Release V3.8 source, illustrated manual and Spectrum workflow
This commit is contained in:
@@ -0,0 +1,213 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import importlib.util
|
||||
from pathlib import Path
|
||||
import sys
|
||||
|
||||
import torch
|
||||
|
||||
|
||||
TOOLS = Path(__file__).parents[1] / "tools"
|
||||
|
||||
|
||||
def _load(name: str, filename: str):
|
||||
path = TOOLS / filename
|
||||
spec = importlib.util.spec_from_file_location(name, path)
|
||||
assert spec is not None and spec.loader is not None
|
||||
module = importlib.util.module_from_spec(spec)
|
||||
sys.modules[name] = module
|
||||
spec.loader.exec_module(module)
|
||||
return module
|
||||
|
||||
|
||||
policy_module = _load(
|
||||
"h3_issue13_r28_continuation_policy",
|
||||
"gpu_diagnostic_node/continuation_policy.py",
|
||||
)
|
||||
runner = _load("h3_issue13_r28_gate_runner", "issue13_r28_gate_runner.py")
|
||||
analyzer = _load("h3_issue13_r28_analyze", "issue13_r28_analyze.py")
|
||||
|
||||
|
||||
class FakeNested:
|
||||
def __init__(self, tensors):
|
||||
self.tensors = tuple(tensors)
|
||||
|
||||
def unbind(self):
|
||||
return self.tensors
|
||||
|
||||
|
||||
def _summary(tensor: torch.Tensor) -> dict:
|
||||
return {
|
||||
"shape": [int(value) for value in tensor.shape],
|
||||
"mean": float(tensor.float().mean().item()),
|
||||
}
|
||||
|
||||
|
||||
def _latent():
|
||||
video = torch.randn((1, 24, 42, 4, 4), generator=torch.Generator().manual_seed(1))
|
||||
audio = torch.randn((1, 32, 2, 235), generator=torch.Generator().manual_seed(2))
|
||||
video_mask = torch.ones((1, 1, 42, 4, 4), dtype=torch.float32)
|
||||
video_mask[:, :, :7] = 0.0
|
||||
audio_mask = torch.ones((1, 1, 2, 235), dtype=torch.float32)
|
||||
audio_mask[..., :37] = 0.0
|
||||
return {
|
||||
"samples": FakeNested((video, audio)),
|
||||
"noise_mask": FakeNested((video_mask, audio_mask)),
|
||||
}
|
||||
|
||||
|
||||
def _apply(mode: str):
|
||||
latent = _latent()
|
||||
original_samples = latent["samples"]
|
||||
original_video_mask, original_audio_mask = [
|
||||
value.clone() for value in latent["noise_mask"].unbind()
|
||||
]
|
||||
policy = policy_module.R28ContinuationPolicy(mode, tensor_summary=_summary)
|
||||
result = policy.prepare_masked_latent(
|
||||
latent=latent,
|
||||
physical_group=2,
|
||||
logical_chunks=(2,),
|
||||
context_frames=22,
|
||||
video_context=torch.zeros((1, 24, 7, 4, 4)),
|
||||
audio_context=torch.zeros((1, 32, 2, 37)),
|
||||
)
|
||||
assert result["samples"] is original_samples
|
||||
assert torch.equal(latent["noise_mask"].unbind()[0], original_video_mask)
|
||||
assert torch.equal(latent["noise_mask"].unbind()[1], original_audio_mask)
|
||||
return policy, result
|
||||
|
||||
|
||||
def test_depth_13_uses_last_four_slots_and_preserves_audio_contract():
|
||||
policy, result = _apply(policy_module.MODE_DEPTH_13)
|
||||
video_mask, audio_mask = result["noise_mask"].unbind()
|
||||
assert policy.active_video_slots == 4
|
||||
assert policy.active_video_frames == 13
|
||||
assert torch.all(video_mask[:, :, :3] == 1)
|
||||
assert torch.all(video_mask[:, :, 3:7] == 0)
|
||||
assert torch.all(video_mask[:, :, 7:] == 1)
|
||||
assert torch.all(audio_mask[..., :37] == 0)
|
||||
assert torch.all(audio_mask[..., 37:] == 1)
|
||||
|
||||
|
||||
def test_depth_9_uses_last_three_slots_only():
|
||||
policy, result = _apply(policy_module.MODE_DEPTH_9)
|
||||
video_mask, _ = result["noise_mask"].unbind()
|
||||
assert policy.active_video_slots == 3
|
||||
assert policy.active_video_frames == 9
|
||||
assert torch.all(video_mask[:, :, :4] == 1)
|
||||
assert torch.all(video_mask[:, :, 4:7] == 0)
|
||||
|
||||
|
||||
def test_strength_75_uses_core_fractional_mask_without_scaling_latent():
|
||||
latent = _latent()
|
||||
source_video, source_audio = latent["samples"].unbind()
|
||||
source_video_before = source_video.clone()
|
||||
source_audio_before = source_audio.clone()
|
||||
policy = policy_module.R28ContinuationPolicy(
|
||||
policy_module.MODE_STRENGTH_75, tensor_summary=_summary
