
expected = torch.ones(2, 3)
|
|
pipeline.text_encoder.encode_ids.return_value = expected
|
|
|
|
actual = pipeline._encode_text_hidden(torch.tensor([1, 2]), {})
|
|
|
|
assert actual is expected
|
|
pipeline.text_encoder.load_to_device.assert_called_once_with()
|
|
pipeline.text_encoder.offload_to_cpu.assert_not_called()
|
|
|
|
|
|
def test_dit_encoder_selection_keeps_vae_resident():
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
|
|
pipeline = object.__new__(MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
pipeline.od_config = SimpleNamespace(
|
|
enable_cpu_offload=False,
|
|
enable_layerwise_offload=False,
|
|
enable_distributed_layerwise_offload=False,
|
|
diffusion_offload_config={
|
|
"mode": "layer",
|
|
"components": ["dit", "text_encoder"],
|
|
},
|
|
)
|
|
component = Mock()
|
|
|
|
with pipeline._component_on_device(component):
|
|
pass
|
|
|
|
component.load_to_device.assert_not_called()
|
|
component.offload_to_cpu.assert_not_called()
|
|
|
|
|
|
def test_distributed_layerwise_resident_blocks_are_stage_scoped():
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
|
|
pipeline = object.__new__(MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
controller = Mock()
|
|
pipeline._dlo_residency_controller = controller
|
|
|
|
with pipeline._resident_dit_layers_on_device():
|
|
controller.load_resident_layers.assert_called_once_with()
|
|
controller.offload_resident_layers.assert_not_called()
|
|
|
|
controller.offload_resident_layers.assert_called_once_with()
|
|
|
|
|
|
def test_distributed_layerwise_resident_blocks_release_on_failure():
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
|
|
pipeline = object.__new__(MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
controller = Mock()
|
|
pipeline._dlo_residency_controller = controller
|
|
|
|
with pytest.raises(RuntimeError, match="denoise failed"):
|
|
with pipeline._resident_dit_layers_on_device():
|
|
raise RuntimeError("denoise failed")
|
|
|
|
controller.load_resident_layers.assert_called_once_with()
|
|
controller.offload_resident_layers.assert_called_once_with()
|
|
|
|
|
|
def test_distributed_layerwise_resident_blocks_can_be_skipped():
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
|
|
pipeline = object.__new__(MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
controller = Mock()
|
|
pipeline._dlo_residency_controller = controller
|
|
|
|
with pipeline._resident_dit_layers_on_device(enabled=False):
|
|
pass
|
|
|
|
controller.load_resident_layers.assert_not_called()
|
|
controller.offload_resident_layers.assert_not_called()
|
|
|
|
|
|
@pytest.mark.skip(reason="encoder layerwise API is not part of this branch")
|
|
def test_encoder_layerwise_offload_keeps_tp_blocks_rank_local(monkeypatch):
|
|
from vllm_omni.diffusion.models.minimax_h3.encoder import (
|
|
MiniMaxH3Qwen3VLEncoder,
|
|
)
|
|
|
|
class Stack(torch.nn.Module):
|
|
def __init__(self, count):
|
|
super().__init__()
|
|
self.blocks = torch.nn.ModuleList([torch.nn.Linear(2, 2) for _ in range(count)])
|
|
|
|
class TextStack(torch.nn.Module):
|
|
def __init__(self, count):
|
|
super().__init__()
|
|
self.layers = torch.nn.ModuleList([torch.nn.Linear(2, 2) for _ in range(count)])
|
|
|
|
hooks = []
|
|
|
|
def fake_apply(block, next_block, device, stream, pin_memory):
|
|
hook = SimpleNamespace(
|
|
block=block,
|
|
next_block=next_block,
|
|
device=device,
|
|
stream=stream,
|
|
pin_memory=pin_memory,
|
|
_prev_hook=None,
|
|
offload_layer=Mock(),
|
|
)
|
|
hooks.append(hook)
|
|
return hook
|
|
|
|
monkeypatch.setattr(
|
|
"vllm_omni.diffusion.offloader.layerwise_backend.apply_block_hook",
|
|
fake_apply,
|
|
)
|
|
monkeypatch.setattr(
|
|
"vllm_omni.platforms.current_omni_platform.Stream",
|
|
Mock(return_value="copy-stream"),
|
|
)
|
|
encoder = object.__new__(MiniMaxH3Qwen3VLEncoder)
|
|
torch.nn.Module.__init__(encoder)
|
|
encoder.device_target = torch.device("cpu")
|
|
encoder.vision = Stack(2)
|
|
encoder.text_model = TextStack(3)
|
|
|
|
encoder.enable_omni_layerwise_offload(pin_memory=False)
|
|
|
|
assert len(hooks) == 5
|
|
assert all(hook._prev_hook is not None for hook in hooks)
|
|
assert all(hook.device == torch.device("cpu") for hook in hooks)
|
|
|
|
|
|
def test_video_vae_keeps_reference_fp32_weights(monkeypatch):
|
|
from vllm_omni.diffusion.models.minimax_h3 import vae as vae_module
|
|
|
|
class FakeRemote(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.model = torch.nn.Linear(1, 1).half()
|
|
|
|
monkeypatch.setattr(
|
|
vae_module,
|
|
"_load_component_config",
|
|
lambda _path: {
|
|
"latent_channels": 1,
|
|
"latents_mean": [0.0],
|
|
"latents_std": [1.0],
|
|
},
|
|
)
|
|
monkeypatch.setattr(
|
|
vae_module,
|
|
"_load_remote_component",
|
|
lambda _path, _config, *, trust_remote_code: FakeRemote(),
|
|
)
|
|
|
|
video_vae = vae_module.MiniMaxH3VideoVAE(
|
|
"unused",
|
|
device=torch.device("cpu"),
|
|
)
|
|
|
|
assert next(video_vae.parameters()).dtype == torch.float32
|
|
|
|
|
|
@pytest.mark.parametrize("fail_encode", [False, True])
|
|
def test_keyframe_encode_enters_backend_context(monkeypatch, fail_encode):
|
|
from PIL import Image
|
|
|
|
from vllm_omni.diffusion.models.minimax_h3 import vae as vae_module
|
|
|
|
events = []
|
|
input_image = Image.new("RGB", (32, 32))
|
|
|
|
@contextmanager
|
|
def track_context(device):
|
|
assert device == torch.device("cpu")
|
|
events.append("enter")
|
|
try:
|
|
yield
|
|
finally:
|
|
events.append("exit")
|
|
|
|
class FakeModel(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.weight = torch.nn.Parameter(torch.ones(1))
|
|
self.parallel_tiling = True
|
|
|
|
def encode_images(self, image, *, use_fp16_latent):
|
|
assert events == ["enter"]
|
|
assert image is input_image
|
|
assert use_fp16_latent
|
|
events.append("encode")
|
|
if fail_encode:
|
|
raise RuntimeError("encode failed")
|
|
return [torch.ones(1, 1, 2, 2)]
|
|
|
|
video_vae = object.__new__(vae_module.MiniMaxH3VideoVAE)
|
|
torch.nn.Module.__init__(video_vae)
|
|
video_vae.model = FakeModel()
|
|
video_vae.config_dict = {
|
|
"latent_channels": 1,
|
|
"latents_mean": [0.0],
|
|
"latents_std": [1.0],
|
|
}
|
|
monkeypatch.setattr(vae_module, "_minimax_h3_keyframe_encode_context", track_context)
|
|
|
|
if fail_encode:
|
|
with pytest.raises(RuntimeError, match="encode failed"):
|
|
video_vae.encode_image(input_image)
|
|
else:
|
|
rows = video_vae.encode_image(input_image)
|
|
torch.testing.assert_close(rows, torch.ones(1, 4))
|
|
|
|
