File size: 10,863 Bytes
9fc9873 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 | from __future__ import annotations
import json
import tempfile
import unittest
import wave
import sqlite3
from pathlib import Path
from types import SimpleNamespace
from unittest import mock
import torch
from safetensors import safe_open
from scripts.data.cache_omni_hidden_states import (
SourcePreset,
assign_source_to_least_loaded_rank,
build_rank_source_sqlite_index,
complete_audio_window_count,
iter_audio_items,
iter_complete_audio_window_frame_ids,
instruction_for_item,
save_hidden_state_source_cpu,
source_hidden_file_is_complete,
source_index_row,
)
from psi.data.seed.full_duplex_audio_contract import AUDIO_COMMAND_STREAMING_PROMPT
def _write_pcm16_wav(path: Path, *, duration_seconds: float, sampling_rate: int = 24_000) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
sample_count = int(round(duration_seconds * sampling_rate))
with wave.open(str(path), "wb") as wav_file:
wav_file.setnchannels(1)
wav_file.setsampwidth(2)
wav_file.setframerate(sampling_rate)
wav_file.writeframes(b"\0\0" * sample_count)
def _window_args(
*,
full_audio: bool,
storage_format: str = "file",
limit_sources: int = 0,
) -> SimpleNamespace:
return SimpleNamespace(
audio_sampling_rate=16_000,
audio_command_full_audio=full_audio,
audio_command_window_seconds=0.6,
audio_command_sample_step_frames=20,
motion_fps=50,
storage_format=storage_format,
limit_sources=limit_sources,
)
class AudioCommandWindowTest(unittest.TestCase):
def test_complete_sources_are_balanced_by_window_count(self) -> None:
loads = [0, 0]
owners = [
assign_source_to_least_loaded_rank(loads, count)
for count in (100, 1, 100, 1)
]
self.assertEqual(owners, [0, 1, 1, 0])
self.assertEqual(loads, [101, 101])
def test_audio_command_prompt_is_identical_for_cache_and_live_inference(self) -> None:
first = instruction_for_item(
"qwen3_tts_seed_audio_command",
"audio_command_text",
{"filename": "walk", "content_short_description": "walk forward"},
)
second = instruction_for_item(
"hfd_white_noise",
"audio_command_text",
{"filename": "background"},
)
self.assertEqual(first, AUDIO_COMMAND_STREAMING_PROMPT)
self.assertEqual(second, AUDIO_COMMAND_STREAMING_PROMPT)
def test_complete_audio_windows_drop_incomplete_tail(self) -> None:
with tempfile.TemporaryDirectory() as tmp_dir:
tmp_path = Path(tmp_dir)
wav_path = tmp_path / "one_second.wav"
_write_pcm16_wav(wav_path, duration_seconds=1.0)
kwargs = {
"sampling_rate": 16_000,
"window_seconds": 0.6,
"step_frames": 20,
"motion_fps": 50,
}
self.assertEqual(complete_audio_window_count(wav_path, **kwargs), 2)
self.assertEqual(list(iter_complete_audio_window_frame_ids(wav_path, **kwargs)), [0, 20])
short_path = tmp_path / "short.wav"
_write_pcm16_wav(short_path, duration_seconds=0.4)
self.assertEqual(complete_audio_window_count(short_path, **kwargs), 0)
def test_audio_command_manifest_expands_unique_wavs_into_windows(self) -> None:
with tempfile.TemporaryDirectory() as tmp_dir:
tmp_path = Path(tmp_dir)
audio_dir = tmp_path / "seed_audio" / "000000"
_write_pcm16_wav(audio_dir / "long.wav", duration_seconds=1.0)
_write_pcm16_wav(audio_dir / "short.wav", duration_seconds=0.4)
manifest_path = tmp_path / "metadata.jsonl"
rows = [
{
"filename": "motion_a",
"content_short_description": "walk forward",
"audio_path": "seed_audio/000000",
},
{
"filename": "motion_b",
"content_short_description": "walk forward",
"audio_path": "seed_audio/000000",
},
]
manifest_path.write_text("".join(json.dumps(row) + "\n" for row in rows), encoding="utf-8")
preset = SourcePreset(
name="qwen3_tts_seed_audio_command",
modality="audio",
dataset_root=tmp_path,
manifest_path=manifest_path,
allowed_tasks=("audio_command_text",),
instruction_key="content_short_description",
