Spaces:
Running on Zero
Running on Zero
fix: the decoder choices are named by what they do: "Auto", "Steadier (diffusion decoder)" and "Sharper (standard decoder)", in the choice, its line, the Upscale lock line, the Done card and the backend's refusals; the values sent stay auto, diffusion and conv (#27)
1c8a58d unverified Download scripts/space_matrix.py from techfreakworm/LTX2.5-Studio: direct link, hf CLI and curl.
- Browser
- Download file 145 kB
-
https://huggingface.co/spaces/techfreakworm/LTX2.5-Studio/resolve/main/scripts/space_matrix.py
- Command line
-
hf download hf://spaces/techfreakworm/LTX2.5-Studio/scripts/space_matrix.py
-
curl -L -o space_matrix.py https://huggingface.co/spaces/techfreakworm/LTX2.5-Studio/resolve/main/scripts/space_matrix.py
145 kB
#!/usr/bin/env python
"""Run a plan of calls against the private LTX-2.5 Studio Space under a hard daily GPU budget.
# DRY RUN (the default): resolve, validate and price every call; contacts nothing
python scripts/space_matrix.py --plan scripts/space_plan.example.json \\
--assets-dir ~/assets --quota-ledger ~/space-ledger.jsonl
# the same plan, executed; when it stops at the ceiling (exit 3), the same command with
# --resume on the next UTC day runs what is left (calls with an outcome in --out are skipped)
python scripts/space_matrix.py --plan plan.json --assets-dir ~/assets \\
--quota-ledger ~/space-ledger.jsonl --out ~/space-runs/matrix --live [--resume]
# one call from flags; the historical 11-mode loop is --plan default
python scripts/space_matrix.py --mode text_to_video --width 704 --height 480 --num-frames 49
# read-only, one GET each, no GPU time
python scripts/space_matrix.py status
python scripts/space_matrix.py info
python scripts/space_matrix.py logs --out run.log
Exit codes: 0 plan complete · 2 illegal plan or contract mismatch (nothing spent) · 3 stopped at
the ceiling · 4 the platform refused on quota · 5 a call failed (with --keep-going the rest of the
plan still runs, and the exit is still 5).
Cost figures are BOUNDS. There is no readable ZeroGPU quota endpoint, so billed seconds cannot be
measured, only bounded: the conservative figure is the wall clock from the moment the Space starts
the event to its completion, times the size multiplier. The ceiling check before each call adds
that call's full reservation (``zerogpu.estimate_duration``, times the multiplier) to today's
conservative total.
The token is read when a command needs it (``HF_TOKEN``, else the Hub's cached token file), is
sent only as the ``Authorization`` header to the Space and the Hub API, and is redacted from
every line, file and error this script writes.
"""
from __future__ import annotations
import argparse
import hashlib
import html
import json
import os
import re
import shutil
import sys
import time
import traceback
import urllib.error
import urllib.parse
import urllib.request
import uuid
from contextlib import contextmanager
from dataclasses import dataclass, field, replace
from dataclasses import fields as dataclass_fields
from datetime import UTC, datetime, timedelta
from pathlib import Path
from typing import Any
REPO = Path(__file__).resolve().parent.parent
# `tests/` for `verify_streams`, the registry-driven stream checker.
for _path in (REPO, REPO / "tests"):
if str(_path) not in sys.path:
sys.path.insert(0, str(_path))
import verify_streams # noqa: E402
from ltx25_studio import constraints, modes, outputs, zerogpu # noqa: E402
from ltx25_studio.ui import fields # noqa: E402
__all__ = [
"API_FIELD_KEYS",
"DEFAULT_MODE_ORDER",
"DEFAULT_QUOTA_CEILING_SECONDS",
"ENDPOINT",
"EXIT_CALL_FAILED",
"EXIT_CEILING",
"EXIT_ILLEGAL",
"EXIT_OK",
"EXIT_QUOTA",
"GATE_ACCEPTED",
"GATE_DEFERRED",
"LORA_ENDPOINT",
"OUTPUT_SLOTS",
"UNBILLED_CATEGORIES",
"Call",
"Ledger",
"PlanError",
"QuotaLedgerUnreadable",
"Secret",
"app_gate",
"as_the_space",
"classify",
"count_onsets",
"default_plan",
"generate_contract_problems",
"load_plan",
"main",
"measure_audio",
"measure_video",
"platform_refusal",
"quota_multiplier",
"quota_refusal",
"read_token",
"redact",
"reservation_seconds",
"resume_split",
"returned_files",
"schedule_days",
"utc_day",
"verify_call",
]
# ---------------------------------------------------------------------------
# the Space and its published signature
# ---------------------------------------------------------------------------
DEFAULT_SPACE_ID = "techfreakworm/LTX2.5-Studio"
DEFAULT_HUB_URL = "https://huggingface.co"
ENDPOINT = "/on_generate"
LORA_ENDPOINT = "/on_validate_lora"
LORA_WEIGHT_ENDPOINT = "/on_lora_weight"
API_PREFIX = "/gradio_api"
#: ``/on_generate``'s file outputs, in order; the report follows them (``[Video, Audio, Markdown]``).
OUTPUT_SLOTS = ("video", "audio")
#: ``/on_generate``'s parameters after ``mode_id``: the field spec minus ``gr.State``, which
#: Gradio does not publish. Read again from ``fields.FIELDS`` at call time (see :func:`api_fields`)
#: so a field appended after import is still sent.
API_FIELD_KEYS: tuple[str, ...] = tuple(f.key for f in fields.FIELDS if f.kind != "state")
#: Field kind -> the component name Gradio publishes. Only ``text`` differs from its kind.
_COMPONENT_FOR_KIND = {"text": "textbox"}
# ---------------------------------------------------------------------------
# budget
# ---------------------------------------------------------------------------
DAILY_QUOTA_SECONDS = 2400.0
DEFAULT_QUOTA_CEILING_SECONDS = 1920.0
#: The Space's ``LTX25_ZEROGPU_SIZE``; ``zerogpu.quota_multiplier`` turns it into the quota factor.
DEFAULT_ZEROGPU_SIZE = "xlarge"
DEFAULT_CALL_TIMEOUT_SECONDS = 1800.0
DEFAULT_READ_TIMEOUT_SECONDS = 120.0
EXIT_OK = 0
EXIT_ILLEGAL = 2
EXIT_CEILING = 3
EXIT_QUOTA = 4
EXIT_CALL_FAILED = 5
#: How the ZeroGPU platform words its refusals (``spaces`` ``zero/client.py``, 0.50-0.52): each
#: error's title and the message under it. The app's failure card quotes only the MESSAGE; a
#: title reaches the API only when an error escapes the handler. Checked in this order.
_PLATFORM_REFUSALS: tuple[tuple[str, re.Pattern], ...] = (
# the PRO/free/IP quota, the runs limit, the Space-wide limit, pending credits
("quota", re.compile(
r"ZeroGPU quota exceeded"
r"|exceeded your [\w ]*?ZeroGPU quota"
r"|exceeded your ZeroGPU runs limit"
r"|exceeded [\w ]*GPU quota"
r"|reached its GPU limit"
r"|ZeroGPU pending credits exceeded"
r"|too many ZeroGPU credits allocated",
re.IGNORECASE,
)),
# the reservation is more than one call may ask for; nothing ran
("limit", re.compile(
r"ZeroGPU illegal duration|requested GPU duration \([\d.]+s\) is larger than the maximum allowed",
re.IGNORECASE,
)),
# no GPU was free in time; nothing ran
("queue", re.compile(r"ZeroGPU queue timeout|No GPU was available", re.IGNORECASE)),
)
#: Failure categories by their string values (``ltx25_studio.errors``). These two mean the
#: platform refused before a GPU was used: recorded with zero billed seconds.
UNBILLED_CATEGORIES = ("limit", "queue")
# ---------------------------------------------------------------------------
# content thresholds (measurements, reported with their thresholds; never stop a plan)
# ---------------------------------------------------------------------------
MIN_PIXEL_STD = 1.0
MIN_DISTINCT_FRAMES = 2
#: mean |Δ| over the frames' own spatial std; frozen probes measured <= 0.000523, the stillest
#: real post-fix render 0.004948.
MIN_RELATIVE_FRAME_DELTA = 0.001
LOW_MEAN_FRAME_DELTA = 0.5
#: The pinned vocoder's rate; ``expect.sample_rate`` overrides it per call.
PINNED_VOCODER_SAMPLING_RATE = 48000
# ---------------------------------------------------------------------------
# the historical default plan (--plan default)
# ---------------------------------------------------------------------------
CHEAPEST_SIZE = modes.MIN_RESOLUTION
CHEAPEST_FRAMES = 9
SEED = 1234
PROMPT = (
"A red sports car driving fast along a winding coastal road at sunset, "
"waves breaking on rocks below, cinematic camera following alongside"
)
#: Producers first: text_to_video makes the video (and, by frame, the image) the conditioned
#: modes chain from; text_to_audio makes the audio. Each chain is explicit in the plan.
DEFAULT_MODE_ORDER: tuple[str, ...] = (
"text_to_video",
"text_to_audio",
"text_to_audio_video",
"image_to_video",
"image_text_to_video",
"image_to_audio_video",
"image_text_to_audio_video",
"video_to_video",
"video_to_audio",
"audio_to_audio",
"audio_to_video",
)
MEDIA_KINDS = ("image", "video", "audio")
_CALL_KEYS = frozenset({"label", "mode", "loras", "expect", "note"})
_EXPECT_KEYS = {
"width": int,
"height": int,
"frames": int,
"has_audio": bool,
"has_video": bool,
"sample_rate": int,
"frame_rate": float,
}
_LABEL = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._-]{0,63}$")
_HUB_ID = re.compile(r"^[A-Za-z0-9][\w.-]*/[\w.-]+(/[\w./-]+)?$")
# ---------------------------------------------------------------------------
# clocks (module-level so tests can drive them)
# ---------------------------------------------------------------------------
def _utcnow() -> datetime:
return datetime.now(UTC)
def _monotonic() -> float:
return time.monotonic()
def utc_day(now: datetime | None = None) -> str:
return (now or _utcnow()).strftime("%Y-%m-%d")
def next_utc_midnight(now: datetime | None = None) -> datetime:
moment = now or _utcnow()
return datetime(moment.year, moment.month, moment.day, tzinfo=UTC) + timedelta(days=1)
# ---------------------------------------------------------------------------
# the token
# ---------------------------------------------------------------------------
_SECRETS: list[str] = []
REDACTED = "<redacted>"
class Secret:
"""The token. Never printed, pickled or formatted; :meth:`reveal` is read by the transport."""
__slots__ = ("_value",)
def __init__(self, value: str):
self._value = value
if value and value not in _SECRETS:
_SECRETS.append(value)
def reveal(self) -> str:
return self._value
def __repr__(self) -> str:
return f"Secret({REDACTED})"
__str__ = __repr__
def __reduce__(self):
raise TypeError("a Secret is not picklable")
def token_path() -> Path:
if os.environ.get("HF_TOKEN_PATH"):
return Path(os.environ["HF_TOKEN_PATH"]).expanduser()
home = os.environ.get("HF_HOME") or "~/.cache/huggingface"
return Path(home).expanduser() / "token"
def read_token() -> Secret | None:
"""``HF_TOKEN``, else the cached token file. Read at call time, never at import."""
value = (os.environ.get("HF_TOKEN") or "").strip()
if not value:
try:
value = token_path().read_text().strip()
except OSError:
value = ""
return Secret(value) if value else None
def redact(text: Any) -> str:
out = "" if text is None else str(text)
for secret in _SECRETS:
out = out.replace(secret, REDACTED)
return out
def _redacted_json(payload: Any) -> str:
return redact(json.dumps(payload, indent=2, sort_keys=False, default=str)) + "\n"
# ---------------------------------------------------------------------------
# errors
# ---------------------------------------------------------------------------
class PlanError(ValueError):
"""The plan cannot run as written. Nothing was sent."""
class ContractMismatch(RuntimeError):
"""The live Space's published signature is not what the local field spec builds."""
class QuotaLedgerUnreadable(RuntimeError):
"""The ledger exists and is not a ledger. Refused rather than read as zero."""
class HttpError(RuntimeError):
def __init__(self, code: int, url: str, body: str):
self.code = code
super().__init__(redact(f"HTTP {code} from {_strip_query(url)}: {body[:500]}"))
class TransportError(RuntimeError):
pass
class CallTimeout(RuntimeError):
pass
def _strip_query(url: str) -> str:
return url.split("?", 1)[0]
# ---------------------------------------------------------------------------
# HTTP
# ---------------------------------------------------------------------------
def _origin(url: str) -> str:
parts = urllib.parse.urlsplit(url)
return f"{parts.scheme}://{parts.netloc}".lower()
class Http:
"""urllib with the token attached only to trusted origins and never carried across a redirect."""
def __init__(self, secret: Secret | None, trusted: tuple[str, ...] = ()):
self._secret = secret
self._trusted = {_origin(u) for u in trusted}
self._opener = urllib.request.build_opener()
def trust(self, url: str) -> None:
self._trusted.add(_origin(url))
def open(self, method: str, url: str, *, data: bytes | None = None, headers=None, timeout: float = 60.0):
request = urllib.request.Request(url, data=data, method=method, headers=dict(headers or {}))
if self._secret is not None and _origin(url) in self._trusted:
# Unredirected: urllib drops it on any redirect, to any host.
request.add_unredirected_header("Authorization", f"Bearer {self._secret.reveal()}")
try:
return self._opener.open(request, timeout=timeout)
except urllib.error.HTTPError as exc:
try:
body = exc.read(2000).decode("utf-8", "replace")
except Exception: # noqa: BLE001
body = ""
raise HttpError(exc.code, url, body) from None
except (urllib.error.URLError, OSError) as exc:
raise TransportError(redact(f"{type(exc).__name__} reaching {_strip_query(url)}: {exc}")) from None
def get_json(self, url: str, *, timeout: float = 60.0) -> Any:
with self.open("GET", url, headers={"Accept": "application/json"}, timeout=timeout) as response:
return json.loads(response.read().decode("utf-8"))
def post_json(self, url: str, payload: Any, *, timeout: float = 60.0) -> Any:
body = json.dumps(payload).encode()
headers = {"Content-Type": "application/json", "Accept": "application/json"}
with self.open("POST", url, data=body, headers=headers, timeout=timeout) as response:
return json.loads(response.read().decode("utf-8"))
def iter_sse(response):
"""``data:`` payloads of an event stream, one parsed JSON value per event."""
buffer: list[str] = []
while True:
raw = response.readline()
if not raw:
if buffer:
yield json.loads("\n".join(buffer))
return
line = raw.decode("utf-8", "replace").rstrip("\r\n")
if not line:
if buffer:
text = "\n".join(buffer)
buffer = []
yield json.loads(text)
continue
if line.startswith("data:"):
buffer.append(line[5:].lstrip())
def space_host(space_id: str) -> str:
"""``owner/Name.x`` -> ``https://owner-name-x.hf.space``, the Space's own subdomain."""
slug = re.sub(r"[^a-z0-9]+", "-", space_id.lower()).strip("-")
return f"https://{slug}.hf.space"
class SpaceApp:
"""The Gradio app's HTTP surface: config, info, upload, queue, files."""