assert events == ["enter", "encode", "exit"]
|
|
assert video_vae.model.parallel_tiling
|
|
|
|
|
|
def test_video_vae_can_load_on_cpu_for_staged_gpu_residency(monkeypatch):
|
|
from vllm_omni.diffusion.models.minimax_h3 import vae as vae_module
|
|
|
|
class FakeRemote(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.model = torch.nn.Linear(1, 1)
|
|
|
|
monkeypatch.setattr(
|
|
vae_module,
|
|
"_load_component_config",
|
|
lambda _path: {
|
|
"latent_channels": 1,
|
|
"latents_mean": [0.0],
|
|
"latents_std": [1.0],
|
|
},
|
|
)
|
|
monkeypatch.setattr(
|
|
vae_module,
|
|
"_load_remote_component",
|
|
lambda _path, _config, *, trust_remote_code: FakeRemote(),
|
|
)
|
|
|
|
video_vae = vae_module.MiniMaxH3VideoVAE(
|
|
"unused",
|
|
device=torch.device("meta"),
|
|
load_device=torch.device("cpu"),
|
|
)
|
|
|
|
assert next(video_vae.parameters()).device.type == "cpu"
|
|
assert video_vae._device_target.type == "meta"
|
|
|
|
|
|
def test_video_vae_uses_snapshot_stager_for_accelerator_target(monkeypatch):
|
|
from vllm_omni.diffusion.models.minimax_h3 import vae as vae_module
|
|
|
|
class FakeRemote(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.model = torch.nn.Linear(1, 1)
|
|
|
|
stager = Mock()
|
|
stager_cls = Mock(return_value=stager)
|
|
monkeypatch.setattr(
|
|
vae_module,
|
|
"_load_component_config",
|
|
lambda _path: {
|
|
"latent_channels": 1,
|
|
"latents_mean": [0.0],
|
|
"latents_std": [1.0],
|
|
},
|
|
)
|
|
monkeypatch.setattr(
|
|
vae_module,
|
|
"_load_remote_component",
|
|
lambda _path, _config, *, trust_remote_code: FakeRemote(),
|
|
)
|
|
monkeypatch.setattr(vae_module, "PinnedModuleStager", stager_cls)
|
|
|
|
video_vae = vae_module.MiniMaxH3VideoVAE(
|
|
"unused",
|
|
device=torch.device("cuda"),
|
|
load_device=torch.device("cpu"),
|
|
)
|
|
video_vae.load_to_device()
|
|
video_vae.offload_to_cpu()
|
|
|
|
stager_cls.assert_called_once_with(
|
|
video_vae.remote,
|
|
torch.device("cuda"),
|
|
pin_memory=True,
|
|
)
|
|
stager.load.assert_called_once_with()
|
|
stager.offload.assert_called_once_with()
|
|
|
|
|
|
def test_remote_component_refuses_untrusted_checkpoint_code(monkeypatch):
|
|
from vllm_omni.diffusion.models.minimax_h3 import vae as vae_module
|
|
|
|
loaded = []
|
|
|
|
def spy_get_class(class_reference, component_path):
|
|
loaded.append((class_reference, component_path))
|
|
raise AssertionError("checkpoint code must not be resolved when untrusted")
|
|
|
|
monkeypatch.setattr(vae_module, "get_class_from_dynamic_module", spy_get_class)
|
|
|
|
with pytest.raises(ValueError, match="--trust-remote-code"):
|
|
vae_module._load_remote_component(
|
|
"/checkpoints/video_vae",
|
|
{"auto_map": {"AutoModel": "modeling_vae.VideoVAE"}},
|
|
trust_remote_code=False,
|
|
)
|
|
|
|
assert loaded == []
|
|
|
|
|
|
def test_remote_component_resolves_checkpoint_code_when_trusted(monkeypatch):
|
|
from vllm_omni.diffusion.models.minimax_h3 import vae as vae_module
|
|
|
|
class FakeComponent(torch.nn.Module):
|
|
@classmethod
|
|
def from_pretrained(cls, component_path):
|
|
module = cls()
|
|
module.loaded_from = component_path
|
|
return module
|
|
|
|
loaded = []
|
|
|
|
def spy_get_class(class_reference, component_path):
|
|
loaded.append((class_reference, component_path))
|
|
return FakeComponent
|
|
|
|
monkeypatch.setattr(vae_module, "get_class_from_dynamic_module", spy_get_class)
|
|
|
|
module = vae_module._load_remote_component(
|
|
"/checkpoints/video_vae",
|
|
{"auto_map": {"AutoModel": "modeling_vae.VideoVAE"}},
|
|
trust_remote_code=True,
|
|
)
|
|
|
|
# Negative control: the trusted path stays byte-for-byte the pre-gate call.
|
|
assert loaded == [("modeling_vae.VideoVAE", "/checkpoints/video_vae")]
|
|
assert module.loaded_from == "/checkpoints/video_vae"
|
|
|
|
|
|
@pytest.mark.parametrize("vae_cls_name", ["MiniMaxH3VideoVAE", "MiniMaxH3AudioVAE"])
|
|
@pytest.mark.parametrize("trusted", [True, False])
|
|
def test_vae_forwards_trust_remote_code_to_loader(monkeypatch, vae_cls_name, trusted):
|
|
from vllm_omni.diffusion.models.minimax_h3 import vae as vae_module
|
|
|
|
class FakeRemote(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.model = torch.nn.Linear(1, 1)
|
|
|
|
forwarded = []
|
|
|
|
def spy_load(_component_path, _config, *, trust_remote_code):
|
|
forwarded.append(trust_remote_code)
|
|
return FakeRemote()
|
|
|
|
monkeypatch.setattr(
|
|
vae_module,
|
|
"_load_component_config",
|
|
lambda _path: {
|
|
"latent_channels": 1,
|
|
"latents_mean": [0.0],
|
|
"latents_std": [1.0],
|
|
"sample_rate": 32000,
|
|
},
|
|
)
|
|
monkeypatch.setattr(vae_module, "_load_remote_component", spy_load)
|
|
|
|
vae_cls = getattr(vae_module, vae_cls_name)
|
|
vae_cls(
|
|
"unused",
|
|
device=torch.device("cpu"),
|
|
trust_remote_code=trusted,
|
|
)
|
|
|
|
assert forwarded == [trusted]
|
|
|
|
|
|
@pytest.mark.parametrize("vae_cls_name", ["MiniMaxH3VideoVAE", "MiniMaxH3AudioVAE"])
|
|
def test_vae_defaults_to_refusing_checkpoint_code(monkeypatch, vae_cls_name):
|
|
from vllm_omni.diffusion.models.minimax_h3 import vae as vae_module
|
|
|
|
forwarded = []
|
|
|
|
def spy_load(_component_path, _config, *, trust_remote_code):
|
|
forwarded.append(trust_remote_code)
|
|
raise AssertionError("unreachable")
|
|
|
|
monkeypatch.setattr(
|
|
vae_module,
|
|
"_load_component_config",
|
|
lambda _path: {
|
|
"latent_channels": 1,
|
|
"latents_mean": [0.0],
|
|
"latents_std": [1.0],
|
|
"sample_rate": 32000,
|
|
},
|
|
)
|
|
monkeypatch.setattr(vae_module, "_load_remote_component", spy_load)
|
|
|
|
vae_cls = getattr(vae_module, vae_cls_name)
|
|
with pytest.raises(AssertionError, match="unreachable"):
|
|
vae_cls("unused", device=torch.device("cpu"))
|
|
|
|
assert forwarded == [False]
|
|
|
|
|
|
def test_video_vae_encode_uses_configured_parallel_tiling():
|
|
from vllm_omni.diffusion.models.minimax_h3.vae import (
|
|
MiniMaxH3VideoVAE,
|
|
)
|
|
|
|
class FakeModel(torch.nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.weight = torch.nn.Parameter(torch.ones(1))
|
|
self.parallel_tiling = True
|
|
self.encode_calls = []
|
|
|
|
def encode_videos(self, frames, *, use_fp16_latent):
|
|
assert self.parallel_tiling
|
|
self.encode_calls.append((frames, use_fp16_latent))
|
|
return [torch.ones(1, 1, 2, 2, 2)]
|
|
|
|
video_vae = object.__new__(MiniMaxH3VideoVAE)
|
|
torch.nn.Module.__init__(video_vae)
|
|
video_vae.model = FakeModel()
|
|
video_vae.config_dict = {
|
|
"latent_channels": 1,
|
|
"latents_mean": [0.0],
|
|
