313 lines
10 KiB
Python
313 lines
10 KiB
Python
from __future__ import annotations
|
|
|
|
import copy
|
|
import hashlib
|
|
|
|
import pytest
|
|
import torch
|
|
|
|
from ComfyUI_H3_Continuum_Join.v3.refine_schedule import (
|
|
AUDIO_POLICY_LOCKED_PASSTHROUGH,
|
|
NOISE_MODE_VIDEO_RANDOM_AUDIO_ZERO,
|
|
make_tail_schedule,
|
|
)
|
|
from ComfyUI_H3_Continuum_Join.v3.refine_target import (
|
|
AUDIO_ONLY_SEED_NAMESPACE,
|
|
MODE_AUDIO_ONLY,
|
|
MODE_VIDEO_AUDIO,
|
|
MODE_VIDEO_ONLY,
|
|
OUTPUT_INPUT_PASSTHROUGH,
|
|
OUTPUT_SAMPLED,
|
|
VIDEO_AUDIO_SEED_NAMESPACE,
|
|
VIDEO_ONLY_SEED_NAMESPACE,
|
|
RefineTargetError,
|
|
build_target_execution_contract,
|
|
derive_target_refine_seed,
|
|
make_refine_target,
|
|
)
|
|
from ComfyUI_H3_Continuum_Join.v3.second_pass import (
|
|
derive_refine_seed,
|
|
run_second_pass_groups,
|
|
)
|
|
from ComfyUI_H3_Continuum_Join.v3.second_pass_nodes import (
|
|
H3ContinuumSecondPassV35,
|
|
)
|
|
from ComfyUI_H3_Continuum_Join.v3.targeted_refine_sampling import (
|
|
sample_targeted_refine_chunk,
|
|
)
|
|
from ComfyUI_H3_Continuum_Join.v3.targeted_second_pass import (
|
|
run_targeted_second_pass_groups,
|
|
)
|
|
|
|
|
|
def _sha(tensor: torch.Tensor) -> str:
|
|
value = tensor.detach().to(device="cpu").contiguous()
|
|
return hashlib.sha256(value.view(torch.uint8).numpy().tobytes()).hexdigest()
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
"mode,video_policy,audio_policy,namespace",
|
|
[
|
|
(
|
|
MODE_VIDEO_ONLY,
|
|
("seeded_random", 1, OUTPUT_SAMPLED),
|
|
("zero", 0, OUTPUT_INPUT_PASSTHROUGH),
|
|
VIDEO_ONLY_SEED_NAMESPACE,
|
|
),
|
|
(
|
|
MODE_AUDIO_ONLY,
|
|
("zero", 0, OUTPUT_INPUT_PASSTHROUGH),
|
|
("seeded_random", 1, OUTPUT_SAMPLED),
|
|
AUDIO_ONLY_SEED_NAMESPACE,
|
|
),
|
|
(
|
|
MODE_VIDEO_AUDIO,
|
|
("seeded_random", 1, OUTPUT_SAMPLED),
|
|
("seeded_random", 1, OUTPUT_SAMPLED),
|
|
VIDEO_AUDIO_SEED_NAMESPACE,
|
|
),
|
|
],
|
|
)
|
|
def test_refine_target_v1_freezes_all_three_stream_policies(
|
|
mode,
|
|
video_policy,
|
|
audio_policy,
|
|
namespace,
|
|
):
|
|
target = make_refine_target(mode)
|
|
assert target.contract["magic"] == "H3_CONTINUUM_REFINE_TARGET"
|
|
assert target.contract["schema_version"] == 1
|
|
assert target.contract["mode"] == mode
|
|
assert (
|
|
target.contract["video"]["noise"],
|
|
target.contract["video"]["mask"],
|
|
target.contract["video"]["output"],
|
|
) == video_policy
|
|
assert (
|
|
target.contract["audio"]["noise"],
|
|
target.contract["audio"]["mask"],
|
|
target.contract["audio"]["output"],
|
|
) == audio_policy
|
|
assert target.contract["seed_namespace"] == namespace
|
|
assert len(target.contract["target_hash"]) == 64
|
|
|
|
|
|
def test_target_seed_namespaces_are_stable_and_video_only_is_legacy_exact():
|
|
assert derive_target_refine_seed(42, 0, MODE_VIDEO_ONLY) == derive_refine_seed(42, 0)
|
|
seeds = {
|
|
derive_target_refine_seed(42, 0, mode)
|
|
for mode in (MODE_VIDEO_ONLY, MODE_AUDIO_ONLY, MODE_VIDEO_AUDIO)
|
|
}
|
|
assert len(seeds) == 3
|
|
assert derive_target_refine_seed(42, 1, MODE_VIDEO_ONLY) != derive_refine_seed(42, 0)
|
|
|
|
|
|
def test_target_execution_uses_sigma_range_without_mutating_schedule_v1():
|
|
sigmas = torch.tensor([0.9, 0.6, 0.3, 0.0], dtype=torch.float32)
|
|
schedule = make_tail_schedule(sigmas, evaluation_count=2)
|
|
before = copy.deepcopy(dict(schedule.contract))
|
|
execution = build_target_execution_contract(
|
|
make_refine_target(MODE_VIDEO_ONLY),
|
|
schedule.contract,