def __init__(self, base_url: str, http: Http, *, api_prefix: str = API_PREFIX):
self.base = base_url.rstrip("/")
self.http = http
self.prefix = "/" + api_prefix.strip("/") if api_prefix else ""
http.trust(self.base)
def url(self, path: str) -> str:
return f"{self.base}{self.prefix}{path}"
def config(self) -> dict:
return self.http.get_json(f"{self.base}/config")
def info(self) -> dict:
return self.http.get_json(self.url("/info"))
def upload(self, local: Path) -> dict:
boundary = uuid.uuid4().hex
content = Path(local).read_bytes()
name = Path(local).name
body = (
f"--{boundary}\r\nContent-Disposition: form-data; name=\"files\"; filename=\"{name}\"\r\n"
f"Content-Type: application/octet-stream\r\n\r\n"
).encode() + content + f"\r\n--{boundary}--\r\n".encode()
headers = {"Content-Type": f"multipart/form-data; boundary={boundary}", "Accept": "application/json"}
with self.http.open("POST", self.url("/upload"), data=body, headers=headers, timeout=600) as response:
paths = json.loads(response.read().decode("utf-8"))
return {"path": paths[0], "orig_name": name, "meta": {"_type": "gradio.FileData"}}
def join(self, fn_index: int, data: list, session_hash: str) -> str:
payload = {
"data": data,
"event_data": None,
"fn_index": fn_index,
"trigger_id": None,
"session_hash": session_hash,
}
return str(self.http.post_json(self.url("/queue/join"), payload)["event_id"])
def stream(self, session_hash: str, *, read_timeout: float):
query = urllib.parse.urlencode({"session_hash": session_hash})
return self.http.open(
"GET", self.url(f"/queue/data?{query}"), headers={"Accept": "text/event-stream"}, timeout=read_timeout
)
def download(self, server_path: str, dest: Path) -> Path:
url = self.url("/file=" + urllib.parse.quote(server_path, safe="/:"))
dest.parent.mkdir(parents=True, exist_ok=True)
with self.http.open("GET", url, timeout=600) as response, dest.open("wb") as handle:
shutil.copyfileobj(response, handle)
return dest
class Hub:
"""The Hub API's read-only Space endpoints."""
def __init__(self, hub_url: str, http: Http):
self.base = hub_url.rstrip("/")
self.http = http
http.trust(self.base)
def space(self, space_id: str) -> dict:
return self.http.get_json(f"{self.base}/api/spaces/{space_id}")
def logs(self, space_id: str, *, kind: str, max_seconds: float) -> list[str]:
lines: list[str] = []
deadline = _monotonic() + max_seconds
url = f"{self.base}/api/spaces/{space_id}/logs/{kind}"
try:
with self.http.open("GET", url, headers={"Accept": "text/event-stream"}, timeout=min(max_seconds, 10.0)) as response:
for event in iter_sse(response):
if isinstance(event, dict):
lines.append(f"{event.get('timestamp', '')} {event.get('data', '')}".strip())
else:
lines.append(str(event))
if _monotonic() >= deadline:
break
except TimeoutError:
pass
except TransportError as exc:
if "timed out" not in str(exc):
raise
return lines
# ---------------------------------------------------------------------------
# the plan
# ---------------------------------------------------------------------------
def api_fields() -> tuple:
"""Non-state fields, from the spec as it stands now."""
return tuple(f for f in fields.FIELDS if f.kind != "state")
def _field(key: str):
return next((f for f in fields.FIELDS if f.key == key), None)
def _accepts(spec, kind: str) -> bool:
predicate = getattr(modes, "accepts", None)
return bool(predicate(spec, kind)) if predicate else kind in spec.inputs
def _requires(spec, kind: str) -> bool:
predicate = getattr(modes, "requires", None)
return bool(predicate(spec, kind)) if predicate else kind in spec.inputs
def _skip_stage_2(spec) -> bool:
return bool(getattr(spec, "skip_stage_2", False))
def _varies_clip(spec) -> bool:
"""The clip is a variation source (mode 04): its length is the run's length."""
return getattr(spec, "video_conditioning", None) == "variation"
def _varies_audio(spec) -> bool:
return getattr(spec, "audio_conditioning", None) == "variation"
@dataclass(frozen=True)
class MediaInput:
kind: str
path: Path | None = None
from_call: str | None = None
#: image chains only: a frame index, ``"middle"`` or ``"last"``.
frame: int | str | None = None
@property
def chained(self) -> bool:
return self.from_call is not None
def describe(self) -> dict:
if self.chained:
out = {"chained": True, "from_call": self.from_call}
if self.kind == "image":
out["frame"] = self.frame
return out
size = self.path.stat().st_size if self.path and self.path.exists() else None
return {"chained": False, "path": str(self.path), "bytes": size}
#: ``lora.staging.DEFAULT_WEIGHT``: the strength ``/on_validate_lora`` stages an adapter at.
#: Read lazily (the LoRA package imports the Hub client); 1.0 where it cannot be read.
_STAGED_WEIGHT: list[float] = []
def staged_lora_weight() -> float:
if not _STAGED_WEIGHT:
try:
from ltx25_studio.lora import staging
_STAGED_WEIGHT.append(float(staging.DEFAULT_WEIGHT))
except Exception: # noqa: BLE001
_STAGED_WEIGHT.append(1.0)
return _STAGED_WEIGHT[0]
@dataclass(frozen=True)
class LoraInput:
source: str
weight: float = 1.0
local: Path | None = None
@property
def needs_weight_route(self) -> bool:
"""A strength other than the one ``/on_validate_lora`` stages at."""
return self.weight != staged_lora_weight()
@property
def label(self) -> str:
return self.local.name if self.local else self.source
def describe(self) -> dict:
return {
"source": self.source,
"route": "upload" if self.local else "hub",
"local": str(self.local) if self.local else None,
"weight": self.weight,
}
#: ``Call.gate``: the app's own gate accepted the call as the Space would run it, or the call
#: takes an input an earlier call produces, so the gate runs when that file exists.
GATE_ACCEPTED = "accepted"
GATE_DEFERRED = "deferred"
@dataclass(frozen=True)
class Call:
label: str
mode: str
values: dict
media: dict
loras: tuple = ()
expect: dict = field(default_factory=dict)
note: str = ""
#: :data:`GATE_ACCEPTED`, :data:`GATE_DEFERRED`, or "" before the gate was asked.
gate: str = ""
#: From ``generate.prepare_request``: the frame count the app will run (None: the model
#: chooses) and the notes it will carry ("length set from your audio ...").
app_frames: int | None = None
app_notes: tuple = ()
#: ``(key, sent, what the app runs)`` for every field the app's scoping changes.
app_changes: tuple = ()
#: The variation level the app snaps ``noise_scale`` to (modes 04 and 10), else None.
app_level: float | None = None
@property
def spec(self):
return modes.get_mode(self.mode)
def value(self, key: str) -> Any:
if key in self.values:
return self.values[key]
spec = _field(key)
return None if spec is None else spec.default
def runs(self, key: str) -> Any:
"""The value the app runs with: after its scoping when the gate ran, else as sent."""
for name, _sent, ran in self.app_changes:
if name == key:
return ran
return self.value(key)
@property
def chained(self) -> tuple[str, ...]:
return tuple(kind for kind, item in self.media.items() if item.chained)
@property
def min_resolution(self) -> bool:
return bool(getattr(self.spec, "min_resolution", False))
@property
def render_size(self) -> tuple[int, int]:
if self.min_resolution:
return (modes.MIN_RESOLUTION, modes.MIN_RESOLUTION)
return (int(self.value("width")), int(self.value("height")))
@property
def stage_2(self) -> bool:
return bool(self.value("enable_stage_2")) and not _skip_stage_2(self.spec)
@property
def exports(self) -> tuple[str, ...]:
return tuple(self.spec.exports)
def expected_geometry(self) -> tuple[int, int] | None:
if "video" not in self.exports:
return None
width, height = self.render_size
if self.stage_2:
width, height = 2 * width, 2 * height
return (int(self.expect.get("width", width)), int(self.expect.get("height", height)))
def expected_frames(self) -> int | None:
if "frames" in self.expect:
return int(self.expect["frames"])
if "video" not in self.exports:
return None
if self.app_frames is not None:
return self.app_frames
if self.value("auto_duration") or _varies_clip(self.spec):
# The model, the audio or the clip sets the length: known once the gate has run.
return None
return int(self.value("num_frames"))
def _check_value(spec, value, problems: list[str], where: str) -> Any:
"""``value`` as it is sent; a choice may be named by its label ("Sharper (standard decoder)")."""
kind = spec.kind
if kind == "text" and not isinstance(value, str):
problems.append(f"{where}: {spec.key} must be text")
elif kind == "checkbox" and not isinstance(value, bool):
problems.append(f"{where}: {spec.key} must be true or false")
elif kind in ("number", "slider"):
if isinstance(value, bool) or not isinstance(value, (int, float)):
problems.append(f"{where}: {spec.key} must be a number")
return value
low, high = spec.options.get("minimum"), spec.options.get("maximum")
if low is not None and value < low or high is not None and value > high:
problems.append(f"{where}: {spec.key}={value} is outside {low}..{high}")
elif kind in ("dropdown", "radio"):
choices = spec.options.get("choices")
if choices:
pairs = [tuple(c) if isinstance(c, (list, tuple)) else (c, c) for c in choices]
by_label = {label: sent for label, sent in pairs}
allowed = [sent for _label, sent in pairs]
if value not in allowed and value in by_label:
return by_label[value]
if value not in allowed:
shown = ", ".join(f"{sent!r} ({label})" if label != sent else repr(sent) for label, sent in pairs)
problems.append(f"{where}: {spec.key}={value!r} is not one of {shown}")
elif kind == "dataframe" and not isinstance(value, list):
problems.append(f"{where}: {spec.key} must be a list of rows")
return value
def _resolve_path(raw: str, base: Path) -> Path:
path = Path(raw).expanduser()
return path if path.is_absolute() else (base / path)
def _parse_call(raw: dict, *, index: int, base: Path, earlier: dict, problems: list[str]) -> Call | None:
before = len(problems)
where = f"call {index + 1}"
if not isinstance(raw, dict):
problems.append(f"{where}: must be an object")
return None
label = raw.get("label")
if not isinstance(label, str) or not _LABEL.match(label):
problems.append(f"{where}: 'label' must match {_LABEL.pattern} (it names the evidence folder)")
return None
where = f"call {label!r}"
if label in earlier:
problems.append(f"{where}: the label is used twice")
return None
try:
spec = modes.get_mode(str(raw.get("mode")))
except Exception as exc: # noqa: BLE001
problems.append(f"{where}: {exc}")
return None
values: dict = {}
media: dict = {}
loras: list[LoraInput] = []
expect: dict = {}
for key, value in raw.items():
if key in ("label", "mode", "note"):
continue
if key == "expect":
if not isinstance(value, dict):
problems.append(f"{where}: 'expect' must be an object")
continue
for name, wanted in value.items():
kind = _EXPECT_KEYS.get(name)
if kind is None:
problems.append(f"{where}: expect.{name} is not one of {sorted(_EXPECT_KEYS)}")
elif kind is bool and not isinstance(wanted, bool):
problems.append(f"{where}: expect.{name} must be true or false")
elif kind is not bool and (isinstance(wanted, bool) or not isinstance(wanted, (int, float))):
problems.append(f"{where}: expect.{name} must be a number")
else:
expect[name] = wanted
continue
if key == "loras":
if not isinstance(value, list):
problems.append(f"{where}: 'loras' must be a list of {{source, weight}}")
continue
for entry in value:
lora = _parse_lora(entry, base=base, where=where, problems=problems)
if lora is not None:
loras.append(lora)
continue
spec_field = _field(key)
if spec_field is None:
known = ", ".join(f.key for f in api_fields())
problems.append(
f"{where}: {key!r} is not a field of this app's {ENDPOINT} (fields: {known}; "
f"call keys: {', '.join(sorted(_CALL_KEYS))})"
)
continue
if spec_field.kind == "state":
problems.append(f"{where}: {key!r} is staged through {LORA_ENDPOINT}; use the call's 'loras' list")
continue
if spec_field.kind in MEDIA_KINDS:
item = _parse_media(spec_field.kind, value, base=base, where=where, earlier=earlier, problems=problems)
if item is not None:
media[spec_field.kind] = item
continue
values[key] = _check_value(spec_field, value, problems, where)
call = Call(label=label, mode=spec.id, values=values, media=media, loras=tuple(loras), expect=expect,
note=str(raw.get("note", "")))
for kind in MEDIA_KINDS:
if kind in media and not _accepts(spec, kind):
problems.append(f"{where}: {spec.id} takes no {kind}")
if kind not in media and _requires(spec, kind):
problems.append(
f"{where}: {spec.id} needs a {kind}: give a file under --assets-dir, or "
f'{{"from_call": "<earlier label>"}} to use an earlier call\'s output'
)
if str(values.get("prompt", "") or "").strip() and not _accepts(spec, "text"):
problems.append(f"{where}: {spec.id} takes no prompt; remove 'prompt'")
try:
width, height = call.render_size
constraints.validate_dimensions(width, height)
constraints.validate_dimension_limit(width, name="width")
constraints.validate_dimension_limit(height, name="height")
except Exception as exc: # noqa: BLE001
problems.append(f"{where}: {exc}")
try:
if not call.value("auto_duration"):
constraints.validate_num_frames(int(call.value("num_frames")))
constraints.validate_num_frames_limit(int(call.value("num_frames")))
else:
if float(call.value("min_seconds")) >= float(call.value("max_seconds")):
problems.append(f"{where}: max_seconds must be greater than min_seconds")
constraints.validate_seconds_limit(float(call.value("max_seconds")))
except Exception as exc: # noqa: BLE001
problems.append(f"{where}: {exc}")
if len(problems) > before:
return call
if call.chained:
# The app's gate opens its inputs; a file an earlier call produces does not exist yet.
# It runs on that file, before anything is sent, once the producer has run.
return replace(call, gate=GATE_DEFERRED)
try:
return gated(call, {kind: item.path for kind, item in media.items()})
except PlanError as exc:
problems.append(f"{where}: {exc}")
except Exception as exc: # noqa: BLE001 -- the app's own words
problems.append(f"{where}: the app would refuse this call before reserving a GPU: {exc}")
return call
@dataclass(frozen=True)
class Gate:
"""What ``generate.prepare_request`` said about one call."""
frames: int | None
notes: tuple[str, ...]
changes: tuple
level: float | None
class _Platform:
"""Stands for the ``spaces`` package the Space imports: ``zerogpu.is_zerogpu`` says yes."""
is_stub = False
@contextmanager
def as_the_space():
"""``device.current()`` answers as on the Space (ZeroGPU, CUDA) while the app's gate runs.
The gate refuses some requests by device (stage 2, and the diffusion decoder off CUDA); left
alone it would apply THIS machine's ``LTX25_DEVICE`` to a call that runs on the Space.