"latents_std": [1.0],
|
|
}
|
|
|
|
rows, shape = video_vae.encode_video("frames")
|
|
|
|
assert video_vae.model.parallel_tiling
|
|
assert video_vae.model.encode_calls == [("frames", True)]
|
|
assert rows.shape == (2, 4)
|
|
assert shape == (2, 2, 2)
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
("case", "extra", "image_count", "expected"),
|
|
[
|
|
("F1", {"frame_index": -1}, 1, [-1]),
|
|
("F2", {"frame_indices": [0, -1]}, 2, [0, -1]),
|
|
],
|
|
)
|
|
def test_f1_f2_official_fl2va_keyframe_matrix(case, extra, image_count, expected):
|
|
from vllm_omni.model_executor.models.minimax_h3.encoder_processing import (
|
|
_resolve_fl2va_keyframe_indices,
|
|
)
|
|
|
|
assert case in {"F1", "F2"}
|
|
assert _resolve_fl2va_keyframe_indices(extra, image_count) == expected
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
("case", "counts"),
|
|
[
|
|
("R1", (3, 0, 0)),
|
|
("R2", (1, 0, 0)),
|
|
("R3", (1, 1, 0)),
|
|
("R4", (0, 1, 1)),
|
|
("R5", (1, 1, 1)),
|
|
("R6", (1, 0, 2)),
|
|
],
|
|
)
|
|
def test_r1_r6_ref2va_reference_count_matrix(case, counts):
|
|
from vllm_omni.model_executor.models.minimax_h3.encoder_processing import (
|
|
_validate_ref2va_reference_counts,
|
|
)
|
|
|
|
assert case.startswith("R")
|
|
_validate_ref2va_reference_counts(*counts)
|
|
|
|
|
|
def test_ref2va_reference_count_validation_preserves_client_error_metadata():
|
|
from vllm_omni.errors import OmniClientError
|
|
from vllm_omni.model_executor.models.minimax_h3.encoder_processing import (
|
|
_validate_ref2va_reference_counts,
|
|
)
|
|
|
|
with pytest.raises(OmniClientError, match="at least one image or video"):
|
|
_validate_ref2va_reference_counts(0, 0, 0)
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
("case", "start_time", "expected_duration"),
|
|
[
|
|
("R7", None, 10.0),
|
|
("R8", 4.0, 6.0),
|
|
],
|
|
)
|
|
def test_r7_r8_ref2va_video_segment_matrix(monkeypatch, tmp_path, case, start_time, expected_duration):
|
|
from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
|
|
|
|
source = tmp_path / f"{case}.mp4"
|
|
source.touch()
|
|
metadata = {
|
|
"width": 1280,
|
|
"height": 720,
|
|
"fps": 24.0,
|
|
"frame_count": 240,
|
|
"duration": 10.0,
|
|
"format_names": ("mp4",),
|
|
"video_codec": "h264",
|
|
"audio_codecs": (),
|
|
"file_size": 1024,
|
|
}
|
|
transcode_calls: list[Any] = []
|
|
monkeypatch.setattr(reference_video_module, "_probe_video", lambda _path: metadata)
|
|
monkeypatch.setattr(
|
|
reference_video_module,
|
|
"_transcode_reference_video",
|
|
lambda source, **kwargs: _append_and_return(transcode_calls, (source, kwargs), "prepared.mp4"),
|
|
)
|
|
|
|
prepared = reference_video_module.prepare_reference_videos(
|
|
[str(source)],
|
|
target_frame_count=209,
|
|
workdir=str(tmp_path / "work"),
|
|
start_time_seconds=start_time,
|
|
)
|
|
|
|
assert prepared[0]["duration_seconds"] == pytest.approx(expected_duration)
|
|
assert prepared[0]["audio_duration_seconds"] == pytest.approx(min(expected_duration, 209 / 24.0))
|
|
assert prepared[0]["start_time_seconds"] == pytest.approx(start_time or 0.0)
|
|
assert transcode_calls[0][1]["duration_seconds"] == pytest.approx(expected_duration)
|
|
assert transcode_calls[0][1]["target_frame_count"] == 209
|
|
|
|
|
|
def test_ref2va_transcode_zero_frame_count_keeps_video_stream(monkeypatch, tmp_path):
|
|
from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
|
|
|
|
commands: list[list[str]] = []
|
|
monkeypatch.setattr(
|
|
reference_video_module.subprocess,
|
|
"run",
|
|
lambda command, **_kwargs: commands.append(command),
|
|
)
|
|
|
|
reference_video_module._transcode_reference_video(
|
|
"input.mp4",
|
|
target_width=768,
|
|
target_height=448,
|
|
target_frame_count=0,
|
|
workdir=str(tmp_path),
|
|
duration_seconds=4.0,
|
|
)
|
|
|
|
assert "-frames:v" not in commands[0]
|
|
|
|
|
|
def test_prepared_reference_video_descriptor_round_trip(tmp_path):
|
|
from vllm_omni.model_executor.models.minimax_h3.reference_video import (
|
|
deserialize_prepared_reference_videos,
|
|
serialize_prepared_reference_videos,
|
|
)
|
|
|
|
item = {
|
|
"original_path": str(tmp_path / "source.mp4"),
|
|
"prepared_path": str(tmp_path / "prepared.mp4"),
|
|
"input_has_audio": True,
|
|
"width": 1344,
|
|
"height": 768,
|
|
"start_time_seconds": 0.0,
|
|
"duration_seconds": 5.2,
|
|
"audio_duration_seconds": 5.2,
|
|
}
|
|
|
|
descriptor = serialize_prepared_reference_videos([item], str(tmp_path))
|
|
|
|
assert deserialize_prepared_reference_videos(descriptor) == (str(tmp_path), [item])
|
|
|
|
|
|
def test_prepared_reference_video_descriptor_rejects_unknown_fields():
|
|
from vllm_omni.model_executor.models.minimax_h3.reference_video import (
|
|
deserialize_prepared_reference_videos,
|
|
)
|
|
|
|
with pytest.raises(ValueError, match="invalid MiniMax H3 prepared-reference-video descriptor"):
|
|
deserialize_prepared_reference_videos('{"artifact_dir":"/tmp","videos":[{"path":"x"}]}')
|
|
|
|
|
|
def test_ref2va_qwen_sampling_uses_one_selective_decode(monkeypatch):
|
|
from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
|
|
|
|
decoded = np.arange(3 * 2 * 2 * 3, dtype=np.uint8).reshape(3, 2, 2, 3)
|
|
decode_calls: list[Any] = []
|
|
monkeypatch.setattr(
|
|
reference_video_module,
|
|
"_probe_video",
|
|
lambda _path: {
|
|
"frame_count": 25,
|
|
"width": 2,
|
|
"height": 2,
|
|
},
|
|
)
|
|
monkeypatch.setattr(
|
|
reference_video_module,
|
|
"_decode_video_frames_ffmpeg",
|
|
lambda path, **kwargs: _append_and_return(decode_calls, (path, kwargs), decoded),
|
|
)
|
|
|
|
sampled = reference_video_module.sample_reference_video_frames("prepared.mp4")
|
|
|
|
assert decode_calls == [
|
|
(
|
|
"prepared.mp4",
|
|
{
|
|
"frame_count": 25,
|
|
"indices": [0, 12, 24],
|
|
"width": 2,
|
|
"height": 2,
|
|
},
|
|
)
|
|
]
|
|
for frame, expected in zip(sampled["frames"], decoded, strict=True):
|
|
np.testing.assert_array_equal(frame, expected)
|
|
assert sampled["block_timestamps"] == [0.25, 1.0]
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
("indices", "frame_count"),
|
|
[([0, 12], 25), (None, 2)],
|
|
)
|
|
def test_ref2va_ffmpeg_decode_reads_exact_rgb_frames(monkeypatch, indices, frame_count):
|
|
from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
|
|
|
|
expected = np.arange(2 * 2 * 3 * 3, dtype=np.uint8).reshape(2, 2, 3, 3)
|
|
observed = []
|
|
|
|
def fake_run(command, **kwargs):
|
|
observed.append((command, kwargs))
|
|
return SimpleNamespace(stdout=expected.tobytes())
|
|
|
|