is_audio_dir_manifest=True,
audio_file_extensions=(".wav",),
)
items = list(
iter_audio_items(
preset,
"audio_command_text",
tmp_path / "hidden",
_window_args(full_audio=False),
)
)
self.assertEqual([item.audio_start_sec for item in items], [0.0, 0.4])
self.assertEqual([item.audio_duration_sec for item in items], [0.6, 0.6])
self.assertEqual(
[item.output_path.name for item in items],
["long.f00000000.safetensors", "long.f00000020.safetensors"],
)
self.assertTrue(all(not item.full_audio for item in items))
legacy_items = list(
iter_audio_items(
preset,
"audio_command_text",
tmp_path / "hidden",
_window_args(full_audio=True),
)
)
self.assertEqual(
[item.output_path.name for item in legacy_items],
["long.safetensors", "short.safetensors"],
)
self.assertTrue(all(item.full_audio for item in legacy_items))
def test_source_layout_keeps_one_complete_wav_in_one_file(self) -> None:
with tempfile.TemporaryDirectory() as tmp_dir:
tmp_path = Path(tmp_dir)
audio_dir = tmp_path / "seed_audio" / "000000"
_write_pcm16_wav(audio_dir / "first.wav", duration_seconds=1.0)
_write_pcm16_wav(audio_dir / "second.wav", duration_seconds=1.0)
manifest_path = tmp_path / "metadata.jsonl"
manifest_path.write_text(
json.dumps(
{
"filename": "motion_a",
"content_short_description": "walk forward",
"audio_path": "seed_audio/000000",
}
)
+ "\n",
encoding="utf-8",
)
preset = SourcePreset(
name="qwen3_tts_seed_audio_command",
modality="audio",
dataset_root=tmp_path,
manifest_path=manifest_path,
allowed_tasks=("audio_command_text",),
instruction_key="content_short_description",
is_audio_dir_manifest=True,
audio_file_extensions=(".wav",),
)
output_root = tmp_path / "hidden"
items = list(
iter_audio_items(
preset,
"audio_command_text",
output_root,
_window_args(
full_audio=False,
storage_format="source",
limit_sources=1,
),
)
)
self.assertEqual(len(items), 2)
self.assertEqual([item.frame_id for item in items], [0, 20])
self.assertEqual(
[item.source_item_index for item in items],
[0, 1],
)
self.assertTrue(
all(item.source_item_count == 2 for item in items)
)
self.assertEqual(
{item.source_key for item in items},
{"seed_audio/000000/first"},
)
self.assertEqual(
len({item.source_output_path for item in items}),
1,
)
source_path = items[0].source_output_path
assert source_path is not None
self.assertEqual(
source_path.relative_to(output_root).as_posix(),
(
"audio/qwen3_tts_seed_audio_command/audio_command_text/"
"sources/seed_audio/000000/first.safetensors"
),
)
hiddens = [
torch.arange(12, dtype=torch.bfloat16).reshape(3, 4),
torch.arange(20, dtype=torch.bfloat16).reshape(5, 4),
]
with mock.patch(
"scripts.data.cache_omni_hidden_states.ACTIVE_MODEL_FINGERPRINT",
"unit-test-model",
):
stats = save_hidden_state_source_cpu(
items,
hiddens,
overwrite=False,
)
self.assertEqual(stats["written"], 2)
self.assertTrue(source_hidden_file_is_complete(items[0]))
with safe_open(
str(source_path),
framework="pt",
device="cpu",
) as reader:
self.assertEqual(
reader.get_tensor("offsets").tolist(),
[0, 3, 8],
)
self.assertEqual(
reader.get_tensor("frame_ids").tolist(),
[0, 20],
)
self.assertEqual(
tuple(reader.get_tensor("hidden_states").shape),
(8, 4),
)
row = source_index_row(items[0], output_root)
index_path = build_rank_source_sqlite_index(
output_root,
preset.name,
"audio_command_text",
"audio",
[row],
shard_rank=0,
)
with sqlite3.connect(
f"file:{index_path}?mode=ro",
uri=True,
) as conn:
indexed = conn.execute(
"SELECT key, path, count, window_seconds, step_frames "
"FROM sources"
).fetchone()
self.assertEqual(
indexed,
(
"seed_audio/000000/first",
"seed_audio/000000/first.safetensors",
2,
0.6,
20,
),
)
self.assertEqual(list(source_path.parent.glob("*.jsonl")), [])
if __name__ == "__main__":
unittest.main()
|