"""
from ltx25_studio import config, device
resolve = getattr(device, "resolve", None)
if resolve is None or not hasattr(device, "current"):
yield None
return
try:
settings = replace(config.load_config(), device="auto")
resolution = resolve(spaces_module=_Platform(), runtime_config=settings)
except Exception as exc: # noqa: BLE001 -- a local setting, not the Space's
raise PlanError(
f"the app's gate cannot be run as the Space runs it, because this machine's LTX25_* "
f"settings do not load: {exc}"
) from None
original = device.current
device.current = lambda: resolution
try:
yield resolution
finally:
device.current = original
def _space_input(kind: str, path) -> Any:
"""A media file as the Space's handler receives it: images as PIL (``gr.Image(type="pil")``),
video and audio as file paths."""
if kind != "image":
return str(path)
from PIL import Image
try:
with Image.open(path) as image:
return image.convert("RGB")
except Exception as exc: # noqa: BLE001
raise PlanError(
f"image {Path(path).name} could not be read as a picture ({type(exc).__name__}); the "
f"Space's image input would refuse it"
) from None
def _app_request(call: Call, media: dict):
"""The ``GenerationRequest`` the Space's handler would build from this call's fields."""
from ltx25_studio import generate
known = {f.name for f in dataclass_fields(generate.GenerationRequest)}
width, height = call.render_size
kwargs: dict = {"mode": call.mode}
for spec in api_fields():
if spec.key not in known:
continue
if spec.kind in MEDIA_KINDS:
path = media.get(spec.kind)
kwargs[spec.key] = None if path is None else _space_input(spec.kind, path)
elif spec.key in ("width", "height"):
kwargs[spec.key] = width if spec.key == "width" else height
elif spec.kind == "dataframe":
kwargs[spec.key] = tuple(call.value(spec.key) or ())
else:
kwargs[spec.key] = call.value(spec.key)
return generate.GenerationRequest(**kwargs)
def _variation_level(spec, scoped) -> float | None:
"""The level a variation mode runs at, as the mode's own ``build_kwargs`` decides it (the
nearest supported level to ``noise_scale``, or the mode's default); None for other modes."""
if not (_varies_clip(spec) or _varies_audio(spec)):
return None
kwargs = spec.build_kwargs(scoped)
clip = kwargs.get("video_variation")
level = getattr(clip, "level", None) if clip is not None else kwargs.get("audio_noise_level")
return None if level is None else float(level)
def app_gate(call: Call, media: dict) -> Gate:
"""``generate.prepare_request``, the checks the app runs before it reserves a GPU, run as
the Space runs them (:func:`as_the_space`), on the files themselves.
Raises what the Space would refuse the call with.
"""
from ltx25_studio import generate
request = _app_request(call, media)
with as_the_space():
scoped, spec, notes = generate.prepare_request(request)
changes = []
for key in (f.key for f in api_fields() if f.kind not in MEDIA_KINDS):
if not hasattr(request, key):
continue
sent, ran = getattr(request, key), getattr(scoped, key)
if sent != ran:
changes.append((key, sent, ran))
return Gate(
frames=None if scoped.auto_duration else int(scoped.num_frames),
notes=tuple(str(note) for note in notes),
changes=tuple(changes),
level=_variation_level(spec, scoped),
)
def gated(call: Call, media: dict) -> Call:
"""``call`` with what the app's gate said; raises its refusal."""
gate = app_gate(call, media)
return replace(call, gate=GATE_ACCEPTED, app_frames=gate.frames, app_notes=gate.notes,
app_changes=gate.changes, app_level=gate.level)
def _parse_media(kind, value, *, base, where, earlier, problems) -> MediaInput | None:
if isinstance(value, str):
path = _resolve_path(value, base)
if not path.is_file():
problems.append(f"{where}: {kind} file {path} does not exist")
return None
return MediaInput(kind=kind, path=path)
if isinstance(value, dict) and "from_call" in value:
source = value["from_call"]
extra = set(value) - {"from_call", "frame"}
if extra:
problems.append(f"{where}: {kind}: unknown key(s) {sorted(extra)}")
producer = earlier.get(source)
if producer is None:
problems.append(f"{where}: {kind} from_call {source!r} is not an earlier call in this plan")
return None
track = "audio" if kind == "audio" else "video"
if track not in producer.exports:
problems.append(f"{where}: {kind} from_call {source!r}: {producer.mode} exports no {track}")
return None
frame = value.get("frame", "middle") if kind == "image" else None
if kind == "image" and not (isinstance(frame, int) and not isinstance(frame, bool) and frame >= 0
or frame in ("middle", "last")):
problems.append(f"{where}: image frame must be an index >= 0, 'middle' or 'last'")
return None
if kind != "image" and "frame" in value:
problems.append(f"{where}: 'frame' applies to an image chain only")
return MediaInput(kind=kind, from_call=source, frame=frame)
problems.append(f"{where}: {kind} must be a file path or {{\"from_call\": \"<label>\"}}")
return None
def _parse_lora(entry, *, base, where, problems) -> LoraInput | None:
if isinstance(entry, str):
entry = {"source": entry}
if not isinstance(entry, dict) or not isinstance(entry.get("source"), str):
problems.append(f"{where}: each LoRA needs a 'source' (a Hub repo id or a local .safetensors)")
return None
extra = set(entry) - {"source", "weight"}
if extra:
problems.append(f"{where}: LoRA: unknown key(s) {sorted(extra)}")
weight = entry.get("weight", staged_lora_weight())
if isinstance(weight, bool) or not isinstance(weight, (int, float)) or weight <= 0:
problems.append(f"{where}: LoRA weight must be a positive number")
return None
if weight != staged_lora_weight():
try:
from ltx25_studio.lora import staging
staging.check_weight(weight, source=entry["source"])
except ImportError:
pass
except Exception as exc: # noqa: BLE001 -- the app's own words
problems.append(f"{where}: {exc}")
return None
source = entry["source"].strip()
candidate = _resolve_path(source, base)
if source.endswith(".safetensors") and candidate.is_file():
return LoraInput(source=source, weight=float(weight), local=candidate)
if _HUB_ID.match(source) and not source.startswith((".", "/", "~")):
return LoraInput(source=source, weight=float(weight))
problems.append(f"{where}: LoRA {source!r} is neither a file under --assets-dir nor a Hub repo id")
return None
def parse_plan(document: Any, *, base: Path) -> list[Call]:
problems: list[str] = []
defaults: dict = {}
if isinstance(document, dict):
unknown = set(document) - {"defaults", "calls", "note"}
if unknown:
problems.append(f"plan: unknown top-level key(s) {sorted(unknown)}")
defaults = document.get("defaults") or {}
entries = document.get("calls")
if not isinstance(defaults, dict):
problems.append("plan: 'defaults' must be an object")
defaults = {}
else:
entries = document
if not isinstance(entries, list) or not entries:
raise PlanError("the plan must be a non-empty list of calls (or {'defaults': {...}, 'calls': [...]})")
calls: dict[str, Call] = {}
for index, raw in enumerate(entries):
merged = {**defaults, **raw} if isinstance(raw, dict) else raw
if isinstance(raw, dict) and isinstance(defaults.get("expect"), dict) and isinstance(raw.get("expect"), dict):
merged["expect"] = {**defaults["expect"], **raw["expect"]}
call = _parse_call(merged, index=index, base=base, earlier=calls, problems=problems)
if call is not None:
calls[call.label] = call
if problems:
raise PlanError("\n".join(problems))
return list(calls.values())
def load_plan(path: str | Path, *, assets_dir: Path | None = None) -> list[Call]:
plan_path = Path(path).expanduser()
try:
document = json.loads(plan_path.read_text())
except (OSError, ValueError) as exc:
raise PlanError(f"{plan_path}: {exc}") from None
return parse_plan(document, base=assets_dir or plan_path.resolve().parent)
def default_plan(mode_ids=None, overrides: dict | None = None) -> list[Call]:
"""The historical 11-mode loop, with every chain written out."""
chosen = list(DEFAULT_MODE_ORDER if mode_ids is None else mode_ids)
entries: list[dict] = []
labels: dict[str, int] = {}
producers: dict[str, str] = {}
problems: list[str] = []
for mode_id in chosen:
try:
spec = modes.get_mode(mode_id)
except Exception as exc: # noqa: BLE001
raise PlanError(str(exc)) from None
labels[mode_id] = labels.get(mode_id, 0) + 1
label = mode_id if labels[mode_id] == 1 else f"{mode_id}-{labels[mode_id]}"
entry: dict = {"label": label, "mode": spec.id, "width": CHEAPEST_SIZE, "height": CHEAPEST_SIZE,
"num_frames": CHEAPEST_FRAMES, "seed": SEED}
if _accepts(spec, "text"):
entry["prompt"] = PROMPT
entry.update(overrides or {})
if not _accepts(spec, "text"):
entry.pop("prompt", None)
for kind in MEDIA_KINDS:
if not _requires(spec, kind):
continue
track = "audio" if kind == "audio" else "video"
source = producers.get(track)
if source is None:
problems.append(f"{label} needs a {kind}, and nothing before it in this plan produces {track}")
continue
entry[kind] = {"from_call": source, "frame": "middle"} if kind == "image" else {"from_call": source}
for track in spec.exports:
producers.setdefault(track, label)
entries.append(entry)
if problems:
raise PlanError("\n".join(problems))
return parse_plan(entries, base=Path.cwd())
# ---------------------------------------------------------------------------
# what a call costs
# ---------------------------------------------------------------------------
def quota_multiplier(size: str = DEFAULT_ZEROGPU_SIZE) -> float:
"""The app's own quota factor for the Space's GPU size."""
return float(zerogpu.quota_multiplier(size))
def _priced_request(call: Call):
"""The request the Space's estimator sees for this call: its fields as the app runs them,
with the length the app's gate set (a recording's or a clip's) once the gate has run."""
from ltx25_studio import generate
width, height = call.render_size
fields = {
"mode": call.mode,
"width": width,
"height": height,
"num_frames": int(call.runs("num_frames")),
"auto_duration": bool(call.runs("auto_duration")),
"min_seconds": float(call.runs("min_seconds")),
"max_seconds": float(call.runs("max_seconds")),
"noise_scale": call.app_level if call.app_level is not None else call.runs("noise_scale"),
}
if call.app_frames is not None:
fields.update(num_frames=int(call.app_frames), auto_duration=False)
if _varies_clip(call.spec):
fields["extras"] = {"length_source": generate.LENGTH_FROM_VIDEO}
return generate.GenerationRequest(**fields)
def reservation_seconds(call: Call, *, floor_seconds: int | None = None) -> int:
"""The app's own estimator for the request as the app runs it: its size and length, with
the Space's ``LTX25_GPU_DURATION_FLOOR_SECONDS`` as ``floor_seconds`` (unset by default)."""
return int(
zerogpu.estimate_duration(
request=_priced_request(call),
enable_stage_2=call.stage_2 and bool(call.runs("enable_stage_2")),
loras=[lora.describe() for lora in call.loras] or None,
enhance_prompt=bool(call.runs("enhance_prompt")),
gpu_duration_floor_seconds=floor_seconds,
)
)
def _reservation_parts(call: Call, *, floor_seconds: int | None) -> list[str]:
"""What the reservation is made of, in the estimator's own terms."""
request = _priced_request(call)
shape = zerogpu.run_shape(request=request)
plain = {"request": request, "loras": None, "enhance_prompt": False, "gpu_duration_floor_seconds": None}
run = zerogpu.estimate_duration(enable_stage_2=False, **plain)
parts = [f"{run} for {shape.width}x{shape.height}x{shape.frames}"]
if call.stage_2 and call.runs("enable_stage_2"):
parts.append(f"+{zerogpu.estimate_duration(enable_stage_2=True, **plain) - run} Upscale")
if call.loras:
parts.append(f"+{zerogpu.LORA_DURATION_SECONDS} LoRA")
if call.runs("enhance_prompt"):
parts.append(f"+{zerogpu.ENHANCE_DURATION_SECONDS} enhancer")
if floor_seconds:
parts.append(f"at least {floor_seconds}")
return parts
def platform_refusal(*texts: Any) -> tuple[str, str] | None:
"""``(category, the platform's words)`` for the first refusal wording in ``texts``, or ``None``.
Quota is looked for first, in every text, because it ends the whole invocation.
"""
plain = [html.unescape(str(text)) for text in texts if text]
for category, pattern in _PLATFORM_REFUSALS:
for text in plain:
found = pattern.search(text)
if found:
return category, found.group(0)
return None
def quota_refusal(*texts: Any) -> str | None:
"""The first platform quota-refusal wording in ``texts``, or ``None``."""
found = platform_refusal(*texts)
return found[1] if found and found[0] == "quota" else None
# ---------------------------------------------------------------------------
# the ledger: one JSON record per line, append-only, keyed by UTC day
# ---------------------------------------------------------------------------
class Ledger:
"""Append-only spend records. The day's total is the sum of ``billed_conservative``.
A call that may spend is written down BEFORE it is sent (``call_start``, at its full
reservation) and settled by its ``call`` record afterwards. A start with no settling record
(the invocation was killed mid-call) counts at its full reservation.
"""
SCHEMA = 3
UNIT = "GPU-seconds: an UPPER BOUND from wall clock x size multiplier, NOT billed seconds"
_SPENDING = ("call", "call_start")
def __init__(self, path: str | Path):
self.path = Path(path).expanduser()
def records(self) -> list[dict]:
if not self.path.exists():
return []
try:
text = self.path.read_text()
except OSError as exc:
raise QuotaLedgerUnreadable(f"{self.path}: {exc}") from None
records = []
for number, line in enumerate(text.splitlines(), start=1):
if not line.strip():
continue
try:
record = json.loads(line)
if not isinstance(record, dict) or "kind" not in record or "utc_day" not in record:
raise ValueError("not a ledger record")
if record["kind"] in self._SPENDING:
float(record["billed_conservative"])
except (ValueError, KeyError, TypeError) as exc:
raise QuotaLedgerUnreadable(
f"{self.path} line {number} is not a ledger record ({exc}). Refusing to read it "
f"as zero spend: a damaged ledger is not evidence that nothing was spent. This "
f"version writes one JSON record per line; fix or move the file by hand."