monkeypatch.setattr(reference_video_module.subprocess, "run", fake_run)
|
|
|
|
actual = reference_video_module._decode_video_frames_ffmpeg(
|
|
"prepared.mp4",
|
|
frame_count=frame_count,
|
|
indices=indices,
|
|
width=3,
|
|
height=2,
|
|
)
|
|
|
|
np.testing.assert_array_equal(actual, expected)
|
|
command, kwargs = observed[0]
|
|
assert command[0] == "ffmpeg"
|
|
assert command[command.index("-frames:v") + 1] == "2"
|
|
if indices is None:
|
|
assert "-vf" not in command
|
|
else:
|
|
assert command[command.index("-vf") + 1] == "select=eq(n\\,0)+eq(n\\,12)"
|
|
assert command[-1] == "pipe:1"
|
|
assert kwargs == {"check": True, "capture_output": True}
|
|
|
|
|
|
def test_ref2va_load_video_frames_uses_ffmpeg_without_decord(monkeypatch):
|
|
from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
|
|
|
|
expected = np.zeros((2, 2, 3, 3), dtype=np.uint8)
|
|
decode_calls: list[Any] = []
|
|
monkeypatch.setitem(sys.modules, "decord", None)
|
|
monkeypatch.setattr(
|
|
reference_video_module,
|
|
"_probe_video",
|
|
lambda _path: {
|
|
"frame_count": 2,
|
|
"width": 3,
|
|
"height": 2,
|
|
},
|
|
)
|
|
monkeypatch.setattr(
|
|
reference_video_module,
|
|
"_decode_video_frames_ffmpeg",
|
|
lambda path, **kwargs: _append_and_return(decode_calls, (path, kwargs), expected),
|
|
)
|
|
|
|
actual = reference_video_module.load_video_frames("prepared.mp4")
|
|
|
|
assert actual is expected
|
|
assert decode_calls == [
|
|
(
|
|
"prepared.mp4",
|
|
{
|
|
"frame_count": 2,
|
|
"width": 3,
|
|
"height": 2,
|
|
},
|
|
)
|
|
]
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
("kwargs", "message"),
|
|
[
|
|
({"frame_count": 0, "width": 3, "height": 2}, "video has no frames"),
|
|
({"frame_count": 1, "width": 0, "height": 2}, "invalid dimensions"),
|
|
(
|
|
{"frame_count": 1, "indices": [], "width": 3, "height": 2},
|
|
"invalid frame indices",
|
|
),
|
|
],
|
|
)
|
|
def test_ref2va_ffmpeg_decode_rejects_invalid_metadata(kwargs, message):
|
|
from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
|
|
from vllm_omni.errors import OmniClientError
|
|
|
|
with pytest.raises(OmniClientError, match=message):
|
|
reference_video_module._decode_video_frames_ffmpeg("prepared.mp4", **kwargs)
|
|
|
|
|
|
def test_ref2va_probe_uses_container_frame_count_without_scan(monkeypatch):
|
|
from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
|
|
|
|
calls: list[Any] = []
|
|
probe = {
|
|
"streams": [
|
|
{
|
|
"codec_type": "video",
|
|
"width": 4,
|
|
"height": 2,
|
|
"r_frame_rate": "24/1",
|
|
"nb_frames": "3",
|
|
"duration": "0.125",
|
|
"sample_aspect_ratio": "1:1",
|
|
}
|
|
],
|
|
"format": {"duration": "0.125", "size": "123"},
|
|
}
|
|
monkeypatch.setattr(
|
|
reference_video_module.subprocess,
|
|
"run",
|
|
lambda command, **kwargs: _append_and_return(
|
|
calls,
|
|
(command, kwargs),
|
|
SimpleNamespace(stdout=json.dumps(probe)),
|
|
),
|
|
)
|
|
|
|
metadata = reference_video_module._probe_video("prepared.mp4")
|
|
|
|
assert metadata["frame_count"] == 3
|
|
assert len(calls) == 1
|
|
assert "-count_frames" not in calls[0][0]
|
|
|
|
|
|
def test_ref2va_probe_counts_frames_only_when_metadata_is_missing(monkeypatch):
|
|
from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
|
|
|
|
calls: list[Any] = []
|
|
stream = {
|
|
"codec_type": "video",
|
|
"width": 4,
|
|
"height": 2,
|
|
"r_frame_rate": "24/1",
|
|
"duration": "0.125",
|
|
"sample_aspect_ratio": "1:1",
|
|
}
|
|
first_probe = {
|
|
"streams": [stream],
|
|
"format": {"duration": "0.125", "size": "123"},
|
|
}
|
|
second_probe = {
|
|
"streams": [{**stream, "nb_read_frames": "3"}],
|
|
"format": {"duration": "0.125", "size": "123"},
|
|
}
|
|
monkeypatch.setattr(
|
|
reference_video_module.subprocess,
|
|
"run",
|
|
lambda command, **kwargs: _append_and_return(
|
|
calls,
|
|
(command, kwargs),
|
|
SimpleNamespace(stdout=json.dumps(first_probe if not calls else second_probe)),
|
|
),
|
|
)
|
|
|
|
metadata = reference_video_module._probe_video("prepared.mp4")
|
|
|
|
assert metadata["frame_count"] == 3
|
|
assert len(calls) == 2
|
|
assert "-count_frames" in calls[1][0]
|
|
|
|
|
|
def test_ref2va_ffmpeg_decode_rejects_incomplete_output(monkeypatch):
|
|
from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
|
|
from vllm_omni.errors import OmniClientError
|
|
|
|
monkeypatch.setattr(
|
|
reference_video_module.subprocess,
|
|
"run",
|
|
lambda *_args, **_kwargs: SimpleNamespace(stdout=b"short"),
|
|
)
|
|
|
|
with pytest.raises(OmniClientError, match="decoded 5 bytes.*expected 18"):
|
|
reference_video_module._decode_video_frames_ffmpeg(
|
|
"prepared.mp4",
|
|
frame_count=1,
|
|
width=3,
|
|
height=2,
|
|
)
|
|
|
|
|
|
def test_ref2va_two_video_recipe_tolerates_container_rounding(monkeypatch, tmp_path):
|
|
from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
|
|
|
|
first = tmp_path / "first.mp4"
|
|
second = tmp_path / "second.mp4"
|
|
first.touch()
|
|
second.touch()
|
|
metadata = {
|
|
str(first): {
|
|
"width": 1280,
|
|
"height": 720,
|
|
"fps": 24.0,
|
|
"frame_count": 241,
|
|
"duration": 10.041667,
|
|
"format_names": ("mp4",),
|
|
"video_codec": "h264",
|
|
"audio_codecs": (),
|
|
"file_size": 1024,
|
|
},
|
|
str(second): {
|
|
"width": 1280,
|
|
"height": 720,
|
|
"fps": 24.0,
|
|
"frame_count": 120,
|
|
"duration": 4.966667,
|
|
"format_names": ("mp4",),
|
|
"video_codec": "h264",
|
|
"audio_codecs": (),
|
|
"file_size": 1024,
|
|
},
|
|
}
|
|
transcode_calls: list[Any] = []
|
|
monkeypatch.setattr(reference_video_module, "_probe_video", lambda path: metadata[str(path)])
|
|
monkeypatch.setattr(
|
|
reference_video_module,
|
|
"_transcode_reference_video",
|
|
lambda source, **kwargs: _append_and_return(
|
|
transcode_calls,
|
|
(source, kwargs),
|
|
f"{source}.prepared.mp4",
|
|
),
|
|
)
|
|
|
|
prepared = reference_video_module.prepare_reference_videos(
|
|
[str(first), str(second)],
|
|
target_frame_count=124,
|
|
workdir=str(tmp_path / "work"),
|
|
)
|
|
|
|
assert [item["duration_seconds"] for item in prepared] == pytest.approx([10.041667, 4.958333])
|
|
assert sum(item["duration_seconds"] for item in prepared) == pytest.approx(15.0)
|
|
assert transcode_calls[1][1]["duration_seconds"] == pytest.approx(4.958333)
|
|
|
|
|
|
def test_ref2va_two_video_recipe_rejects_real_duration_overflow(monkeypatch, tmp_path):
|
|
from vllm_omni.diffusion.models.minimax_h3 import reference_video as reference_video_module
|
|
from vllm_omni.errors import OmniClientError
|
|
|
|