|
|
)
|
|
|
|
assert dict(schedule.contract) == before
|
|
assert schedule.contract["noise_mode"] == NOISE_MODE_VIDEO_RANDOM_AUDIO_ZERO
|
|
assert schedule.contract["audio_policy"] == AUDIO_POLICY_LOCKED_PASSTHROUGH
|
|
assert execution["policy_authority"] == "refine_target_contract_v1"
|
|
assert execution["sigma_range"]["sigma_hash"] == schedule.contract["sigma_hash"]
|
|
assert execution["sigma_range"]["source_schedule_hash"] == schedule.contract["schedule_hash"]
|
|
assert "noise_mode" not in execution["sigma_range"]
|
|
assert "audio_policy" not in execution["sigma_range"]
|
|
|
|
|
|
def test_invalid_target_contracts_fail_before_sampling():
|
|
with pytest.raises(RefineTargetError, match="unsupported"):
|
|
make_refine_target("unknown")
|
|
|
|
called = False
|
|
|
|
def forbidden(**_kwargs):
|
|
nonlocal called
|
|
called = True
|
|
raise AssertionError("future target must stop before Sampling")
|
|
|
|
with pytest.raises(RefineTargetError, match="unsupported"):
|
|
sample_targeted_refine_chunk(
|
|
model=object(),
|
|
conditioning=[],
|
|
latent={"samples": object()},
|
|
sampler=object(),
|
|
sigmas=torch.tensor([0.2, 0.0]),
|
|
seed=1,
|
|
refine_target="unknown",
|
|
legacy_sample_fn=forbidden,
|
|
)
|
|
assert called is False
|
|
|
|
|
|
def test_targeted_video_only_sampler_delegates_every_argument_unchanged():
|
|
captured = {}
|
|
sentinel = {"samples": object()}
|
|
inputs = {
|
|
"model": object(),
|
|
"conditioning": [[torch.zeros((1, 1, 1)), {}]],
|
|
"latent": {"samples": object()},
|
|
"sampler": object(),
|
|
"sigmas": torch.tensor([0.4, 0.0]),
|
|
"seed": 99,
|
|
"enable_preview": False,
|
|
}
|
|
|
|
def delegate(**kwargs):
|
|
captured.update(kwargs)
|
|
return sentinel
|
|
|
|
result = sample_targeted_refine_chunk(
|
|
**inputs,
|
|
refine_target=MODE_VIDEO_ONLY,
|
|
legacy_sample_fn=delegate,
|
|
)
|
|
assert result is sentinel
|
|
assert captured == inputs
|
|
|
|
|
|
def _group(group_id: int, *, temporal: int, audio_t: int, terminal: bool) -> dict:
|
|
return {
|
|
"group_id": group_id,
|
|
"logical_chunks": [1] if group_id == 0 else [2, 3],
|
|
"physical_prompt": f"group {group_id + 1}",
|
|
"prompt_policy": "paired_timeline_v1" if terminal else "single",
|
|
"physical_frames": 260 if terminal else 124,
|
|
"trim_prefix_frames": 22 if terminal else 0,
|
|
"terminal_merged": terminal,
|
|
"source_width": 128,
|
|
"source_height": 128,
|
|
"source_batch": 1,
|
|
"latent_channels": 24,
|
|
"source_latent_t": temporal,
|
|
"source_latent_h": 8,
|
|
"source_latent_w": 8,
|
|
"source_audio_shape": [1, 32, 2, audio_t],
|
|
}
|
|
|
|
|
|
def _fixture():
|
|
groups = [
|
|
_group(0, temporal=31, audio_t=207, terminal=False),
|
|
_group(1, temporal=65, audio_t=433, terminal=True),
|
|
]
|
|
videos = [
|
|
{"samples": torch.arange(24 * group["source_latent_t"] * 8 * 8, dtype=torch.float32).reshape(1, 24, group["source_latent_t"], 8, 8)}
|
|
for group in groups
|
|
]
|
|
audios = [
|
|
{"samples": torch.arange(32 * 2 * group["source_audio_shape"][-1], dtype=torch.float32).reshape(group["source_audio_shape"])}
|
|
for group in groups
|
|
]
|
|
plan = {
|
|
"width": 128,
|
|
"height": 128,
|
|
"second_pass_contract": {"version": 1, "physical_groups": groups},
|
|
}
|
|
return groups, videos, audios, plan
|
|
|
|
|
|
def _run(*, targeted: bool):
|
|
groups, videos, audios, plan = _fixture()
|
|
calls = []
|
|
|
|
def sample(**kwargs):
|
|
calls.append(kwargs)
|
|
return {
|
|