) from None
records.append(record)
return records
def _spending(self) -> list[dict]:
"""Settled calls, plus every start that was never settled."""
records = self.records()
settled = {r.get("call_id") for r in records if r["kind"] == "call" and r.get("call_id")}
return [r for r in records if r["kind"] == "call"
or r["kind"] == "call_start" and r.get("call_id") not in settled]
def unsettled(self) -> list[dict]:
return [r for r in self._spending() if r["kind"] == "call_start"]
def spent(self, day: str) -> float:
return sum(float(r["billed_conservative"]) for r in self._spending() if r["utc_day"] == day)
def previous_day(self, day: str) -> tuple[str, float] | None:
days: dict[str, float] = {}
for record in self._spending():
if record["utc_day"] < day:
days[record["utc_day"]] = days.get(record["utc_day"], 0.0) + float(record["billed_conservative"])
if not days:
return None
last = max(days)
return last, days[last]
def append(self, record: dict) -> None:
line = redact(json.dumps({"schema": self.SCHEMA, **record}, sort_keys=True)) + "\n"
self.path.parent.mkdir(parents=True, exist_ok=True)
descriptor = os.open(self.path, os.O_WRONLY | os.O_APPEND | os.O_CREAT, 0o600)
try:
os.write(descriptor, line.encode())
os.fsync(descriptor)
finally:
os.close(descriptor)
# ---------------------------------------------------------------------------
# the live contract
# ---------------------------------------------------------------------------
@dataclass
class Endpoint:
name: str
fn_index: int
input_types: list[str]
parameters: list[dict]
returns: list[dict]
def find_endpoint(config: dict, info: dict, name: str) -> Endpoint | None:
published = (info.get("named_endpoints") or {}).get(name)
dependency = next(
(d for d in config.get("dependencies", []) if "/" + str(d.get("api_name") or "") == name), None
)
if published is None or dependency is None:
return None
types = {c["id"]: str(c.get("type", "")).lower() for c in config.get("components", [])}
return Endpoint(
name=name,
fn_index=int(dependency["id"]),
input_types=[types.get(i, "?") for i in dependency.get("inputs", [])],
parameters=list(published.get("parameters") or []),
returns=list(published.get("returns") or []),
)
def _expected_component(kind: str) -> str:
return _COMPONENT_FOR_KIND.get(kind, kind)
def generate_contract_problems(parameters: list[dict], *, input_types: list[str] | None = None,
plan_modes=()) -> tuple[list[str], list[str]]:
"""``(problems, table)`` comparing ``/on_generate``'s published signature with the field spec."""
expected = [("mode", None, None)] + [(f.key, f.kind, f.label) for f in api_fields()]
problems: list[str] = []
table = [f" {'#':>2} {'expected':<44} live"]
for slot in range(max(len(expected), len(parameters))):
want = expected[slot] if slot < len(expected) else None
have = parameters[slot] if slot < len(parameters) else None
live = "-" if have is None else f"{have.get('component')} {have.get('label')!r}"
if want is None:
problems.append(f"slot {slot}: the Space publishes {live} that the field spec does not have")
table.append(f" {slot:>2} {'-':<44} {live} ✗")
continue
key, kind, label = want
shown = "mode (a choice)" if kind is None else f"{key} ({_expected_component(kind)} {label!r})"
mark = "✓"
if have is None:
problems.append(f"slot {slot}: the field spec sends {key!r} and the Space has no parameter there")
mark = "✗"
elif kind is None:
if str(have.get("component", "")).lower() not in ("radio", "dropdown"):
problems.append(f"slot 0 should be the mode selector, the Space has {live}")
mark = "✗"
else:
component = str(have.get("component", "")).lower()
if component != _expected_component(kind):
problems.append(f"slot {slot}: {key!r} is a {kind}; the Space has {live}")
mark = "✗"
elif label and have.get("label") and not str(have.get("label")).strip().startswith(str(label).strip()):
# A prefix, because the UI may add detail to a label (stage 2 says what size it returns).
# Labels are the only thing that tells two same-kind neighbours (width, height) apart.
problems.append(f"slot {slot}: {key!r} is labelled {label!r} here; the Space has {live}")
mark = "✗"
table.append(f" {slot:>2} {shown:<44} {live} {mark}")
if parameters:
choices = (parameters[0].get("type") or {}).get("enum")
if choices:
for mode_id in plan_modes:
if mode_id not in choices:
problems.append(f"the Space does not offer mode {mode_id!r} (it offers {choices})")
if input_types is not None:
want_types = ["mode", *(f.kind for f in fields.FIELDS)]
if len(input_types) != len(want_types):
problems.append(
f"{ENDPOINT} takes {len(input_types)} inputs on the Space; the field spec builds {len(want_types)}"
)
else:
want_state = [i for i, kind in enumerate(want_types) if kind == "state"]
have_state = [i for i, kind in enumerate(input_types) if kind == "state"]
if want_state != have_state:
problems.append(f"{ENDPOINT}'s State inputs sit at {have_state} on the Space, {want_state} here")
return problems, table
@dataclass
class LoraRoute:
endpoint: Endpoint
source_slot: int | None
upload_slot: int | None
weight_slot: int | None
weight_endpoint: Endpoint | None = None
_SOURCE_NAMES = {"repo_id", "repo", "source", "hub_id", "repo_or_url"}
def _lora_route(config: dict, info: dict) -> LoraRoute | None:
endpoint = find_endpoint(config, info, LORA_ENDPOINT)
if endpoint is None:
return None
source = upload = weight = None
for slot, parameter in enumerate(endpoint.parameters):
name = str(parameter.get("parameter_name") or "").lower()
component = str(parameter.get("component", "")).lower()
label = str(parameter.get("label") or "").lower()
if weight is None and component in ("slider", "number") and ("weight" in name or "strength" in name
or "strength" in label):
weight = slot
elif upload is None and (component in ("file", "uploadbutton") or "upload" in name):
upload = slot
elif source is None and component == "textbox" and (name in _SOURCE_NAMES or "repo" in label):
source = slot
weight_endpoint = find_endpoint(config, info, LORA_WEIGHT_ENDPOINT)
if weight_endpoint is not None:
names = [str(p.get("parameter_name") or "").lower() for p in weight_endpoint.parameters]
if "index" not in names or "weight" not in names:
weight_endpoint = None
return LoraRoute(endpoint, source, upload, weight, weight_endpoint)
def lora_contract_problems(route: LoraRoute | None, calls: list[Call]) -> list[str]:
problems: list[str] = []
with_loras = [c for c in calls if c.loras]
if not with_loras:
return problems
if route is None:
return [f"{', '.join(c.label for c in with_loras)} stage LoRAs, and the Space publishes no {LORA_ENDPOINT}"]
for call in with_loras:
for lora in call.loras:
if lora.local is not None and route.upload_slot is None:
problems.append(f"{call.label}: {LORA_ENDPOINT} has no upload parameter for {lora.source}")
if lora.local is None and route.source_slot is None:
problems.append(f"{call.label}: {LORA_ENDPOINT} has no Hub repo id parameter for {lora.source}")
if lora.needs_weight_route and route.weight_slot is None and route.weight_endpoint is None:
names = [p.get("parameter_name") for p in route.endpoint.parameters]
problems.append(
f"{call.label}: {lora.source} asks for strength {lora.weight:g}, and the Space exposes no "
f"public way to set an adapter's strength ({LORA_ENDPOINT} takes {names} and stages at "
f"{staged_lora_weight():g}; there is no {LORA_WEIGHT_ENDPOINT} with 'index' and 'weight'). "
f"Run the rest with --skip {call.label}, set its weight to {staged_lora_weight():g}, or "
f"publish that route."
)
return problems
def build_data(input_types: list[str], values: list) -> list:
"""Values for the non-State inputs in order, ``None`` in every State slot."""
remaining = list(values)
data = []
for kind in input_types:
if kind == "state":
data.append(None)
else:
if not remaining:
raise ContractMismatch(f"{len(values)} values for {input_types}")
data.append(remaining.pop(0))
if remaining:
raise ContractMismatch(f"{len(values)} values for {input_types}")
return data
def resolved_request(call: Call) -> dict:
"""Every non-state field by name, as it will be sent (media described, not uploaded)."""
out: dict = {}
width, height = call.render_size
for spec in api_fields():
if spec.kind in MEDIA_KINDS:
item = call.media.get(spec.kind)
out[spec.key] = None if item is None else item.describe()
elif spec.key == "width":
out[spec.key] = width
elif spec.key == "height":
out[spec.key] = height
else:
out[spec.key] = call.value(spec.key)
return out
def generate_values(call: Call, uploaded: dict) -> list:
values: list = [call.mode]
request = resolved_request(call)
for spec in api_fields():
if spec.kind in MEDIA_KINDS:
values.append(uploaded.get(spec.kind))
elif spec.kind == "dataframe":
values.append({"headers": list(spec.options.get("headers") or []), "data": request[spec.key] or [],
"metadata": None})
else:
values.append(request[spec.key])
return values
# ---------------------------------------------------------------------------
# measurements (reported with thresholds; none of them stops a plan)
# ---------------------------------------------------------------------------
def _decode_video_frames(path, *, limit=512):
import av
frames = []
with av.open(str(path)) as container:
if not container.streams.video:
return frames
for frame in container.decode(video=0):
frames.append(frame.to_ndarray(format="rgb24"))
if len(frames) >= limit:
break
return frames
def _check(name, value, threshold, verdict, detail=""):
return {"name": name, "value": value, "threshold": threshold, "verdict": verdict, "detail": detail}
def _verbatim(value: Any) -> Any:
"""Any value as JSON-able data, field for field, without interpreting it."""
import dataclasses
import enum
if dataclasses.is_dataclass(value) and not isinstance(value, type):
return {f.name: _verbatim(getattr(value, f.name)) for f in dataclasses.fields(value)}
if isinstance(value, enum.Enum):
return _verbatim(value.value)
if isinstance(value, dict):
return {str(k): _verbatim(v) for k, v in value.items()}
if isinstance(value, (list, tuple, set, frozenset)):
return [_verbatim(v) for v in value]
if value is None or isinstance(value, (str, int, float, bool)):
return value
if hasattr(value, "item"):
try:
return value.item()
except Exception: # noqa: BLE001
pass
if hasattr(value, "__dict__"):
return {k: _verbatim(v) for k, v in vars(value).items() if not k.startswith("_")}
return str(value)
def picture_health(frames) -> Any:
"""The app's own picture check (``outputs.assess_picture``), recorded verbatim.
The single call site. Nothing here judges what it returns: a fired condition is shown in the
summary and never fails a call.
"""
return _verbatim(outputs.assess_picture(frames))
def fired_conditions(health: Any) -> list[str]:
conditions = health.get("conditions") if isinstance(health, dict) else None
return [str(c) for c in conditions or [] if c]
def measure_video(path) -> dict:
"""Frame statistics (this harness's own thresholds) and the app's picture-health record."""
import numpy as np
try:
frames = _decode_video_frames(path)
except Exception as exc: # noqa: BLE001
return {"decoded_frames": 0, "checks": [_check("decodes", 0, "> 0 frames", "fail", f"{type(exc).__name__}: {exc}")]}
if not frames:
return {"decoded_frames": 0, "checks": [_check("decodes", 0, "> 0 frames", "fail", "no video frames")]}
stack = np.stack(frames).astype("float32")
deltas = [float(np.abs(stack[i + 1] - stack[i]).mean()) for i in range(len(frames) - 1)] or [0.0]
spatial_std = float(np.median([frame.std() for frame in stack]))
mean_delta = float(sum(deltas) / len(deltas))
relative = None if spatial_std < MIN_PIXEL_STD else mean_delta / spatial_std
distinct = len({frame.tobytes() for frame in frames})
checks = [
_check("flatness", round(spatial_std, 4), f">= {MIN_PIXEL_STD} median per-frame std",
"pass" if spatial_std >= MIN_PIXEL_STD else "fail"),
_check("distinct_frames", distinct, f">= {MIN_DISTINCT_FRAMES}",
"pass" if distinct >= MIN_DISTINCT_FRAMES or len(frames) == 1 else "fail"),
_check("freeze_floor", None if relative is None else round(relative, 6),
f">= {MIN_RELATIVE_FRAME_DELTA} mean |delta| / spatial std",
"n/a" if relative is None or len(frames) < 2 else
("pass" if relative >= MIN_RELATIVE_FRAME_DELTA else "fail")),
_check("motion", round(mean_delta, 4), f">= {LOW_MEAN_FRAME_DELTA} mean |delta| (warn only)",
"pass" if mean_delta >= LOW_MEAN_FRAME_DELTA else "warn",
"a still-conditioned mode is expected to land low"),
]
return {
"decoded_frames": len(frames),
"distinct_frames": distinct,
"spatial_std": spatial_std,
"stack_std": float(stack.std()),
"mean_abs_delta": mean_delta,
"min_abs_delta": min(deltas),
"max_abs_delta": max(deltas),
"relative_delta": relative,
"checks": checks,
"picture_health": picture_health(frames),
}
#: Audio measurements. Reported, never judged: only digital silence fails a call.
SILENT_DBFS = -70.0
SILENCE_WINDOW_SECONDS = 0.02
SILENCE_WINDOW_DBFS = -60.0
CLIP_LEVEL = 0.999
SPECTRUM_FFT = 2048
SPECTRUM_MAX_FRAMES = 4096
ENERGY_BANDS_HZ = ((0, 250), (250, 2000), (2000, 6000), (6000, None))
#: An onset: the level of a short window rises at least ONSET_RISE_DB above the lowest level of
#: the windows just before it. Steady noise moves well under 2 dB window to window at this size;
#: speech, hits and chimes rise far more.
ONSET_WINDOW_SECONDS = 0.01
ONSET_LOOKBACK_SECONDS = 0.05
ONSET_RISE_DB = 8.0
ONSET_MIN_GAP_SECONDS = 0.1
def _dbfs(rms: float) -> float:
import math
return 20.0 * math.log10(rms + 1e-12)
def _decode_audio_track(path, index: int):
"""One audio stream as float32 ``[channels, samples]`` and its rate."""
import av
import numpy as np
with av.open(str(path)) as container:
stream = container.streams.audio[index]
layout = getattr(getattr(stream, "layout", None), "name", None) or "stereo"
resampler = av.AudioResampler(format="fltp", layout=layout, rate=int(stream.rate))
blocks = []
for frame in container.decode(stream):
for out in resampler.resample(frame):
blocks.append(out.to_ndarray())
for out in resampler.resample(None):
blocks.append(out.to_ndarray())
rate = int(stream.rate)
if not blocks:
return np.zeros((1, 0), dtype="float32"), rate
return np.concatenate(blocks, axis=-1).astype("float32"), rate
def _spectral(mono, rate: int) -> dict:
import numpy as np
if mono.size < SPECTRUM_FFT:
mono = np.pad(mono, (0, SPECTRUM_FFT - mono.size))
hop = SPECTRUM_FFT // 2
count = 1 + (mono.size - SPECTRUM_FFT) // hop
starts = hop * np.unique(np.linspace(0, count - 1, min(count, SPECTRUM_MAX_FRAMES)).astype(int))
frames = mono[starts[:, None] + np.arange(SPECTRUM_FFT)[None, :]] * np.hanning(SPECTRUM_FFT)
power = (np.abs(np.fft.rfft(frames, axis=1)) ** 2)[:, 1:]
freqs = np.fft.rfftfreq(SPECTRUM_FFT, 1.0 / rate)[1:]
mean = power.mean(axis=0)
total = float(mean.sum())
if total <= 0.0:
return {"spectral_flatness": None, "spectral_flatness_frame_median": None,
"spectral_centroid_hz": None, "band_energy_share": None, "band_level_dbfs": None}
tiny = 1e-20
def flatness(spectrum):
return float(np.exp(np.mean(np.log(spectrum + tiny))) / (np.mean(spectrum) + tiny))
live = power[power.sum(axis=1) > total * 1e-6]
mean_square = float(np.mean(mono.astype("float64") ** 2))
bands, levels = {}, {}
for low, high in ENERGY_BANDS_HZ:
mask = (freqs >= low) & (freqs < high) if high else (freqs >= low)
name = f"{low}-{high or 'nyquist'}"
share = float(mean[mask].sum()) / total
bands[name] = round(share, 4)
levels[name] = round(10.0 * np.log10(share * mean_square + 1e-12), 2)
return {
"spectral_flatness": round(flatness(mean), 4),
"spectral_flatness_frame_median": round(float(np.median([flatness(f) for f in live])), 4) if len(live) else None,
"spectral_centroid_hz": round(float((freqs * mean).sum() / total), 1),
"band_energy_share": bands,
"band_level_dbfs": levels,
}
def count_onsets(mono, rate: int) -> int:
"""Rises of the short-window level of at least :data:`ONSET_RISE_DB` over the lowest level of
the preceding :data:`ONSET_LOOKBACK_SECONDS`, out of silence or not, at least
:data:`ONSET_MIN_GAP_SECONDS` apart."""