first = tmp_path / "first.mp4"
|
|
second = tmp_path / "second.mp4"
|
|
first.touch()
|
|
second.touch()
|
|
metadata = {
|
|
str(first): {
|
|
"width": 1280,
|
|
"height": 720,
|
|
"fps": 24.0,
|
|
"frame_count": 240,
|
|
"duration": 10.0,
|
|
"format_names": ("mp4",),
|
|
"video_codec": "h264",
|
|
"audio_codecs": (),
|
|
"file_size": 1024,
|
|
},
|
|
str(second): {
|
|
"width": 1280,
|
|
"height": 720,
|
|
"fps": 24.0,
|
|
"frame_count": 120,
|
|
"duration": 5.02,
|
|
"format_names": ("mp4",),
|
|
"video_codec": "h264",
|
|
"audio_codecs": (),
|
|
"file_size": 1024,
|
|
},
|
|
}
|
|
monkeypatch.setattr(reference_video_module, "_probe_video", lambda path: metadata[str(path)])
|
|
monkeypatch.setattr(reference_video_module, "_transcode_reference_video", lambda source, **kwargs: "prepared.mp4")
|
|
|
|
with pytest.raises(OmniClientError, match="15 seconds"):
|
|
reference_video_module.prepare_reference_videos(
|
|
[str(first), str(second)],
|
|
target_frame_count=124,
|
|
workdir=str(tmp_path / "work"),
|
|
)
|
|
|
|
|
|
@pytest.mark.parametrize("duration", [4.0, 15.0])
|
|
def test_g2_output_duration_accepts_official_boundaries(duration):
|
|
from vllm_omni.model_executor.models.minimax_h3.encoder_processing import resolve_minimax_h3_shape
|
|
|
|
sampling = SimpleNamespace(
|
|
fps=24,
|
|
num_frames=1,
|
|
height=None,
|
|
width=None,
|
|
extra_args={"duration": duration},
|
|
)
|
|
|
|
height, width, *_ = resolve_minimax_h3_shape("ref2va", sampling, None)
|
|
assert height > 0 and width > 0
|
|
|
|
|
|
@pytest.mark.parametrize("duration", [3.99, 15.01, float("nan"), "not-a-duration"])
|
|
def test_g2_output_duration_rejects_out_of_contract_values(duration):
|
|
from vllm_omni.model_executor.models.minimax_h3.encoder_processing import resolve_minimax_h3_shape
|
|
|
|
sampling = SimpleNamespace(
|
|
fps=24,
|
|
num_frames=1,
|
|
height=None,
|
|
width=None,
|
|
extra_args={"duration": duration},
|
|
)
|
|
with pytest.raises(ValueError, match="duration"):
|
|
resolve_minimax_h3_shape("ref2va", sampling, None)
|
|
|
|
|
|
def test_g1_fanout_uses_incrementing_output_seeds():
|
|
from vllm_omni.diffusion.models.minimax_h3.pipeline_minimax_h3 import (
|
|
_minimax_h3_output_seeds,
|
|
_resolve_minimax_h3_num_outputs,
|
|
)
|
|
|
|
assert _resolve_minimax_h3_num_outputs(3) == 3
|
|
assert _minimax_h3_output_seeds(100, 3) == [100, 101, 102]
|
|
with pytest.raises(ValueError, match="num_outputs_per_prompt"):
|
|
_resolve_minimax_h3_num_outputs(11)
|
|
|
|
|
|
def test_g3_task_specific_aspect_ratio_policy():
|
|
from PIL import Image
|
|
|
|
from vllm_omni.model_executor.models.minimax_h3.preprocessing import (
|
|
resolve_minimax_h3_aspect_ratio,
|
|
)
|
|
|
|
image = Image.new("RGB", (1280, 720))
|
|
assert resolve_minimax_h3_aspect_ratio("fl2va", "9:16", image) == pytest.approx(16 / 9)
|
|
assert resolve_minimax_h3_aspect_ratio("ref2va", None, None) == pytest.approx(16 / 9)
|
|
assert resolve_minimax_h3_aspect_ratio("ref2va", "auto", None) == pytest.approx(16 / 9)
|
|
with pytest.raises(ValueError, match="requires an explicit"):
|
|
resolve_minimax_h3_aspect_ratio("t2va", None, None)
|
|
with pytest.raises(ValueError, match="one of"):
|
|
resolve_minimax_h3_aspect_ratio("t2va", "2:1", None)
|
|
|
|
|
|
def test_g3_t2va_shape_requires_a_named_ratio_and_fl2va_ignores_override():
|
|
from PIL import Image
|
|
|
|
from vllm_omni.model_executor.models.minimax_h3.encoder_processing import resolve_minimax_h3_shape
|
|
|
|
t2va_sampling = SimpleNamespace(
|
|
fps=24,
|
|
num_frames=1,
|
|
height=None,
|
|
width=None,
|
|
extra_args={"duration": 5.0},
|
|
)
|
|
with pytest.raises(ValueError, match="requires an explicit aspect_ratio"):
|
|
resolve_minimax_h3_shape("t2va", t2va_sampling, None)
|
|
|
|
fl2va_sampling = SimpleNamespace(
|
|
fps=24,
|
|
num_frames=1,
|
|
height=None,
|
|
width=None,
|
|
extra_args={"duration": 5.0, "aspect_ratio": "9:16"},
|
|
)
|
|
height, width, *_ = resolve_minimax_h3_shape("fl2va", fl2va_sampling, Image.new("RGB", (1280, 720)))
|
|
assert (height, width) == (768, 1344)
|
|
|
|
|
|
def test_g4_reference_image_boundaries_and_aspect_ratio():
|
|
from PIL import Image
|
|
|
|
from vllm_omni.model_executor.models.minimax_h3.encoder_processing import (
|
|
_validate_reference_image,
|
|
)
|
|
|
|
_validate_reference_image(Image.new("RGB", (256, 256)))
|
|
_validate_reference_image(Image.new("RGB", (5760, 2304)))
|
|
with pytest.raises(ValueError, match="dimensions"):
|
|
_validate_reference_image(Image.new("RGB", (255, 255)))
|
|
with pytest.raises(ValueError, match="aspect ratio"):
|
|
_validate_reference_image(Image.new("RGB", (256, 641)))
|
|
|
|
|
|
def test_g4_reference_image_file_format_and_size_contract(tmp_path):
|
|
from PIL import Image
|
|
|
|
from vllm_omni.model_executor.models.minimax_h3.preprocessing import (
|
|
load_minimax_h3_images,
|
|
)
|
|
|
|
valid_path = tmp_path / "reference.png"
|
|
Image.new("RGB", (256, 256)).save(valid_path)
|
|
assert load_minimax_h3_images(valid_path)[0].size == (256, 256)
|
|
|
|
invalid_path = tmp_path / "reference.bmp"
|
|
Image.new("RGB", (256, 256)).save(invalid_path)
|
|
with pytest.raises(ValueError, match="must use"):
|
|
load_minimax_h3_images(invalid_path)
|
|
|
|
|
|
def test_g4_standalone_audio_duration_and_total_duration_contract():
|
|
from vllm_omni.model_executor.models.minimax_h3.reference_video import (
|
|
validate_reference_audio_waveforms,
|
|
)
|
|
|
|
sample_rate = 16000
|
|
validate_reference_audio_waveforms(
|
|
[
|
|
(torch.zeros(1, 2 * sample_rate), sample_rate),
|
|
(torch.zeros(1, 13 * sample_rate), sample_rate),
|
|
]
|
|
)
|
|
with pytest.raises(ValueError, match="duration"):
|
|
validate_reference_audio_waveforms([(torch.zeros(1, int(1.9 * sample_rate)), sample_rate)])
|
|
with pytest.raises(ValueError, match="at most 15 seconds in total"):
|
|
validate_reference_audio_waveforms(
|
|
[
|
|
(torch.zeros(1, 8 * sample_rate), sample_rate),
|
|
(torch.zeros(1, 8 * sample_rate), sample_rate),
|
|
]
|
|
)
|
|
|
|
|
|
def test_ref2va_video_soundtrack_does_not_consume_standalone_audio_budget():
|
|
from vllm_omni.inputs.data import OmniDiffusionSamplingParams
|
|
from vllm_omni.model_executor.models.minimax_h3.conditioning import (
|
|
MiniMaxH3EncoderConditioning,
|
|
MiniMaxH3EncoderMediaInput,
|
|
MiniMaxH3TextConditioning,
|
|
)
|
|
from vllm_omni.model_executor.models.minimax_h3.encoder_processing import encode_media
|
|
|
|
pipeline = _distilled_pipeline([], {"ref2va": None})