"video": kwargs["latent"]["video"] + float(len(calls)),
|
|
"audio": kwargs["latent"]["audio"] + 100.0,
|
|
}
|
|
|
|
common = dict(
|
|
model="model",
|
|
clip="clip",
|
|
sampler="sampler",
|
|
sigmas=torch.tensor([0.6, 0.3, 0.0], dtype=torch.float32),
|
|
video_latents=videos,
|
|
audio_latents=audios,
|
|
assembly_plan=plan,
|
|
refine_seed=73,
|
|
encode_prompt_fn=lambda _clip, prompt, **_kwargs: [prompt],
|
|
latent_builder=lambda video, audio: {"video": video, "audio": audio},
|
|
sample_fn=sample,
|
|
stream_extractor=lambda value: (value["video"], value["audio"]),
|
|
clone_model_fn=lambda source, **kwargs: (source, kwargs),
|
|
enable_preview=False,
|
|
)
|
|
if targeted:
|
|
result = run_targeted_second_pass_groups(
|
|
**common,
|
|
refine_target=MODE_VIDEO_ONLY,
|
|
)
|
|
else:
|
|
result = run_second_pass_groups(**common)
|
|
return groups, videos, audios, common["sigmas"], calls, result
|
|
|
|
|
|
def test_targeted_video_only_is_exact_legacy_second_pass_parity():
|
|
legacy = _run(targeted=False)
|
|
targeted = _run(targeted=True)
|
|
legacy_groups, _legacy_inputs, legacy_audio_inputs, legacy_sigmas, legacy_calls, legacy_result = legacy
|
|
targeted_groups, _targeted_inputs, targeted_audio_inputs, targeted_sigmas, targeted_calls, targeted_result = targeted
|
|
legacy_videos, legacy_audios, legacy_plan, legacy_status = legacy_result
|
|
targeted_videos, targeted_audios, targeted_plan, targeted_status = targeted_result
|
|
|
|
assert [_sha(item["samples"]) for item in targeted_videos] == [
|
|
_sha(item["samples"]) for item in legacy_videos
|
|
]
|
|
assert all(
|
|
output is source
|
|
for output, source in zip(targeted_audios, targeted_audio_inputs, strict=True)
|
|
)
|
|
assert all(
|
|
output is source
|
|
for output, source in zip(legacy_audios, legacy_audio_inputs, strict=True)
|
|
)
|
|
assert [call["seed"] for call in targeted_calls] == [
|
|
call["seed"] for call in legacy_calls
|
|
] == [derive_refine_seed(73, 0), derive_refine_seed(73, 1)]
|
|
assert all(call["sigmas"] is targeted_sigmas for call in targeted_calls)
|
|
assert all(call["sigmas"] is legacy_sigmas for call in legacy_calls)
|
|
assert [call["conditioning"] for call in targeted_calls] == [
|
|
call["conditioning"] for call in legacy_calls
|
|
] == [["group 1"], ["group 2"]]
|
|
assert len(targeted_calls) == len(legacy_calls) == len(legacy_groups) == len(targeted_groups) == 2
|
|
assert legacy_status == targeted_status
|
|
|
|
normalized_targeted = copy.deepcopy(targeted_plan)
|
|
target_contract = normalized_targeted["second_pass_contract"].pop(
|
|
"refine_target_contract"
|
|
)
|
|
execution_contract = normalized_targeted["second_pass_contract"].pop(
|
|
"refine_execution_contract"
|
|
)
|
|
assert normalized_targeted == legacy_plan
|
|
assert target_contract["mode"] == MODE_VIDEO_ONLY
|
|
assert execution_contract["target"] == target_contract
|
|
assert execution_contract["sigma_range"]["sigma_hash"] == (
|
|
legacy_plan["second_pass_contract"]["refine_schedule"]["sigma_hash"]
|
|
)
|
|
|
|
|
|
def test_existing_v35_public_schema_remains_frozen():
|
|
schema = H3ContinuumSecondPassV35.INPUT_TYPES()
|
|
assert list(schema["required"]) == [
|
|
"model",
|
|
"clip",
|
|
"sampler",
|
|
"sigmas",
|
|
"video_latents",
|
|
"audio_latents",
|
|
"assembly_plan",
|
|
"refine_seed",
|
|
]
|
|
assert list(schema["optional"]) == ["refine_context", "video_vae"]
|
|
assert "refine_target" not in schema["required"]
|
|
assert "refine_target" not in schema["optional"]
|