import numpy as np
window = max(1, int(round(ONSET_WINDOW_SECONDS * rate)))
count = mono.size // window
if count < 2:
return 0
frames = mono[: count * window].reshape(count, window).astype("float64")
level = 10.0 * np.log10(np.mean(frames**2, axis=1) + 1e-12)
back = max(1, int(round(ONSET_LOOKBACK_SECONDS / ONSET_WINDOW_SECONDS)))
gap = max(1, int(round(ONSET_MIN_GAP_SECONDS / ONSET_WINDOW_SECONDS)))
onsets, last = 0, -gap
for index in range(1, count):
floor = level[max(0, index - back) : index].min()
if level[index] >= SILENCE_WINDOW_DBFS and level[index] - floor >= ONSET_RISE_DB and index - last >= gap:
onsets += 1
last = index
return onsets
def measure_audio(path) -> dict:
"""Every audio stream of the file, as measurements. No verdicts."""
import av
import numpy as np
try:
with av.open(str(path)) as container:
count = len(container.streams.audio)
except Exception as exc: # noqa: BLE001
return {"tracks": [], "error": f"{type(exc).__name__}: {exc}"}
tracks = []
for index in range(count):
try:
samples, rate = _decode_audio_track(path, index)
except Exception as exc: # noqa: BLE001
tracks.append({"index": index, "error": f"{type(exc).__name__}: {exc}"})
continue
channels, length = samples.shape
if length == 0:
tracks.append({"index": index, "sample_rate": rate, "channels": channels, "samples": 0,
"level_dbfs": _dbfs(0.0)})
continue
rms = float(np.sqrt(np.mean(samples.astype("float64") ** 2)))
peak = float(np.abs(samples).max())
window = max(1, int(round(SILENCE_WINDOW_SECONDS * rate)))
usable = (length // window) * window
if usable:
windows = samples[:, :usable].reshape(channels, -1, window)
window_rms = np.sqrt(np.mean(windows.astype("float64") ** 2, axis=(0, 2)))
silence = float(np.mean(20.0 * np.log10(window_rms + 1e-12) < SILENCE_WINDOW_DBFS))
else:
silence = None
clipped = int(np.count_nonzero(np.abs(samples) >= CLIP_LEVEL))
mono = samples.mean(axis=0)
onsets = count_onsets(mono, rate)
tracks.append({
"index": index,
"sample_rate": rate,
"channels": channels,
"samples": length,
"seconds": round(length / rate, 4),
"level_dbfs": round(_dbfs(rms), 2),
"rms": rms,
"peak": peak,
"peak_dbfs": round(_dbfs(peak), 2),
"clipped_samples": clipped,
"clipped_ratio": clipped / float(samples.size),
"silence_ratio": silence,
"onsets": onsets,
"onsets_per_second": round(onsets / (length / rate), 3),
**_spectral(mono, rate),
})
return {
"tracks": tracks,
"definitions": {
"level_dbfs": "RMS over every sample of every channel",
"clipped_samples": f"samples with |x| >= {CLIP_LEVEL} of full scale",
"silence_ratio": f"share of {SILENCE_WINDOW_SECONDS * 1000:g} ms windows below {SILENCE_WINDOW_DBFS:g} dBFS",
"onsets": f"rises of the {ONSET_WINDOW_SECONDS * 1000:g} ms level by >= {ONSET_RISE_DB:g} dB over the lowest "
f"level of the preceding {ONSET_LOOKBACK_SECONDS * 1000:g} ms, >= "
f"{ONSET_MIN_GAP_SECONDS * 1000:g} ms apart (steady noise has none; speech and hits many)",
"spectral_flatness": "geometric / arithmetic mean of the mean power spectrum of the channels' mean "
"(about 1 for white noise, far below 1 for speech or music)",
"spectral_flatness_frame_median": f"the same per {SPECTRUM_FFT}-sample frame, median over frames",
"spectral_centroid_hz": "power-weighted mean frequency",
"band_energy_share": "share of spectral energy per band, Hz",
"band_level_dbfs": "each band's energy as a level, dBFS (the bands add up to the channels' mean)",
},
"note": "Measurements only: nothing here is a verdict. Only digital silence fails a call.",
}
def _audio_line(measured: dict) -> str:
parts = []
for track in measured.get("tracks", []):
if "error" in track:
parts.append(f"track {track['index']}: {track['error']}")
continue
flat = track.get("spectral_flatness")
centroid = track.get("spectral_centroid_hz")
silence = track.get("silence_ratio")
bands = track.get("band_energy_share") or {}
parts.append(
f"{track['level_dbfs']:.1f} dBFS · peak {track.get('peak_dbfs', 0):.1f} · clipped "
f"{track.get('clipped_samples', 0)} · silence {0 if silence is None else 100 * silence:.0f}% · flatness "
f"{'n/a' if flat is None else f'{flat:.3f}'} · centroid "
f"{'n/a' if centroid is None else f'{centroid:.0f} Hz'} · onsets {track.get('onsets', 'n/a')}"
+ (" · bands " + " ".join(f"{name} {100 * share:.0f}%" for name, share in bands.items()) if bands else "")
)
return " | ".join(parts)
def _count_frames(path) -> int | None:
try:
return len(_decode_video_frames(path, limit=100_000))
except Exception: # noqa: BLE001
return None
def returned_files(spec) -> dict[str, tuple[str, ...]]:
"""The file the app returns in each output slot for this mode, by the tracks it carries.
From the registry's ``exports``, as the app returns them: a slot holds a file exactly when
the mode exports that track (``ui.handlers``). The video file carries every exported track;
beside it, a mode that also exports audio returns the same sound as a file of its own,
audio only (``outputs.write_output``). An audio-only mode's one file is its audio.
"""
exported = [slot for slot in OUTPUT_SLOTS if slot in spec.exports]
return {slot: tuple(exported) if index == 0 else (slot,) for index, slot in enumerate(exported)}
def _sound_on_its_own(path, *, sampling_rate: int | None) -> dict:
"""The sound beside the video file, judged as audio only: an audio stream and no video,
stereo, at the expected rate. A file that cannot be probed is never a pass."""
expected = {"video": False, "audio": True, "audio_channels": verify_streams.EXPECTED_AUDIO_CHANNELS}
if sampling_rate is not None:
expected["sampling_rate"] = sampling_rate
actual = verify_streams.probe_streams(path)
if actual is None:
return {"probed": False, "ok": False, "expected": expected, "actual": {}, "problems": [
"it could not be opened as a media container, so nothing about the sound on its own has been "
"established. This is NOT a pass"]}
problems = []
if not actual["audio"]:
problems.append("the sound on its own has NO audio stream")
if actual["video"]:
problems.append("the sound on its own has a video stream: the audio output must be the sound by itself")
channels = actual.get("audio_channels")
if actual["audio"] and channels is not None and channels != verify_streams.EXPECTED_AUDIO_CHANNELS:
problems.append(f"audio has {channels} channel(s), expected {verify_streams.EXPECTED_AUDIO_CHANNELS} — "
f"something downmixed the vocoder's stereo output")
if actual["audio"] and sampling_rate is not None and actual.get("sampling_rate") not in (None, sampling_rate):
problems.append(f"audio sample rate is {actual.get('sampling_rate')}, expected {sampling_rate} — the file "
f"disagrees with the vocoder")
return {"probed": True, "ok": not problems, "expected": expected, "actual": actual, "problems": problems}
def _described(slot: str, tracks: tuple[str, ...], *, beside: bool) -> str:
if beside:
return "its sound on its own beside the video, as an audio-only file"
return f"a {slot} file with " + " and ".join(tracks)
def _seconds(measured: dict | None) -> float | None:
"""The first audio track's length, from its sample count and rate."""
tracks = (measured or {}).get("tracks") or [{}]
samples, rate = tracks[0].get("samples"), tracks[0].get("sample_rate")
return samples / rate if samples is not None and rate else None
def verify_call(call: Call, artifacts: dict[str, Path], *, sampling_rate: int | None) -> dict:
"""What the mode promises, judged over the set of files the call returned, plus measurements.
Each file is judged by what its slot holds (:func:`returned_files`): the video file against
every track the mode exports (``verify_streams``, with the geometry); the sound beside it as
audio only, and its length against the video's own sound. A promised file that did not come
back, and a file the mode never returns, are mismatches.
"""
problems: list[str] = []
streams = []
rate = call.expect.get("sample_rate", sampling_rate)
geometry = call.expected_geometry()
files = returned_files(call.spec)
primary = next(iter(files))
for slot in dict.fromkeys([*OUTPUT_SLOTS, *artifacts]):
path, tracks = artifacts.get(slot), files.get(slot)
if path is None:
if tracks:
problems.append(f"no {slot} file came back: {call.mode} returns "
f"{_described(slot, tracks, beside=slot != primary)}")
continue
if tracks is None:
actual = verify_streams.probe_streams(path)
entry = {"judged_as": f"a file {call.mode} never returns", "probed": actual is not None, "ok": False,
"expected": None, "actual": actual or {}, "problems": [
f"{call.mode} returns no {slot} file (it exports {' and '.join(call.exports)}), and this "
f"one came back as its {slot} output: a file the mode never returns is not a pass"]}
elif slot == primary:
verdict = verify_streams.verify(
call.mode,
path,
frame_rate=call.expect.get("frame_rate"),
sampling_rate=rate if "audio" in tracks else None,
geometry=geometry if "video" in tracks else None,
)
entry = {"judged_as": f"every track {call.mode} exports: {', '.join(tracks)}", "probed": verdict.probed,
"ok": verdict.ok, "expected": verdict.expected, "actual": verdict.actual,
"problems": verdict.problems}
elif path == artifacts.get(primary):
actual = verify_streams.probe_streams(path)
entry = {"judged_as": "the sound on its own: audio only", "probed": actual is not None, "ok": False,
"expected": None, "actual": actual or {}, "problems": [
f"the {primary} file came back as the {slot} output too: {call.mode} returns its sound "
f"there as a file of its own, audio only"]}
else:
entry = {"judged_as": "the sound on its own: audio only", **_sound_on_its_own(path, sampling_rate=rate)}
streams.append({"slot": slot, "path": str(path), **entry})
problems += [f"{Path(path).name}: {p}" for p in entry["problems"]]
if primary in artifacts:
probe = next(s["actual"] for s in streams if s["slot"] == primary) or {}
for track in ("video", "audio"):
wanted = call.expect.get(f"has_{track}")
if wanted is not None and bool(probe.get(track)) != bool(wanted):
problems.append(f"expected has_{track}={wanted}, the {primary} file has {track}={bool(probe.get(track))}")
frames = None
if "video" in artifacts:
frames = _count_frames(artifacts["video"])
wanted_frames = call.expected_frames()
if wanted_frames is not None and frames != wanted_frames:
problems.append(f"decoded {frames} frames, expected {wanted_frames}")
content: dict = {}
if "video" in artifacts:
content["video"] = measure_video(artifacts["video"])
if "audio" in artifacts:
content["audio"] = measure_audio(artifacts["audio"])
if primary == "video" and "audio" in files[primary] and primary in artifacts \
and artifacts[primary] != artifacts.get("audio"):
# The video file's own sound, which the sound on its own must match.
content["audio_in_video"] = measure_audio(artifacts[primary])
for key, slot in (("audio", "audio"), ("audio_in_video", primary)):
for track in (content.get(key) or {}).get("tracks", []):
level = track.get("level_dbfs")
if level is not None and level < SILENT_DBFS:
problems.append(f"{Path(artifacts[slot]).name}: audio track {track['index']} is digital silence: "
f"{level:.1f} dBFS < {SILENT_DBFS:g}")
lengths = None
own, inside = _seconds(content.get("audio")), _seconds(content.get("audio_in_video"))
if own is not None and inside is not None:
tolerance = constraints.DEFAULT_LENGTH_TOLERANCE_SECONDS
lengths = {"sound_on_its_own_seconds": round(own, 4), "sound_in_video_seconds": round(inside, 4),
"difference_seconds": round(abs(own - inside), 4), "tolerance_seconds": tolerance}
if abs(own - inside) > tolerance:
problems.append(
f"{Path(artifacts['audio']).name} is {own:.3f} s and the sound in {Path(artifacts[primary]).name} "
f"{inside:.3f} s: the sound on its own does not match the video's sound (the app allows "
f"{tolerance:g} s between a run's audio and video)"
)
return {
"ok": not problems,
"problems": problems,
"expected": {"geometry": list(geometry) if geometry else None, "frames": call.expected_frames(),
"sample_rate": rate if "audio" in call.exports else None,
"files": {slot: list(tracks) for slot, tracks in files.items()},
**{k: v for k, v in call.expect.items() if k.startswith("has_")}},
"decoded_frames": frames,
"streams": streams,
"sound_lengths": lengths,
"content": content,
"content_note": (
"Content figures are measurements. The frame checks carry this harness's thresholds, the "
"picture-health record is the app's own check verbatim, and audio has no verdict. None of "
"them stops a plan; no artifact is a pass until a frame has been looked at."