|
|
# A 2.35-second video soundtrack and a 15-second standalone reference
|
|
# each satisfy their modality's separate 15-second duration budget.
|
|
embedded = torch.full((188, 32), 1.0)
|
|
standalone = torch.full((1200, 32), 2.0)
|
|
video_vae = Mock()
|
|
video_vae.encode_image.return_value = torch.ones(1, 96)
|
|
video_vae.encode_video.return_value = (torch.ones(1, 96), (1, 2, 2))
|
|
audio_vae = Mock()
|
|
audio_vae.encode_waveform.side_effect = [(embedded, 94), (standalone, 600)]
|
|
base = MiniMaxH3EncoderConditioning.from_omni_payload(_encoder_payload(task="ref2va", num_frames=362))
|
|
media = MiniMaxH3EncoderMediaInput(
|
|
task=base.task,
|
|
height=base.height,
|
|
width=base.width,
|
|
num_frames=base.num_frames,
|
|
latent_t=base.latent_t,
|
|
audio_t=base.audio_t,
|
|
images=(torch.zeros(32, 32, 3, dtype=torch.uint8),),
|
|
videos=(torch.zeros(1, 32, 32, 3, dtype=torch.uint8),),
|
|
video_audios=((torch.zeros(1, 37600), 16000),),
|
|
audios=((torch.zeros(1, 15 * 16000), 16000),),
|
|
)
|
|
encoded = encode_media(media, video_vae=video_vae, audio_vae=audio_vae, emit_conditioning=True)
|
|
conditioning = MiniMaxH3EncoderConditioning.from_components(
|
|
MiniMaxH3TextConditioning(base.hidden_states, base.token_tags), encoded
|
|
)
|
|
sampling = OmniDiffusionSamplingParams(
|
|
width=1344,
|
|
height=768,
|
|
fps=24,
|
|
num_inference_steps=50,
|
|
extra_args={"task": "ref2va", "duration": 15.0},
|
|
)
|
|
|
|
context = pipeline._prepare_request_inputs(
|
|
{"additional_information": {"encoder_output": conditioning.to_omni_payload()}},
|
|
sampling=sampling,
|
|
)
|
|
|
|
assert context["audio_condition_lengths"] == [94, 600]
|
|
assert [block["kind"] for block in context["ref_blocks"]] == ["image", "video_audio", "audio"]
|
|
torch.testing.assert_close(context["audio_condition"], torch.cat([embedded, standalone]))
|
|
|
|
|
|
def test_ref2va_audio_duration_validation_precedes_rank_branch():
|
|
from vllm_omni.model_executor.models.minimax_h3.conditioning import MiniMaxH3EncoderMediaInput
|
|
from vllm_omni.model_executor.models.minimax_h3.encoder_processing import encode_media
|
|
|
|
media = MiniMaxH3EncoderMediaInput(
|
|
task="ref2va",
|
|
height=32,
|
|
width=32,
|
|
num_frames=0,
|
|
latent_t=1,
|
|
audio_t=1,
|
|
)
|
|
with pytest.raises(ValueError, match="max_standalone_seconds must be positive"):
|
|
encode_media(
|
|
media,
|
|
video_vae=None,
|
|
audio_vae=None,
|
|
emit_conditioning=False,
|
|
)
|
|
|
|
|
|
def test_ref2va_standalone_audio_condition_is_bounded_to_output_duration():
|
|
from vllm_omni.model_executor.models.minimax_h3.conditioning import MiniMaxH3EncoderMediaInput
|
|
from vllm_omni.model_executor.models.minimax_h3.encoder_processing import encode_media
|
|
|
|
observed = []
|
|
|
|
class FakeAudioVAE:
|
|
def encode_waveform(self, waveform, sample_rate):
|
|
observed.append((waveform.clone(), sample_rate))
|
|
return waveform, 100
|
|
|
|
waveform = torch.arange(40, dtype=torch.float32).reshape(1, 40)
|
|
media = MiniMaxH3EncoderMediaInput(
|
|
task="ref2va",
|
|
height=32,
|
|
width=32,
|
|
num_frames=60,
|
|
latent_t=1,
|
|
audio_t=100,
|
|
audios=((waveform, 10),),
|
|
)
|
|
encoded = encode_media(
|
|
media,
|
|
video_vae=None,
|
|
audio_vae=FakeAudioVAE(),
|
|
emit_conditioning=True,
|
|
)
|
|
|
|
assert observed[0][0].shape == (1, 25)
|
|
assert observed[0][1] == 10
|
|
assert encoded.audio_condition.shape == (1, 25)
|
|
assert encoded.audio_condition_lengths == (100,)
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
("field", "value", "message"),
|
|
[
|
|
("fps", 23.0, "FPS"),
|
|
("duration", 1.0, "duration"),
|
|
("format_names", ("avi",), "container"),
|
|
("video_codec", "vp9", "H.264"),
|
|
("audio_codecs", ("opus",), "AAC"),
|
|
("file_size", 50 * 1024 * 1024 + 1, "size"),
|
|
],
|
|
)
|
|
def test_g4_reference_video_metadata_validation(field, value, message, tmp_path):
|
|
from vllm_omni.model_executor.models.minimax_h3.reference_video import (
|
|
_validate_reference_video_metadata,
|
|
)
|
|
|
|
metadata = {
|
|
"width": 1280,
|
|
"height": 720,
|
|
"fps": 24.0,
|
|
"duration": 10.0,
|
|
"format_names": ("mp4",),
|
|
"video_codec": "h264",
|
|
"audio_codecs": ("aac",),
|
|
"file_size": 1024,
|
|
}
|
|
metadata[field] = value
|
|
with pytest.raises(ValueError, match=message):
|
|
_validate_reference_video_metadata(metadata, index=0, source=str(tmp_path / "reference.mp4"))
|
|
|
|
|
|
_ENCODER_HIDDEN = 4
|
|
_ENCODER_HEAD_DIM = 2
|
|
_ENCODER_NUM_HEADS = 2
|
|
_ENCODER_NUM_KV_HEADS = 1
|
|
_ENCODER_INTERMEDIATE = 4
|
|
|
|
# One distinct value per checkpoint tensor, none of them a parameter initializer
|
|
# (fused weights start as `torch.empty`, norms as ones). A uniform fill would
|
|
# prove only that every row was written, not that each source landed in its own
|
|
# row range, which is the invariant the fused loaders have to get right.
|
|
_SOURCE_FILL = {
|
|
"self_attn.q_proj.weight": 3.0,
|
|
"self_attn.k_proj.weight": 4.0,
|
|
"self_attn.v_proj.weight": 5.0,
|
|
"mlp.gate_proj.weight": 6.0,
|
|
"mlp.up_proj.weight": 7.0,
|
|
"input_layernorm.weight": 8.0,
|
|
}
|
|
|
|
_QKV_TARGET = "text_model.layers.0.self_attn.qkv_proj.weight"
|
|
_GATE_UP_TARGET = "text_model.layers.0.mlp.gate_up_proj.weight"
|
|
_NORM_TARGET = "text_model.layers.0.input_layernorm.weight"
|
|
|
|
|
|
def _one_layer_text_encoder():
|
|
"""Encoder stub holding both fused weights plus one plain parameter."""
|
|
from vllm_omni.diffusion.models.minimax_h3.encoder import (
|
|
MiniMaxH3Qwen3VLEncoder,
|
|
MiniMaxH3Qwen3VLMergedColumnParallelLinear,
|
|
MiniMaxH3Qwen3VLQKVParallelLinear,
|
|
MiniMaxH3Qwen3VLRMSNorm,
|
|
)
|
|
|
|
group = SimpleNamespace(rank_in_group=0, world_size=1, device_group=None, ranks=[0])
|
|
layer = nn.Module()
|
|
layer.self_attn = nn.Module()
|
|
layer.mlp = nn.Module()
|
|
with (
|
|
patch("vllm.model_executor.parameter.get_tensor_model_parallel_rank", return_value=0),
|
|
patch("vllm.model_executor.parameter.get_tensor_model_parallel_world_size", return_value=1),
|
|
):
|
|
layer.self_attn.qkv_proj = MiniMaxH3Qwen3VLQKVParallelLinear(
|
|
group,
|
|
hidden_size=_ENCODER_HIDDEN,
|
|
num_heads=_ENCODER_NUM_HEADS,
|
|
num_kv_heads=_ENCODER_NUM_KV_HEADS,
|
|
head_dim=_ENCODER_HEAD_DIM,
|
|
dtype=torch.float32,
|
|
)
|
|
layer.mlp.gate_up_proj = MiniMaxH3Qwen3VLMergedColumnParallelLinear(
|
|
group,
|
|
input_size=_ENCODER_HIDDEN,
|
|
intermediate_size=_ENCODER_INTERMEDIATE,
|
|
dtype=torch.float32,
|
|
)
|
|
layer.input_layernorm = MiniMaxH3Qwen3VLRMSNorm(_ENCODER_HIDDEN)
|
|
|
|
encoder = object.__new__(MiniMaxH3Qwen3VLEncoder)
|
|
nn.Module.__init__(encoder)
|
|
encoder.text_model = nn.Module()
|
|
encoder.text_model.layers = nn.ModuleList([layer])
|
|
return encoder
|
|
|
|
|
|
def _write_text_encoder_checkpoint(path, *, omit=()):
|
|
"""Write a one-layer Qwen3-VL-named checkpoint, one fill value per source."""