|
||||
)
|
||||
result = policy.prepare_masked_latent(
|
||||
latent=latent,
|
||||
physical_group=2,
|
||||
logical_chunks=(2,),
|
||||
context_frames=22,
|
||||
video_context=torch.zeros((1, 24, 7, 4, 4)),
|
||||
audio_context=torch.zeros((1, 32, 2, 37)),
|
||||
)
|
||||
video_mask, audio_mask = result["noise_mask"].unbind()
|
||||
assert torch.all(video_mask[:, :, :7] == 0.25)
|
||||
assert torch.all(audio_mask[..., :37] == 0)
|
||||
assert torch.equal(source_video, source_video_before)
|
||||
assert torch.equal(source_audio, source_audio_before)
|
||||
|
||||
|
||||
def test_baseline_returns_original_latent_and_mask_objects():
|
||||
latent = _latent()
|
||||
policy = policy_module.R28ContinuationPolicy(
|
||||
policy_module.MODE_BASELINE_22, tensor_summary=_summary
|
||||
)
|
||||
result = policy.prepare_masked_latent(
|
||||
latent=latent,
|
||||
physical_group=2,
|
||||
logical_chunks=(2,),
|
||||
context_frames=22,
|
||||
video_context=torch.zeros((1, 24, 7, 4, 4)),
|
||||
audio_context=torch.zeros((1, 32, 2, 37)),
|
||||
)
|
||||
assert result is latent
|
||||
assert policy.active_video_slots == 7
|
||||
assert policy.active_video_frames == 22
|
||||
|
||||
|
||||
def _source_prompt() -> dict:
|
||||
return {
|
||||
"114": {"inputs": {"image": "old.png"}},
|
||||
"119": {"inputs": {}},
|
||||
"139": {"inputs": {"sampler_name": "old"}},
|
||||
"150": {"inputs": {"enabled": True}},
|
||||
"153": {"inputs": {"scheduler": "old", "steps": 20}},
|
||||
"188": {"inputs": {"value": "old"}},
|
||||
"191": {"inputs": {}},
|
||||
"249": {"inputs": {}},
|
||||
"286": {"inputs": {"lora_1": {"on": False}}},
|
||||
"297": {"inputs": {"value": 0.3}},
|
||||
"305": {
|
||||
"class_type": "H3ContinuumMaskedPrefixR1Diagnostic",
|
||||
"inputs": {
|
||||
"continuation_source_mode": "Recursive",
|
||||
"synchronize_sampling": True,
|
||||
"last_frame": ["900", 0],
|
||||
},
|
||||
},
|
||||
"306": {"inputs": {}},
|
||||
}
|
||||
|
||||
|
||||
def test_runner_changes_only_the_approved_r28_experiment_contract():
|
||||
prompt, loras = runner.configure_prompt(
|
||||
_source_prompt(),
|
||||
case="depth_13f",
|
||||
output_prefix="video/issue13_r28/test",
|
||||
image_name="closeup.png",
|
||||
seed=131325,
|
||||
width=576,
|
||||
height=864,
|
||||
diagnostic_nonce=132800,
|
||||
)
|
||||
inputs = prompt["305"]["inputs"]
|
||||
assert prompt["305"]["class_type"] == "H3ContinuumContinuationR28Diagnostic"
|
||||
assert inputs["continuation_experiment"] == "Video Depth 13f"
|
||||
assert inputs["continuation_transport"] == "masked_av_prefix_22_v1"
|
||||
assert inputs["continuity"] == "Balanced — 22 frames"
|
||||
assert inputs["audio_continuity"] is True
|
||||
assert inputs["chunks"] == 6
|
||||
assert inputs["chunk_seconds"] == 5.0
|
||||
assert "continuation_source_mode" not in inputs
|
||||
assert "synchronize_sampling" not in inputs
|
||||
assert "last_frame" not in inputs
|
||||
assert prompt["188"]["inputs"]["value"].count("\n---\n") == 5
|
||||
assert len(set(prompt["188"]["inputs"]["value"].split("\n---\n"))) == 1
|
||||
assert loras == [{"name": runner.r25.TURBO_LORA, "strength": 1.0}]
|
||||
|
||||
|
||||
def _analysis_case(scale: float) -> dict:
|
||||
rows = []
|
||||
for group in range(1, 7):
|
||||
row = {key: 0.0 for key in analyzer.base.METRIC_KEYS}
|
||||
for key in analyzer.DRIFT_KEYS:
|
||||
row[key] = float(group - 1) * scale
|
||||
rows.append(row)
|
||||
return {
|
||||
"end_trend": analyzer.base._trend(rows, analyzer.base.METRIC_KEYS),
|
||||
"boundary_metrics": [
|
||||
{"trajectory_flow_discontinuity": 0.1} for _ in range(5)
|
||||
],
|
||||
}
|
||||
|
||||
|
||||
def test_comparison_marks_a_clean_half_reduction_promising():
|
||||
result = analyzer._metric_comparison(_analysis_case(1.0), _analysis_case(0.5))
|
||||
assert result["screening_mean_reduction_percent"] == 50.0
|
||||
assert result["sharpness_not_worse"] is True
|
||||
assert result["gradient_not_worse"] is True
|
||||
assert result["numeric_promising"] is True
|
||||
Reference in New Issue
Block a user