),
}
def _flags(verify: dict) -> list[str]:
flagged = []
content = verify.get("content") or {}
for check in (content.get("video") or {}).get("checks", []):
if check["verdict"] in ("fail", "warn"):
flagged.append(f"video.{check['name']}={check['verdict']}")
for condition in fired_conditions((content.get("video") or {}).get("picture_health")):
flagged.append(f"picture: {condition}")
return flagged
# ---------------------------------------------------------------------------
# chained inputs
# ---------------------------------------------------------------------------
def extract_image(video: Path, dest: Path, frame) -> int:
from PIL import Image
frames = _decode_video_frames(video, limit=100_000)
if not frames:
raise RuntimeError(f"{video} has no frames to take an image from")
index = {"middle": len(frames) // 2, "last": len(frames) - 1}.get(frame, frame)
if not isinstance(index, int) or not 0 <= index < len(frames):
raise RuntimeError(f"{video} has {len(frames)} frames; frame {frame!r} does not exist")
dest.parent.mkdir(parents=True, exist_ok=True)
Image.fromarray(frames[index]).save(str(dest))
return index
def extract_audio(source: Path, dest: Path) -> Path:
import av
dest.parent.mkdir(parents=True, exist_ok=True)
with av.open(str(source)) as probe:
if not probe.streams.audio:
raise RuntimeError(f"{source} has no audio stream")
stream = probe.streams.audio[0]
already = probe.format.name in ("wav", "wave") and stream.codec_context.name.startswith("pcm")
if already:
shutil.copyfile(source, dest)
return dest
with av.open(str(source)) as container, av.open(str(dest), "w", format="wav") as out:
stream = container.streams.audio[0]
layout = getattr(getattr(stream, "layout", None), "name", None) or "stereo"
out_stream = out.add_stream("pcm_s16le", rate=int(stream.rate), layout=layout)
for frame in container.decode(stream):
frame.pts = None
for packet in out_stream.encode(frame):
out.mux(packet)
for packet in out_stream.encode():
out.mux(packet)
return dest
def _sha256(path: Path) -> str:
digest = hashlib.sha256()
with Path(path).open("rb") as handle:
for chunk in iter(lambda: handle.read(1 << 20), b""):
digest.update(chunk)
return digest.hexdigest()
# ---------------------------------------------------------------------------
# one live call
# ---------------------------------------------------------------------------
class Evidence:
"""``--out/<label>/``: every file written here is redacted."""
def __init__(self, directory: Path):
self.dir = directory
self.dir.mkdir(parents=True, exist_ok=True)
self._events = (self.dir / "events.jsonl").open("a", encoding="utf-8")
def event(self, record: dict) -> None:
self._events.write(redact(json.dumps(record, sort_keys=True, default=str)) + "\n")
self._events.flush()
def write(self, name: str, payload: Any) -> Path:
path = self.dir / name
path.write_text(_redacted_json(payload))
return path
def close(self) -> None:
self._events.close()
@dataclass
class Outcome:
label: str
status: str # ok | mismatch | failed | timeout | quota | skipped
detail: str = ""
category: str | None = None
billed_conservative: float = 0.0
reserved: int = 0
artifacts: dict = field(default_factory=dict)
flags: list = field(default_factory=list)
audio: str = ""
@property
def failed(self) -> bool:
return self.status in ("mismatch", "failed", "timeout")
@dataclass
class Session:
"""Everything one live invocation shares."""
app: SpaceApp
generate: Endpoint
lora: LoraRoute | None
ledger: Ledger
out: Path
log: Any
invocation: str
ceiling: float
floor_seconds: int | None
multiplier: float
call_timeout: float
read_timeout: float
sampling_rate: int
space_sha: str | None
produced: dict = field(default_factory=dict)
class CeilingReached(RuntimeError):
pass
def _find_record(data: list, status: str) -> dict | None:
"""A structured run record, if the API carries one: an output beyond the three published
ones, or JSON embedded in the status. The report's prose is never read for facts."""
for extra in data[3:]:
if isinstance(extra, dict):
return extra
found = re.search(r'<script[^>]*type="application/json"[^>]*>(.*?)</script>', status or "", re.S)
if found is None:
return None
try:
record = json.loads(html.unescape(found.group(1)))
except ValueError:
return None
return record if isinstance(record, dict) else None
def _record_value(record: dict | None, *names) -> Any:
if not record:
return None
for name in names:
value = record.get(name)
if value is None and isinstance(record.get("failure"), dict):
value = record["failure"].get(name)
if value is not None:
return value
return None
class _Reported(Exception):
"""What reached the API, shaped for ``errors.categorise``: the title (when one came) and the
message. The exception's own type does not cross the API."""
def __init__(self, title: str, message: str):
super().__init__(message)
self.title = title
self.message = message
def classify(record: dict | None, *, title: Any = None, error: Any = None, status: str = "") -> dict:
"""A failed call's category, as it came: the run's structured record's when the API carries
one, else the platform's own refusal wording, else ``errors.categorise`` (the app's
classifier) on the text that reached the API. An unknown string is kept as it is."""
value = _record_value(record, "category")
if value:
return {"category": str(value), "source": "the run's structured record"}
found = platform_refusal(title, error, status)
if found:
return {"category": found[0], "source": "the platform's refusal wording", "platform_words": found[1]}
from ltx25_studio import errors
words = " ".join(str(text) for text in (error, _plain(status)) if text)
failure = errors.categorise(_Reported(str(title or ""), words))
return {"category": str(failure.category), "gpu_used": bool(failure.gpu_used),
"source": "errors.categorise on the text that reached the API"}
def _refusal_category(exc: BaseException) -> str:
"""The category of a refusal raised here, before anything was sent: the app's own."""
if isinstance(exc, PlanError):
return "input"
from ltx25_studio import errors
return str(errors.categorise(exc).category)
def _plain(markdown: str) -> str:
text = re.sub(r"<script.*?</script>", "", markdown or "", flags=re.S)
text = re.sub(r"<[^>]+>", " ", text)
return re.sub(r"[ \t]+", " ", html.unescape(text)).strip()
def _file_path(value) -> str | None:
if isinstance(value, str):
return value or None
if isinstance(value, dict):
if isinstance(value.get("path"), str):
return value["path"]
for key in ("video", "audio", "file"):
inner = _file_path(value.get(key))
if inner:
return inner
return None
def _run_events(ctx: Session, evidence: Evidence, endpoint: Endpoint, data: list, session_hash: str,
started: float, deadline: float, timing: dict, key: str) -> dict:
"""Join one event and read its stream to completion. Records every message."""
evidence.event({"t": round(_monotonic() - started, 4), "endpoint": endpoint.name, "client": "join"})
event_id = ctx.app.join(endpoint.fn_index, data, session_hash)
timing[f"{key}_event_id_t"] = round(_monotonic() - started, 4)
evidence.event({"t": timing[f"{key}_event_id_t"], "endpoint": endpoint.name, "client": "event_id",
"event_id": event_id})
try:
with ctx.app.stream(session_hash, read_timeout=min(ctx.read_timeout, ctx.call_timeout)) as response:
for message in iter_sse(response):
now = _monotonic()
t = round(now - started, 4)
evidence.event({"t": t, "utc": _utcnow().isoformat(), "endpoint": endpoint.name, **message})
kind = message.get("msg") if isinstance(message, dict) else None
if kind == "process_starts":
timing.setdefault(f"{key}_process_starts_t", t)
elif kind == "progress":
timing.setdefault(f"{key}_first_progress_t", t)
if key == "generate":
for unit in message.get("progress_data") or []:
ctx.log(f" progress {unit.get('desc') or ''} {unit.get('index')}/{unit.get('length')}"
f" {unit.get('unit') or ''}".rstrip())
elif kind == "log" and key == "generate":
ctx.log(f" space {message.get('title') or ''}: {message.get('log')}")
elif kind in ("process_completed", "unexpected_error") and message.get("event_id") in (event_id, None):
timing[f"{key}_completed_t"] = t
return message
elif kind == "close_stream":
break
if now > deadline:
raise CallTimeout(f"no completion within {ctx.call_timeout:g} s")
except TimeoutError:
raise CallTimeout(f"the event stream went silent for {ctx.read_timeout:g} s") from None
except TransportError as exc:
if "timed out" in str(exc):
raise CallTimeout(f"the event stream went silent for {ctx.read_timeout:g} s") from None
raise
raise TransportError("the event stream closed before the call completed")
def _validate_loras(ctx: Session, call: Call, evidence: Evidence, session_hash: str, started: float,
timing: dict) -> list[dict]:
route = ctx.lora
verdicts = []
for index, lora in enumerate(call.loras):
values: list = [None] * len(route.endpoint.parameters)
if lora.local is not None:
values[route.upload_slot] = ctx.app.upload(lora.local)
if route.source_slot is not None:
values[route.source_slot] = ""
else:
values[route.source_slot] = lora.source
if route.weight_slot is not None:
values[route.weight_slot] = lora.weight
data = build_data(route.endpoint.input_types, values)
deadline = _monotonic() + ctx.call_timeout
message = _run_events(ctx, evidence, route.endpoint, data, session_hash, started, deadline, timing,
f"lora{index}")
output = message.get("output") or {}
texts = [str(v) for v in output.get("data") or [] if isinstance(v, str)]
verdict = texts[0] if texts else str(output.get("error") or message.get("message") or "")
staged_list = texts[1:]
name = Path(lora.label).name
staged = any(name in text for text in staged_list) if staged_list else None
if message.get("success") is False or message.get("msg") == "unexpected_error":
staged = False
entry = {**lora.describe(), "verdict": verdict, "staged_list": staged_list, "staged": staged}
if staged and lora.needs_weight_route and route.weight_slot is None and route.weight_endpoint is not None:
weights = {"index": index, "weight": lora.weight}
names = [str(p.get("parameter_name") or "").lower() for p in route.weight_endpoint.parameters]
data = build_data(route.weight_endpoint.input_types, [weights.get(n) for n in names])
done = _run_events(ctx, evidence, route.weight_endpoint, data, session_hash, started,
_monotonic() + ctx.call_timeout, timing, f"lora{index}_weight")
entry["weight_set"] = done.get("success", True) is not False
verdicts.append(entry)
ctx.log(f" lora {lora.label}: {'staged' if staged else 'NOT staged' if staged is False else 'unconfirmed'}"
f" · {_plain(verdict)[:160]}")
if staged is False:
break
return verdicts
def _resolve_media(ctx: Session, call: Call, evidence: Evidence) -> dict[str, dict]:
resolved: dict[str, dict] = {}
for kind, item in call.media.items():
if not item.chained:
resolved[kind] = {"local": item.path, "chained": False, "sha256": _sha256(item.path)}
continue
artifacts = ctx.produced.get(item.from_call)
track = "audio" if kind == "audio" else "video"
source = (artifacts or {}).get(track)
if source is None:
raise RuntimeError(f"{kind} is chained from {item.from_call!r}, which produced no {track}")
inputs = evidence.dir / "inputs"
if kind == "image":
dest = inputs / "image.png"
index = extract_image(Path(source), dest, item.frame)
elif kind == "video":
dest = inputs / f"video{Path(source).suffix or '.mp4'}"
dest.parent.mkdir(parents=True, exist_ok=True)
shutil.copyfile(source, dest)
index = None
else:
dest = extract_audio(Path(source), inputs / "audio.wav")
index = None
resolved[kind] = {"local": dest, "chained": True, "from_call": item.from_call, "source_artifact": str(source),
"frame_index": index, "sha256": _sha256(dest)}
return resolved
def run_call(ctx: Session, call: Call) -> Outcome:
day = utc_day()
reserved = reservation_seconds(call, floor_seconds=ctx.floor_seconds)
worst = reserved * ctx.multiplier
spent = ctx.ledger.spent(day)
if spent + worst > ctx.ceiling:
rollover = next_utc_midnight()
left = max(0.0, ctx.ceiling - spent)
raise CeilingReached(
f"stopped before {call.label}: today (UTC {day}) has <= {spent:.1f} GPU-s on the ledger, and "
f"this call reserves {reserved} s x {ctx.multiplier:g} = {worst:.1f}, which would pass the "
f"{ctx.ceiling:.0f} GPU-s ceiling. {left:.1f} GPU-s are left today; the UTC day rolls over at "
f"{rollover.isoformat().replace('+00:00', 'Z')} (in {_duration(rollover - _utcnow())}). Nothing was sent."
)
evidence = Evidence(ctx.out / call.label)
started = _monotonic()
timing: dict = {"reserved_seconds": reserved, "multiplier": ctx.multiplier, "worst_case_billed": worst}
request = {
"label": call.label,
"mode": call.mode,
"note": call.note,
"fields": resolved_request(call),
"loras": [lora.describe() for lora in call.loras],
"expect": call.expect,
"reservation_seconds": reserved,
"space_sha": ctx.space_sha,
}
evidence.write("request.json", request)
ctx.log(f"\n>>> {call.label} {call.mode} reserve {reserved} s (<= {worst:.0f} GPU-s worst case)")
session_hash = uuid.uuid4().hex
outcome = Outcome(label=call.label, status="failed", reserved=reserved)
message: dict | None = None
call_id = uuid.uuid4().hex[:16]
ledger_record = {"kind": "call", "call_id": call_id, "utc_day": day, "invocation": ctx.invocation,
"label": call.label, "mode": call.mode, "reserved_seconds": reserved,
"multiplier": ctx.multiplier, "billed_conservative": 0.0, "billed_worker": None,
"space_sha": ctx.space_sha}
sent = False
def settle(status: str, billed: float, **extra) -> None:
"""The call's one closing ledger record."""
outcome.billed_conservative = billed
ctx.ledger.append({**ledger_record, "at": _utcnow().isoformat(), "outcome": status,
"billed_conservative": round(billed, 3), **extra})
def unsent(detail: str, category: str, **extra) -> Outcome:
"""Refused here, before the run was sent: nothing billed."""
outcome.detail, outcome.category = detail, category
settle("failed", 0.0, error=detail, category=category)
evidence.write("result.json", {"outcome": "failed", "sent": False, "detail": detail,
"category": category, **extra})
return outcome
def no_completion() -> float:
"""No completion was seen: the window so far, never less than the reservation."""
start_t = timing.get("generate_process_starts_t", timing.get("generate_event_id_t", 0.0))
return max(_monotonic() - started - start_t, reserved) * ctx.multiplier
try:
try:
media = _resolve_media(ctx, call, evidence)
except Exception as exc: # noqa: BLE001
request["media"] = {"error": redact(str(exc))}
evidence.write("request.json", request)
return unsent(f"input could not be prepared: {redact(exc)}", "input")
request["media"] = {k: {kk: str(vv) if isinstance(vv, Path) else vv for kk, vv in v.items()}
for k, v in media.items()}
try:
# The app's own gate on the real files, as the Space runs it: what a dry run could
# only defer for a chained input is checked here, before anything is sent.
call = gated(call, {k: v["local"] for k, v in media.items()})
except Exception as exc: # noqa: BLE001 -- refused before dispatch: nothing billed
evidence.write("request.json", request)
return unsent(f"the app refuses this call before reserving a GPU: {redact(exc)}",
_refusal_category(exc))
request.update(app_notes=list(call.app_notes), app_frames=call.app_frames, app_level=call.app_level,
app_changes=[{"field": k, "sent": s, "runs": r} for k, s, r in call.app_changes])
evidence.write("request.json", request)
if call.loras:
verdicts = _validate_loras(ctx, call, evidence, session_hash, started, timing)
request["loras"] = verdicts
evidence.write("request.json", request)
rejected = [v for v in verdicts if v["staged"] is False]
if rejected:
return unsent(f"LoRA not staged: {_plain(rejected[0]['verdict'])[:400]}", "adapter", loras=verdicts)
uploaded = {kind: ctx.app.upload(item["local"]) for kind, item in media.items()}
timing["uploaded_t"] = round(_monotonic() - started, 4)
data = build_data(ctx.generate.input_types, generate_values(call, uploaded))