|
|
import json
|
|
|
|
import safetensors.torch
|
|
|
|
prefix = "model.language_model.layers.0"
|
|
q_rows = _ENCODER_NUM_HEADS * _ENCODER_HEAD_DIM
|
|
kv_rows = _ENCODER_NUM_KV_HEADS * _ENCODER_HEAD_DIM
|
|
shapes = {
|
|
"self_attn.q_proj.weight": (q_rows, _ENCODER_HIDDEN),
|
|
"self_attn.k_proj.weight": (kv_rows, _ENCODER_HIDDEN),
|
|
"self_attn.v_proj.weight": (kv_rows, _ENCODER_HIDDEN),
|
|
"mlp.gate_proj.weight": (_ENCODER_INTERMEDIATE, _ENCODER_HIDDEN),
|
|
"mlp.up_proj.weight": (_ENCODER_INTERMEDIATE, _ENCODER_HIDDEN),
|
|
"input_layernorm.weight": (_ENCODER_HIDDEN,),
|
|
}
|
|
tensors = {f"{prefix}.{source}": torch.full(shape, _SOURCE_FILL[source]) for source, shape in shapes.items()}
|
|
for source in omit:
|
|
del tensors[f"{prefix}.{source}"]
|
|
safetensors.torch.save_file(tensors, str(path / "model.safetensors"))
|
|
index = {"weight_map": dict.fromkeys(tensors, "model.safetensors")}
|
|
(path / "model.safetensors.index.json").write_text(json.dumps(index))
|
|
return tensors
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
("omitted_sources", "expected_detail", "unreported_target"),
|
|
[
|
|
(("self_attn.q_proj.weight",), f"{_QKV_TARGET}: ['q']", _GATE_UP_TARGET),
|
|
(("self_attn.k_proj.weight",), f"{_QKV_TARGET}: ['k']", _GATE_UP_TARGET),
|
|
(("self_attn.v_proj.weight",), f"{_QKV_TARGET}: ['v']", _GATE_UP_TARGET),
|
|
(("mlp.gate_proj.weight",), f"{_GATE_UP_TARGET}: ['gate']", _QKV_TARGET),
|
|
(("mlp.up_proj.weight",), f"{_GATE_UP_TARGET}: ['up']", _QKV_TARGET),
|
|
(
|
|
("self_attn.q_proj.weight", "self_attn.k_proj.weight", "self_attn.v_proj.weight"),
|
|
f"{_QKV_TARGET}: ['q', 'k', 'v']",
|
|
_GATE_UP_TARGET,
|
|
),
|
|
],
|
|
)
|
|
def test_encoder_load_reports_missing_fused_source_shards(
|
|
tmp_path, omitted_sources, expected_detail, unreported_target
|
|
):
|
|
encoder = _one_layer_text_encoder()
|
|
weights = _write_text_encoder_checkpoint(tmp_path, omit=omitted_sources)
|
|
|
|
with pytest.raises(RuntimeError) as excinfo:
|
|
encoder.load_weights(weights.items())
|
|
|
|
message = str(excinfo.value)
|
|
assert expected_detail in message
|
|
assert unreported_target not in message
|
|
|
|
|
|
def test_encoder_load_rejects_unloaded_plain_param(tmp_path):
|
|
encoder = _one_layer_text_encoder()
|
|
weights = _write_text_encoder_checkpoint(tmp_path, omit=("input_layernorm.weight",))
|
|
|
|
with pytest.raises(RuntimeError, match=r"weights not loaded.*input_layernorm"):
|
|
encoder.load_weights(weights.items())
|
|
|
|
|
|
def test_encoder_load_places_every_source_in_its_own_rows(tmp_path):
|
|
encoder = _one_layer_text_encoder()
|
|
weights = _write_text_encoder_checkpoint(tmp_path)
|
|
|
|
encoder.load_weights(weights.items())
|
|
|
|
params = dict(encoder.named_parameters())
|
|
q_rows = _ENCODER_NUM_HEADS * _ENCODER_HEAD_DIM
|
|
kv_rows = _ENCODER_NUM_KV_HEADS * _ENCODER_HEAD_DIM
|
|
qkv = params[_QKV_TARGET]
|
|
gate_up = params[_GATE_UP_TARGET]
|
|
slices = {
|
|
"self_attn.q_proj.weight": qkv[:q_rows],
|
|
"self_attn.k_proj.weight": qkv[q_rows : q_rows + kv_rows],
|
|
"self_attn.v_proj.weight": qkv[q_rows + kv_rows :],
|
|
"mlp.gate_proj.weight": gate_up[:_ENCODER_INTERMEDIATE],
|
|
"mlp.up_proj.weight": gate_up[_ENCODER_INTERMEDIATE:],
|
|
"input_layernorm.weight": params[_NORM_TARGET],
|
|
}
|
|
for source, rows in slices.items():
|
|
assert torch.equal(rows, torch.full_like(rows, _SOURCE_FILL[source])), source
|
|
|
|
|
|
def _pad_seq_len_batch(pad_seq_len):
|
|
"""A t2va request that pins the packed length through ``extra_args``."""
|
|
from vllm_omni.diffusion.request import OmniDiffusionRequest
|
|
from vllm_omni.diffusion.worker.request_batch import DiffusionRequestBatch
|
|
from vllm_omni.inputs.data import OmniDiffusionSamplingParams
|
|
|
|
extra_args = {"task": "t2va", "aspect_ratio": "16:9"}
|
|
if pad_seq_len is not None:
|
|
extra_args["pad_seq_len"] = pad_seq_len
|
|
sampling = OmniDiffusionSamplingParams(
|
|
quality="lossless",
|
|
width=1344,
|
|
height=768,
|
|
fps=24,
|
|
num_frames=124,
|
|
num_inference_steps=None,
|
|
extra_args=extra_args,
|
|
)
|
|
return DiffusionRequestBatch(
|
|
[
|
|
OmniDiffusionRequest(
|
|
prompt=_encoder_prompt("pinned packed length"),
|
|
sampling_params=sampling,
|
|
request_id="pad-seq-len",
|
|
)
|
|
]
|
|
)
|
|
|
|
|
|
@pytest.mark.parametrize("pinned, expected", [(None, None), (54080, 54080)])
|
|
def test_pad_seq_len_reaches_diffuse_from_extra_args(pinned, expected):
|
|
"""The request field must arrive at ``diffuse``, not stop at validation."""
|
|
diffuse_calls: list[dict[str, Any]] = []
|
|
pipeline = _distilled_pipeline(diffuse_calls, {"fl2va": None, "ref2va": None})
|
|
|
|
pipeline.forward(_pad_seq_len_batch(pinned))
|
|
|
|
assert [call["pad_seq_len"] for call in diffuse_calls] == [expected]
|
|
|
|
|
|
@pytest.mark.parametrize("task", ["t2va", "ref2va"])
|
|
def test_pad_seq_len_reaches_the_packer_for_every_task_family(task, monkeypatch):
|
|
"""``diffuse`` must hand the pin to whichever packer the task uses."""
|
|
import vllm_omni.diffusion.models.minimax_h3.pipeline_minimax_h3 as module
|
|
|
|
class _CapturedError(Exception):
|
|
pass
|
|
|
|
recorded = {}
|
|
|
|
def _capture(**kwargs):
|
|
recorded.update(kwargs)
|
|
raise _CapturedError
|
|
|
|
monkeypatch.setattr(module, "minimax_h3_packed_sequence", _capture)
|
|
monkeypatch.setattr(module, "minimax_h3_packed_sequence_ref2va_blocks", _capture)
|
|
|
|
pipeline = _distilled_pipeline([], {"fl2va": None, "ref2va": None})
|
|
pipeline._initial_noise = lambda **kwargs: (torch.zeros(1), torch.zeros(1))
|
|
|
|
with pytest.raises(_CapturedError):
|
|
pipeline._build_denoise_inputs(
|
|
task=task,
|
|
text_embeddings=torch.ones(7, 2),
|
|
text_tags=torch.ones(7, dtype=torch.long),
|
|
seed=0,
|
|
latent_t=2,
|
|
latent_h=4,
|
|
latent_w=6,
|
|
audio_t=3,
|
|
num_frames=5,
|
|
num_steps=2,
|
|
video_shift=12.0,
|
|
audio_shift=3.0,
|
|
base_schedule=None,
|
|
visual_condition=None,
|
|
visual_condition_shape=(1, 4, 6),
|
|
audio_condition=None,
|
|
ref_audio_t=1,
|
|
pad_seq_len=192,
|
|
)
|
|
|
|
assert recorded["seq_len"] == 192
|
|
|
|
|
|
@pytest.mark.parametrize("value", [0, -64, 100, "128", True, 1.5, (1 << 20) + 64])
|
|
def test_pad_seq_len_request_validation_rejects_bad_values(value):
|
|
from vllm_omni.diffusion.models.minimax_h3.pipeline_minimax_h3 import (
|
|
_resolve_pad_seq_len,
|
|
)
|
|
from vllm_omni.errors import OmniClientError
|
|
|
|
with pytest.raises(OmniClientError):
|
|
_resolve_pad_seq_len(value)
|
|
|
|
|
|
def test_pad_seq_len_request_validation_accepts_aligned_values():
|
|
from vllm_omni.diffusion.models.minimax_h3.packed_sequence import (
|
|
MINIMAX_H3_MAX_PAD_SEQ_LEN,
|
|
)
|
|
from vllm_omni.diffusion.models.minimax_h3.pipeline_minimax_h3 import (
|
|
_resolve_pad_seq_len,
|
|
)
|
|
|
|
assert _resolve_pad_seq_len(None) is None
|
|
assert _resolve_pad_seq_len(54080) == 54080
|
|
assert _resolve_pad_seq_len(MINIMAX_H3_MAX_PAD_SEQ_LEN) == MINIMAX_H3_MAX_PAD_SEQ_LEN
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
("capability", "allow_tf32"),
|
|
[((8, 6), True), ((9, 0), False), ((10, 0), True), ((10, 3), True), (None, True)],
|
|
)
|
|
@pytest.mark.parametrize("initial_tf32", [False, True])
|
|
@pytest.mark.parametrize("fail_encode", [False, True])
|
|
def test_keyframe_encode_pins_and_restores_cudnn_settings(
|
|
monkeypatch, capability, allow_tf32, initial_tf32, fail_encode
|
|
):
|
|
from vllm.platforms.interface import DeviceCapability
|
|
|
|
from vllm_omni.diffusion.models.minimax_h3.vae import (
|
|
_minimax_h3_keyframe_encode_context,
|
|
)
|
|
from vllm_omni.platforms import current_omni_platform
|
|
|
|
def get_capability(device_id):
|
|
assert device_id == 1
|
|
return DeviceCapability(*capability) if capability is not None else None
|
|
|
|
monkeypatch.setattr(current_omni_platform, "is_cuda", lambda: True)
|
|
monkeypatch.setattr(current_omni_platform, "get_device_capability", get_capability)
|
|
|
|
cudnn = torch.backends.cudnn
|
|
with cudnn.flags(enabled=False, benchmark=True, deterministic=False, allow_tf32=initial_tf32):
|
|
error_context = pytest.raises(RuntimeError, match="encode failed") if fail_encode else nullcontext()
|
|
with error_context, _minimax_h3_keyframe_encode_context(torch.device("cuda:1")):
|
|
assert cudnn.enabled
|
|
assert not cudnn.benchmark
|
|
assert cudnn.deterministic
|
|
assert cudnn.allow_tf32 is allow_tf32
|
|
if fail_encode:
|
|
raise RuntimeError("encode failed")
|
|
|
|
assert not cudnn.enabled
|
|
assert cudnn.benchmark
|
|
assert not cudnn.deterministic
|
|
assert cudnn.allow_tf32 is initial_tf32
|
|
|
|
|
|
def test_keyframe_encode_cpu_keeps_cudnn_settings(monkeypatch):
|
|
from vllm_omni.diffusion.models.minimax_h3.vae import (
|
|
_minimax_h3_keyframe_encode_context,
|
|
)
|
|
from vllm_omni.platforms import current_omni_platform
|
|
|
|
def unexpected_capability_query(device_id):
|
|
pytest.fail("CPU keyframe encode must not query CUDA capability")
|
|
|
|
monkeypatch.setattr(current_omni_platform, "get_device_capability", unexpected_capability_query)
|
|
cudnn = torch.backends.cudnn
|
|
with cudnn.flags(enabled=False, benchmark=True, deterministic=False, allow_tf32=False):
|
|
with _minimax_h3_keyframe_encode_context(torch.device("cpu")):
|
|
assert not cudnn.enabled
|
|
assert cudnn.benchmark
|
|
assert not cudnn.deterministic
|
|
assert not cudnn.allow_tf32
|
|
|
|
|
|
def _peer_rank_preencode_pipeline(monkeypatch):
|
|
"""An H3 pipeline whose video VAE is a distributed rank that owns no output.