deadline = _monotonic() + ctx.call_timeout
# Written down at its full reservation BEFORE it is sent, settled below: a call this
# process never comes back from still counts against the day.
ctx.ledger.append({**ledger_record, "kind": "call_start", "at": _utcnow().isoformat(),
"billed_conservative": round(worst, 3)})
sent = True
evidence.write("result.json", {"outcome": "sent", "sent": True,
"note": "sent to the Space; replaced when the call completes"})
try:
message = _run_events(ctx, evidence, ctx.generate, data, session_hash, started, deadline, timing,
"generate")
except (CallTimeout, TransportError) as exc:
billed = no_completion()
outcome.status = "timeout" if isinstance(exc, CallTimeout) else "failed"
outcome.detail = redact(str(exc))
outcome.category = "timeout" if isinstance(exc, CallTimeout) else "transport"
timing.update(billed_conservative=round(billed, 3),
billed_rule="max(window, reservation) x multiplier: no completion")
settle(outcome.status, billed, error=outcome.detail, category=outcome.category)
evidence.write("timing.json", timing)
evidence.write("result.json", {"outcome": outcome.status, "sent": True, "detail": outcome.detail,
"category": outcome.category})
return outcome
start_t = timing.get("generate_process_starts_t", timing.get("generate_event_id_t", 0.0))
window = max(0.0, timing["generate_completed_t"] - start_t)
output = message.get("output") or {}
data_out = list(output.get("data") or [])
status = str(data_out[2]) if len(data_out) > 2 and data_out[2] is not None else ""
record = _find_record(data_out, status)
# The worker's own GPU time only from a structured record; the report's prose is not read.
worker = _record_value(record, "elapsed_seconds")
billed_worker = None if worker is None else float(worker) * ctx.multiplier
error = None
if message.get("success") is False or message.get("msg") == "unexpected_error":
error = output.get("error") or message.get("message") or (
"the Space raised an error and did not say what (its show_error is off)"
)
missing = [slot for index, slot in enumerate(OUTPUT_SLOTS)
if slot in returned_files(call.spec) and not (index < len(data_out) and _file_path(data_out[index]))]
failure = classify(record, title=output.get("title"), error=error, status=status) if error or missing \
else None
category = failure["category"] if failure else None
billed = 0.0 if category in UNBILLED_CATEGORIES else window * ctx.multiplier
timing.update(window_seconds=round(window, 4), billed_conservative=round(billed, 3),
billed_worker=None if billed_worker is None else round(billed_worker, 3),
worker_elapsed_seconds=worker,
billed_rule=f"zero: the platform refused ({category}) before a GPU was used"
if category in UNBILLED_CATEGORIES else "(completed - process_starts) x multiplier")
outcome.category = category
words = (failure or {}).get("platform_words")
closing = ("quota_refusal" if category == "quota" else "platform_refusal" if category in UNBILLED_CATEGORIES
else "failed" if failure else "completed")
settle(closing, billed,
billed_worker=None if billed_worker is None else round(billed_worker, 3),
error=redact(words or error or (_plain(status)[:600] if failure else "")) or None,
category=category)
evidence.write("timing.json", timing)
result = {"success": message.get("success"), "status_markdown": status, "status_text": _plain(status),
"error": error, "title": output.get("title"), "record": record, "category": category,
"classified": failure, "outputs": data_out}
if category == "quota":
outcome.status = "quota"
outcome.detail = redact(error or _plain(status) or words)
evidence.write("result.json", {**result, "outcome": "quota"})
return outcome
artifacts: dict[str, Path] = {}
downloaded: dict[str, Path] = {}
for index, track in enumerate(OUTPUT_SLOTS):
server_path = _file_path(data_out[index]) if index < len(data_out) else None
if not server_path:
continue
if server_path not in downloaded:
downloaded[server_path] = ctx.app.download(server_path, evidence.dir / f"{track}_{Path(server_path).name}")
artifacts[track] = downloaded[server_path]
timing["downloaded_t"] = round(_monotonic() - started, 4)
evidence.write("timing.json", timing)
result["artifacts"] = {k: {"path": str(v), "bytes": v.stat().st_size, "sha256": _sha256(v)} for k, v in artifacts.items()}
if failure:
tracks = " and ".join(missing)
outcome.detail = redact(error or f"no {tracks} file; the Space said: {_plain(status)[:600]}")
outcome.status = "failed"
evidence.write("result.json", {**result, "outcome": "failed", "detail": outcome.detail})
return outcome
verify = verify_call(call, artifacts, sampling_rate=ctx.sampling_rate)
evidence.write("verify.json", verify)
outcome.artifacts = artifacts
outcome.flags = _flags(verify)
outcome.audio = _audio_line((verify["content"] or {}).get("audio") or {})
outcome.status = "ok" if verify["ok"] else "mismatch"
outcome.detail = "; ".join(verify["problems"])
evidence.write("result.json", {**result, "outcome": outcome.status, "detail": outcome.detail})
ctx.produced[call.label] = {k: str(v) for k, v in artifacts.items()}
return outcome
except KeyboardInterrupt:
# Recorded, then the invocation stops. A sent call is charged as one that never completed.
if sent and message is None:
settle("interrupted", no_completion(), error="interrupted while the Space was running it")
evidence.write("result.json", {"outcome": "interrupted", "sent": sent,
"completed": message is not None})
raise
except Exception as exc: # noqa: BLE001 -- recorded, then the plan stops (or keeps going)
outcome.status = "failed"
outcome.category = outcome.category or "harness"
outcome.detail = redact(f"{type(exc).__name__}: {exc}")
if sent and message is None:
settle("failed", no_completion(), error=outcome.detail, category=outcome.category)
elif not sent:
settle("failed", 0.0, error=outcome.detail, category=outcome.category)
evidence.write("result.json", {"outcome": "failed", "sent": sent, "detail": outcome.detail,
"traceback": redact("".join(traceback.format_exception(exc)))})
return outcome
finally:
evidence.close()
def _duration(delta: timedelta) -> str:
seconds = max(0, int(delta.total_seconds()))
return f"{seconds // 3600}h {(seconds % 3600) // 60:02d}m"
# ---------------------------------------------------------------------------
# reporting
# ---------------------------------------------------------------------------
def schedule_days(costs: list[float], *, spent: float, ceiling: float) -> list[list[int]]:
"""The calls (by index) each UTC day can run at worst case, every call billed its full
reservation: today first, from what the ledger already holds, then whole days from zero."""
days: list[list[int]] = [[]]
running = spent
for index, cost in enumerate(costs):
if running + cost > ceiling and (days[-1] or running > 0):
days.append([])
running = 0.0
days[-1].append(index)
running += cost
return days
def _span(indexes: list[int], calls: list[Call], offset: int) -> str:
first, last = indexes[0], indexes[-1]
numbers = f"#{first + 1 + offset}" if first == last else f"#{first + 1 + offset}-{last + 1 + offset}"
labels = calls[first].label if first == last else f"{calls[first].label} .. {calls[last].label}"
return f"calls {numbers:<7} {labels}"
def _levels_line(call: Call) -> str | None:
if call.app_level is None:
return None
from ltx25_studio import audio_conditioning
name = "VIDEO_VARIATION_LEVELS" if _varies_clip(call.spec) else "VARIATION_LEVELS"
levels = getattr(audio_conditioning, name, ())
shown = f"; the mode's levels: {', '.join(f'{v:g}' for v in levels)}" if levels else ""
return f" level {call.app_level:g} (noise_scale {call.value('noise_scale')!r}{shown})"
def _gate_line(call: Call) -> str | None:
if call.gate == GATE_ACCEPTED:
return " app gate generate.prepare_request accepts it, run as on the Space (ZeroGPU, CUDA)"
if call.gate == GATE_DEFERRED:
sources = ", ".join(f"{kind} from {call.media[kind].from_call}" for kind in call.chained)
return (f" app gate checked when it exists: its {sources} is made by that call, and "
f"generate.prepare_request runs on the file before anything is sent (a refusal bills nothing)")
return None
def render_plan(calls: list[Call], *, floor_seconds: int | None, multiplier: float, spent: float,
ceiling: float, live: bool, ledger_note: str, resume_command: str = "", skipped=(),
offset: int = 0) -> list[str]:
floor = (f", at least {floor_seconds} s (the Space's LTX25_GPU_DURATION_FLOOR_SECONDS)"
if floor_seconds else "")
lines = [
"=== PLAN — LIVE: this WILL spend GPU-seconds ===" if live
else "=== PLAN — DRY RUN: nothing is contacted and nothing is spent ===",
f" reservation from each call's size and length (the app's estimator){floor} · "
f"multiplier x{multiplier:g} · ceiling {ceiling:.0f} GPU-s/UTC day",
f" {ledger_note}",
]
if skipped:
lines.append(f" resume {len(skipped)} call(s) already have an outcome in --out and are skipped, "
f"never re-run: " + ", ".join(f"{label} ({state})" for label, state in skipped))
costs: list[float] = []
running = spent
for number, call in enumerate(calls, start=1 + offset):
reserved = reservation_seconds(call, floor_seconds=floor_seconds)
worst = reserved * multiplier
costs.append(worst)
running += worst
parts = _reservation_parts(call, floor_seconds=floor_seconds)
lines.append("")
lines.append(f" {number:>2}. {call.label} [{call.mode}]")
lines.append(f" request {json.dumps(resolved_request(call), sort_keys=False, default=str)}")
for kind, item in call.media.items():
if item.chained:
where = f"chained from {item.from_call}" + (f" (frame {item.frame})" if kind == "image" else "")
else:
where = str(item.path)
lines.append(f" {kind:<9} {where}")
for lora in call.loras:
line = f" lora {lora.source} at {lora.weight:g} via {'upload' if lora.local else 'Hub id'}"
if lora.needs_weight_route:
line += (f": a strength other than {staged_lora_weight():g} needs a public route that sets it "
f"({LORA_WEIGHT_ENDPOINT} with index and weight). --live checks the Space for one "
f"before anything is spent and refuses the plan without it (exit 2); "
f"--skip {call.label} runs the rest")
lines.append(line)
gate = _gate_line(call)
if gate:
lines.append(gate)
shown = [f"{key} {ran!r} (sent {sent!r})" for key, sent, ran in call.app_changes
if key in call.values or key in ("num_frames", "auto_duration")]
if shown:
lines.append(f" app runs {' · '.join(shown)}")
level = _levels_line(call)
if level:
lines.append(level)
for note in call.app_notes:
lines.append(f" app {note}")
geometry = call.expected_geometry()
expect = {"geometry": f"{geometry[0]}x{geometry[1]}" if geometry else None,
"frames": call.expected_frames(),
**{k: v for k, v in call.expect.items() if k not in ("width", "height", "frames")}}
lines.append(f" expect {json.dumps({k: v for k, v in expect.items() if v is not None})}")
lines.append(f" reserve {reserved} s ({', '.join(parts)}) -> worst case <= {worst:.0f} GPU-s; "
f"day total <= {running:.0f}")
if worst > ceiling:
lines.append(f" NEVER RUNS under this ceiling: its worst case alone passes {ceiling:.0f} GPU-s")
plan_total = running - spent
days = schedule_days(costs, spent=spent, ceiling=ceiling)
today = _utcnow()
lines += [
"",
f" this plan <= {plan_total:.0f} GPU-s worst case (every call at its full reservation)",
f" today <= {spent:.1f} GPU-s already on the ledger (UTC {utc_day(today)})",
f" remaining {max(0.0, ceiling - spent):.1f} GPU-s before the {ceiling:.0f} ceiling",
f" days {len(days)} UTC day(s) at worst case under the {ceiling:.0f} GPU-s ceiling:",
]
for index, members in enumerate(days):
when = utc_day(today + timedelta(days=index)) + (" (today)" if index == 0 else "")
if not members:
lines.append(f" day {index + 1} UTC {when:<18} nothing fits in what is left today")
continue
total = sum(costs[i] for i in members)
lines.append(f" day {index + 1} UTC {when:<18} <= {total:>4.0f} GPU-s "
f"{_span(members, calls, offset)}")
if len(days) == 1:
lines.append(f" verdict FITS, with {ceiling - running:.0f} GPU-s to spare at worst case")
else:
stop = days[1][0]
lines.append(
f" verdict DOES NOT FIT: call #{stop + 1 + offset} ({calls[stop].label}) is where the ceiling "
f"binds at worst case. A live run checks the ledger before every call and refuses, with exit 3 "
f"and nothing sent, the first call that could pass the ceiling; a call bills its measured window, "
f"usually far less than its reservation, so a live day usually runs further than this. The UTC day "
f"rolls over at {next_utc_midnight(today).isoformat().replace('+00:00', 'Z')}."
)
if resume_command:
lines += [
" resume on the next UTC day, run the same plan with --resume: every call that has an "
"outcome in --out is skipped by label, never re-run, and the rest runs:",
f" {resume_command}",
]
lines.append(" Every figure is an UPPER BOUND, not measured billing.")
return lines
# ---------------------------------------------------------------------------
# CLI
# ---------------------------------------------------------------------------
_FIELD_FLAGS = (
("--prompt", "prompt", str),
("--width", "width", int),
("--height", "height", int),
("--num-frames", "num_frames", int),
("--seed", "seed", int),
("--min-seconds", "min_seconds", float),
("--max-seconds", "max_seconds", float),
("--strength", "strength", float),
("--noise-scale", "noise_scale", float),
)
_BOOL_FLAGS = (
("--enable-stage-2", "enable_stage_2"),
("--enhance-prompt", "enhance_prompt"),
("--auto-duration", "auto_duration"),
)
SUBCOMMANDS = ("status", "info", "logs")
_EPILOG = f"""
read-only subcommands (one GET each, no GPU time; see `<command> --help`):
status the Space's runtime stage, hardware and deployed commit
info compare the Space's /gradio_api/info for {ENDPOINT} with the local field spec
logs --out FILE save the Space's runtime (or build) log
plan file: a JSON list of calls, or {{"defaults": {{...}}, "calls": [...]}}. A call has a "label",
a "mode", any {ENDPOINT} field by its key (a choice by its value or its label, e.g. "decoder":
"Sharper (standard decoder)"), input media as file paths under --assets-dir ("image", "video", "audio") or
{{"from_call": "<earlier label>", "frame": 0}}, optional "loras" ([{{"source": "<hub id>" |
"<file.safetensors>", "weight": 0.8}}]) and optional "expect" ({", ".join(sorted(_EXPECT_KEYS))}).