|
|
|
|
Every rank of the VAE group must call ``decode_with_chunks`` to keep the
|
|
temporal collectives in lockstep, but only the owner receives chunks, so a
|
|
peer's callback is never invoked.
|
|
"""
|
|
from vllm_omni.diffusion.models.minimax_h3 import pipeline_minimax_h3 as mod
|
|
|
|
class FakeAudioVAE:
|
|
def decode_latent(self, latent):
|
|
return torch.zeros(1, 1, 2)
|
|
|
|
class PeerRankVideoVAE:
|
|
chunk_value_range = (0.0, 1.0)
|
|
|
|
def decode_with_chunks(self, latent, *, on_chunk):
|
|
del latent, on_chunk
|
|
|
|
pipeline = object.__new__(mod.MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
pipeline.audio_vae = FakeAudioVAE()
|
|
pipeline.video_vae = PeerRankVideoVAE()
|
|
pipeline.device = torch.device("cpu")
|
|
pipeline.od_config = SimpleNamespace()
|
|
monkeypatch.setattr(mod.MiniMaxH3Pipeline, "_uses_manual_component_offload", lambda self, component: False)
|
|
return pipeline
|
|
|
|
|
|
def test_peer_vae_rank_returns_no_preencoded_output_instead_of_failing(monkeypatch):
|
|
"""A rank that owns no output must not turn its silence into a request error."""
|
|
pipeline = _peer_rank_preencode_pipeline(monkeypatch)
|
|
|
|
output = pipeline.decode_to_mp4(torch.zeros(1), torch.zeros(1), height=2, width=2)
|
|
|
|
assert output == b""
|
|
|
|
|
|
def test_peer_vae_rank_post_decode_reaches_post_processing(monkeypatch):
|
|
"""The real H3 post_decode chain, not just the shared consumer, tolerates a peer rank."""
|
|
from vllm_omni.diffusion.models.minimax_h3 import pipeline_minimax_h3 as mod
|
|
from vllm_omni.diffusion.worker.utils import StepRequestState
|
|
|
|
pipeline = _peer_rank_preencode_pipeline(monkeypatch)
|
|
monkeypatch.setattr(pipeline, "_unpack_denoised_rows", lambda *args, **kwargs: (torch.zeros(1), torch.zeros(1)))
|
|
state = StepRequestState(
|
|
request_id="peer-rank",
|
|
sampling=SimpleNamespace(num_outputs_per_prompt=1),
|
|
prompt="a prompt",
|
|
)
|
|
state.latents = torch.zeros(1)
|
|
state.extra[mod._STEP_BRANCH] = object()
|
|
state.extra[mod._STEP_AUDIO_ROWS] = torch.zeros(1)
|
|
state.extra[mod._STEP_SHAPE] = {
|
|
"latent_t": 1,
|
|
"latent_h": 1,
|
|
"latent_w": 1,
|
|
"audio_t": 1,
|
|
"height": 2,
|
|
"width": 2,
|
|
"preencode_mp4": True,
|
|
}
|
|
|
|
output = pipeline.post_decode(state)
|
|
|
|
assert output.output == (b"", None)
|
|
assert mod._minimax_h3_post_process(output.output)["video"] == [b""]
|
|
|
|
|
|
@pytest.mark.parametrize("duration", [60, 75])
|
|
def test_long_video_shape_requires_explicit_opt_in(duration):
|
|
from vllm_omni.errors import OmniClientError
|
|
from vllm_omni.inputs.data import OmniDiffusionSamplingParams
|
|
from vllm_omni.model_executor.models.minimax_h3.encoder_processing import resolve_minimax_h3_shape
|
|
|
|
sampling = OmniDiffusionSamplingParams(width=960, height=544, fps=24, extra_args={"duration": duration})
|
|
with pytest.raises(OmniClientError, match="15"):
|
|
resolve_minimax_h3_shape("ref2va", sampling, None)
|
|
sampling.extra_args["long_video"] = True
|
|
_, _, frames, video_t, audio_t = resolve_minimax_h3_shape("ref2va", sampling, None)
|
|
assert frames >= duration * 24 and frames % 17 == 5
|
|
assert video_t == (frames - 5) // 17 * 5 + 2
|
|
assert audio_t == round(frames / 24 * 40)
|
|
|
|
|
|
@pytest.mark.parametrize("preencode", [False, True])
|
|
@pytest.mark.parametrize("cancel_phase", ["before_prepare", "prepare", "diffuse"])
|
|
def test_request_cancellation_at_prepare_and_decode_boundaries(preencode, cancel_phase, monkeypatch):
|
|
from vllm_omni.diffusion.cancellation import RequestCancellationRegistry, request_cancellation_scope
|
|
from vllm_omni.diffusion.data import DiffusionRequestAbortedError
|
|
from vllm_omni.diffusion.models.minimax_h3 import MiniMaxH3Pipeline
|
|
from vllm_omni.platforms import current_omni_platform
|
|
|
|
synchronize = Mock()
|
|
monkeypatch.setattr(current_omni_platform, "synchronize", synchronize)
|
|
pipeline = object.__new__(MiniMaxH3Pipeline)
|
|
torch.nn.Module.__init__(pipeline)
|
|
registry = RequestCancellationRegistry()
|
|
signal = registry.create("request")
|
|
|
|
def prepare(*args):
|
|
if cancel_phase == "prepare":
|
|
registry.cancel(["request"])
|
|
return {"num_outputs": 1, "seed": 1101, "preencode_mp4": preencode}
|
|
|
|
def diffuse(**kwargs):
|
|
registry.cancel(["request"])
|
|
return torch.zeros(1), torch.zeros(1)
|
|
|
|
pipeline._prepare_request_inputs = Mock(side_effect=prepare)
|
|
pipeline._denoise_kwargs = Mock(return_value={})
|
|
pipeline.diffuse = Mock(side_effect=diffuse)
|
|
pipeline.decode = Mock()
|
|
pipeline.decode_to_mp4 = Mock()
|
|
try:
|
|
if cancel_phase == "before_prepare":
|
|
registry.cancel(["request"])
|
|
with request_cancellation_scope([signal]), pytest.raises(DiffusionRequestAbortedError):
|
|
pipeline.forward(_t2va_batch())
|
|
synchronize.assert_called_once_with()
|
|
if cancel_phase == "before_prepare":
|
|
pipeline._prepare_request_inputs.assert_not_called()
|
|
if cancel_phase != "diffuse":
|
|
pipeline.diffuse.assert_not_called()
|
|
pipeline.decode.assert_not_called()
|
|
pipeline.decode_to_mp4.assert_not_called()
|
|
finally:
|
|
registry.close()
|