Every call whose inputs exist goes through the app's own gate (generate.prepare_request) in the
dry run, as the Space runs it; a call with an input made by an earlier call is gated when that
file exists, before anything is sent.
a plan longer than one day: run it with --out DIR; when it stops at the ceiling (exit 3), run the
same command with --resume on the next UTC day. Calls with an outcome in DIR are skipped by
label, never re-run; a failed one is re-run only if you move its folder out of DIR.
exit codes: 0 plan complete · 2 illegal plan or contract mismatch (nothing spent) · 3 stopped at
the ceiling · 4 platform quota refusal · 5 a call failed (still 5 with --keep-going)
"""
def _run_parser() -> argparse.ArgumentParser:
parser = argparse.ArgumentParser(
prog="space_matrix.py",
description="Run a plan of calls against the private LTX-2.5 Studio Space inside a hard daily "
"GPU budget. Dry run by default; --live is what spends.",
epilog=_EPILOG,
formatter_class=argparse.RawDescriptionHelpFormatter,
)
plan = parser.add_argument_group("what to run")
plan.add_argument("--plan", help="a plan JSON file, or 'default' for the historical 11-mode loop "
"(the default when no --mode is given)")
plan.add_argument("--modes", help="with --plan default: comma-separated mode ids, run in that order")
plan.add_argument("--only", help="comma-separated call labels: run only these calls of the plan")
plan.add_argument("--skip", help="comma-separated call labels: leave these calls of the plan out")
plan.add_argument("--assets-dir", help="where relative media and LoRA paths resolve (default: the plan "
"file's folder)")
plan.add_argument("--mode", help="a single call from flags: the mode id")
plan.add_argument("--label", help="with --mode: the call's label (default: the mode id)")
for flag, key, kind in _FIELD_FLAGS:
plan.add_argument(flag, type=kind, default=None, help=f"the {key} field (--mode, or --plan default)")
for flag, key in _BOOL_FLAGS:
plan.add_argument(flag, action="store_true", default=None, help=f"set {key} (--mode, or --plan default)")
plan.add_argument("--image", help="with --mode: the input image file")
plan.add_argument("--video", help="with --mode: the input video file")
plan.add_argument("--audio", help="with --mode: the input audio file")
plan.add_argument("--set", action="append", default=[], metavar="KEY=VALUE",
help="with --mode: any field by key (JSON value, else a string); repeatable")
plan.add_argument("--lora", action="append", default=[], metavar="SOURCE[@WEIGHT]",
help="with --mode: stage a LoRA (Hub id or local .safetensors); repeatable")
plan.add_argument("--expect", action="append", default=[], metavar="KEY=VALUE",
help=f"with --mode: an expectation ({', '.join(sorted(_EXPECT_KEYS))}); repeatable")
run = parser.add_argument_group("running and budget")
run.add_argument("--live", action="store_true", help="actually call the Space. Without it nothing is "
"contacted and nothing is spent.")
run.add_argument("--dry-run", action="store_true", help="force the dry run even with --live")
run.add_argument("--quota-ledger", help="the append-only spend ledger (one JSON record per line, keyed by "
"UTC day). Required with --live; read, never written, by a dry run.")
run.add_argument("--quota-ceiling", type=float, default=DEFAULT_QUOTA_CEILING_SECONDS,
help=f"GPU-seconds per UTC day never to pass (default {DEFAULT_QUOTA_CEILING_SECONDS:.0f})")
run.add_argument("--reservation-floor-seconds", type=int, default=None,
help="the Space's LTX25_GPU_DURATION_FLOOR_SECONDS, if it sets one: every call's "
"reservation is at least this (default: unset, as on the Space). The reservation "
"itself is the app's estimate from each call's size and length")
run.add_argument("--zerogpu-size", default=DEFAULT_ZEROGPU_SIZE,
help=f"the Space's LTX25_ZEROGPU_SIZE; zerogpu.quota_multiplier turns it into the quota "
f"factor (default {DEFAULT_ZEROGPU_SIZE}, which counts double)")
run.add_argument("--out", help="evidence directory, one folder per call label. Required with --live.")
run.add_argument("--resume", action="store_true", help="continue a plan in its --out: every call with an "
"outcome there is skipped by label and never re-run (a dry run shows what is left)")
run.add_argument("--keep-going", action="store_true", help="after a failed call, run the rest of the plan "
"(a quota refusal or the ceiling still stops at once)")
run.add_argument("--call-timeout", type=float, default=DEFAULT_CALL_TIMEOUT_SECONDS,
help=f"seconds to wait for one call to complete (default {DEFAULT_CALL_TIMEOUT_SECONDS:.0f}); "
"a timed-out call is recorded, charged at least its reservation, and never retried")
run.add_argument("--read-timeout", type=float, default=DEFAULT_READ_TIMEOUT_SECONDS,
help=f"seconds of silence on the event stream before giving up (default "
f"{DEFAULT_READ_TIMEOUT_SECONDS:.0f})")
run.add_argument("--sampling-rate", type=int, default=PINNED_VOCODER_SAMPLING_RATE,
help=f"expected audio sample rate (default {PINNED_VOCODER_SAMPLING_RATE}); "
"expect.sample_rate overrides it per call")
where = parser.add_argument_group("where")
_where_flags(where)
return parser
def _where_flags(group) -> None:
group.add_argument("--space", default=DEFAULT_SPACE_ID, help=f"the Space id (default {DEFAULT_SPACE_ID})")
group.add_argument("--space-url", help="the Space app's base URL (default: https://<owner>-<name>.hf.space)")
group.add_argument("--hub-url", default=DEFAULT_HUB_URL, help=f"the Hub API base (default {DEFAULT_HUB_URL})")
def _sub_parser(name: str) -> argparse.ArgumentParser:
descriptions = {
"status": "Print the Space's runtime stage, hardware and deployed commit. One GET to the Hub API.",
"info": f"Compare the Space's /gradio_api/info for {ENDPOINT} with the local field spec. One GET to "
"the Space. Exits 2 on a mismatch.",
"logs": "Save the Space's runtime (or build) log to a file. One GET to the Hub API, read for at most "
"--max-seconds.",
}
parser = argparse.ArgumentParser(prog=f"space_matrix.py {name}", description=descriptions[name])
_where_flags(parser)
if name == "logs":
parser.add_argument("--out", required=True, help="the file to write")
parser.add_argument("--kind", choices=("run", "build"), default="run", help="which log (default run)")
parser.add_argument("--max-seconds", type=float, default=20.0,
help="stop reading the (endless) log stream after this long (default 20)")
return parser
def _parse_kv(items: list[str], what: str) -> dict:
out = {}
for item in items:
key, sep, raw = item.partition("=")
if not sep or not key:
raise PlanError(f"{what} {item!r} must be KEY=VALUE")
try:
out[key] = json.loads(raw)
except ValueError:
out[key] = raw
return out
def _flag_values(args) -> dict:
values = {key: getattr(args, flag[2:].replace("-", "_")) for flag, key, _ in _FIELD_FLAGS}
values.update({key: getattr(args, flag[2:].replace("-", "_")) for flag, key in _BOOL_FLAGS})
return {k: v for k, v in values.items() if v is not None}
def _labels(text: str | None) -> list[str]:
return [label.strip() for label in (text or "").split(",") if label.strip()]
def build_calls(args) -> list[Call]:
flags = _flag_values(args)
single = args.mode is not None
singles = [args.image, args.video, args.audio, args.label, *args.set, *args.lora, *args.expect]
if not single and any(singles):
raise PlanError("--image/--video/--audio/--label/--set/--lora/--expect describe a single call: add --mode")
if single and (args.plan or args.modes):
raise PlanError("--mode describes one call; it cannot be combined with --plan or --modes")
if single:
entry: dict = {"label": args.label or args.mode, "mode": args.mode, **flags, **_parse_kv(args.set, "--set")}
for kind in MEDIA_KINDS:
if getattr(args, kind):
entry[kind] = getattr(args, kind)
loras = []
for item in args.lora:
source, sep, weight = item.rpartition("@")
loras.append({"source": source, "weight": float(weight)} if sep and _is_number(weight)
else {"source": item})
if loras:
entry["loras"] = loras
if args.expect:
entry["expect"] = _parse_kv(args.expect, "--expect")
base = Path(args.assets_dir).expanduser() if args.assets_dir else Path.cwd()
calls = parse_plan([entry], base=base)
elif args.plan in (None, "default"):
mode_ids = [m.strip() for m in args.modes.split(",") if m.strip()] if args.modes else None
calls = default_plan(mode_ids, flags)
else:
if args.modes:
raise PlanError("--modes selects from --plan default; use --only with a plan file")
if flags:
raise PlanError(f"{', '.join('--' + k.replace('_', '-') for k in flags)}: a plan file sets its own "
"fields; edit the file")
calls = load_plan(args.plan, assets_dir=Path(args.assets_dir).expanduser() if args.assets_dir else None)
known = {c.label for c in calls}
for flag, wanted in (("--only", _labels(args.only)), ("--skip", _labels(args.skip))):
unknown = [label for label in wanted if label not in known]
if unknown:
raise PlanError(f"{flag}: no call labelled {unknown} (labels: {sorted(known)})")
if args.only:
calls = [c for c in calls if c.label in _labels(args.only)]
if args.skip:
calls = [c for c in calls if c.label not in _labels(args.skip)]
chosen = {c.label for c in calls}
for call in calls:
for item in call.media.values():
if item.chained and item.from_call not in chosen:
raise PlanError(f"{call.label} chains from {item.from_call}, which --only/--skip leaves out")
return calls
def _is_number(text: str) -> bool:
try:
float(text)
except ValueError:
return False
return True
# ---------------------------------------------------------------------------
# resuming a plan in its --out
# ---------------------------------------------------------------------------
def evidence_state(folder: Path) -> dict | None:
"""What an earlier invocation left in ``--out/<label>/``; None when it holds nothing."""
if not folder.is_dir() or not any(folder.iterdir()):
return None
try:
result = json.loads((folder / "result.json").read_text())
except FileNotFoundError:
# A call writes result.json before it is sent: nothing of this one reached the Space.
return {"outcome": "unsent", "sent": False}
except (OSError, ValueError):
return {"outcome": "unreadable", "sent": True}
if not isinstance(result, dict):
return {"outcome": "unreadable", "sent": True}
artifacts = {track: info.get("path") for track, info in (result.get("artifacts") or {}).items()
if isinstance(info, dict) and info.get("path")}
return {"outcome": str(result.get("outcome") or "unknown"), "sent": result.get("sent", True) is not False,
"artifacts": artifacts}
def _runs_again(state: dict | None) -> bool:
"""Only a call nothing of which reached the Space runs on --resume; nothing is retried."""
return state is None or state["outcome"] == "unsent" or (state["outcome"] == "interrupted" and not state["sent"])
def resume_split(out_dir: Path, calls: list[Call]) -> tuple[list[Call], list[tuple[str, str]], dict]:
"""``(calls still to run, (label, outcome) of those skipped, artifacts of the finished ones)``."""
remaining, skipped, produced = [], [], {}
for call in calls:
state = evidence_state(out_dir / call.label)
if _runs_again(state):
remaining.append(call)
continue
skipped.append((call.label, state["outcome"]))
if state["outcome"] in ("ok", "mismatch") and state.get("artifacts"):
produced[call.label] = state["artifacts"]
return remaining, skipped, produced
def _resume_command(argv: list[str]) -> str:
"""The command that continues this plan on the next UTC day."""
import shlex
words = [w for w in argv if w not in ("--dry-run", "--live", "--resume")]
extra = []
if not any(w == "--out" or w.startswith("--out=") for w in words):
extra += ["--out", "<DIR>"]
if not any(w == "--quota-ledger" or w.startswith("--quota-ledger=") for w in words):
extra += ["--quota-ledger", "<LEDGER>"]
return " ".join(["python scripts/space_matrix.py", shlex.join(words), *extra, "--live", "--resume"]).replace(" ", " ")
class _Out:
"""Every line the CLI prints passes through here and is redacted."""
def __init__(self, log):
self._log = log
def __call__(self, line: Any = "") -> None:
self._log(redact(line))
def main(argv=None, *, log=print) -> int:
out = _Out(log)
argv = list(sys.argv[1:] if argv is None else argv)
if argv and argv[0] in SUBCOMMANDS:
return _subcommand(argv[0], argv[1:], out)
args = _run_parser().parse_args(argv)
try:
calls = build_calls(args)
except PlanError as exc:
out(f"the plan is not runnable:\n{exc}")
return EXIT_ILLEGAL
live = args.live and not args.dry_run
try:
multiplier = quota_multiplier(args.zerogpu_size)
except Exception as exc: # noqa: BLE001 -- the app's ConfigError names the legal sizes
out(f"--zerogpu-size: {exc}")
return EXIT_ILLEGAL
if args.reservation_floor_seconds is not None and args.reservation_floor_seconds < 1:
out("--reservation-floor-seconds must be a whole number of seconds >= 1, as on the Space.")
return EXIT_ILLEGAL
ledger = Ledger(args.quota_ledger) if args.quota_ledger else None
if live and ledger is None:
out("--live needs --quota-ledger PATH: the day's spend must be kept outside any checkout.")
return EXIT_ILLEGAL
if live and not args.out:
out("--live needs --out DIR for the per-call evidence.")
return EXIT_ILLEGAL
if args.resume and not args.out:
out("--resume needs --out DIR: the evidence folder of the run it continues.")
return EXIT_ILLEGAL
skipped: list = []
produced: dict = {}
if args.resume:
calls, skipped, produced = resume_split(Path(args.out).expanduser(), calls)
try:
spent = ledger.spent(utc_day()) if ledger else 0.0
previous = ledger.previous_day(utc_day()) if ledger else None
unsettled = ledger.unsettled() if ledger else []
except QuotaLedgerUnreadable as exc:
out(str(exc))
return EXIT_ILLEGAL
if ledger is None:
note = "no --quota-ledger given: this dry run assumes nothing has been spent today"
else:
note = f"ledger {ledger.path}: <= {spent:.1f} GPU-s on UTC {utc_day()}"
if previous:
note += f" (the last earlier day, UTC {previous[0]}, ended at <= {previous[1]:.1f})"
if unsettled:
note += (f"; {len(unsettled)} call(s) sent and never settled ("
f"{', '.join(str(r.get('label')) for r in unsettled)}) count at their full reservation")
if not calls:
out(f" {note}")
out(f"nothing left to run: every call of the plan has an outcome in {args.out}"
if skipped else "nothing to run")
return EXIT_OK
for line in render_plan(calls, floor_seconds=args.reservation_floor_seconds, multiplier=multiplier, spent=spent,
ceiling=args.quota_ceiling, live=live, ledger_note=note,
resume_command=_resume_command(argv), skipped=skipped):
out(line)
if not live:
out("")
out("DRY RUN: every call above was validated offline. Nothing was contacted, ZERO GPU-seconds were "
"spent, and the ledger was read, not written. Add --live to execute.")
return EXIT_OK
return _live(args, calls, ledger, multiplier, out, skipped=skipped, produced=produced,
resume_command=_resume_command(argv))
def _live(args, calls: list[Call], ledger: Ledger, multiplier: float, out, *, skipped=(), produced=None,
resume_command: str = "") -> int:
out_dir = Path(args.out).expanduser()
if not args.resume:
taken = [c.label for c in calls if (out_dir / c.label).exists() and any((out_dir / c.label).iterdir())]
if taken:
out(f"{out_dir} already holds evidence for {taken}; use a fresh --out so runs never mix, or "
f"--resume to run only the calls that have no outcome there.")
return EXIT_ILLEGAL
secret = read_token()
if secret is None:
out(f"no token: set HF_TOKEN or log in so {token_path()} exists. The Space is private.")
return EXIT_ILLEGAL
http = Http(secret)
hub = Hub(args.hub_url, http)
try:
status = hub.space(args.space)
except (HttpError, TransportError) as exc:
out(f"could not read the Space's status: {exc}")
return EXIT_ILLEGAL
runtime = status.get("runtime") or {}
stage = runtime.get("stage")
if stage != "RUNNING":
out(f"the Space is {stage}, not RUNNING. Wake it or wait for the build; nothing was sent.")
return EXIT_ILLEGAL
base_url = args.space_url or status.get("host") or space_host(args.space)
if not base_url.startswith(("http://", "https://")):
base_url = "https://" + base_url
sha = runtime.get("sha") or status.get("sha")
try:
app = SpaceApp(base_url, http)
config = app.config()
app.prefix = "/" + str(config.get("api_prefix") or API_PREFIX).strip("/")
info = app.info()
except (HttpError, TransportError, ValueError) as exc:
out(f"could not read the Space's API: {exc}")
return EXIT_ILLEGAL
generate = find_endpoint(config, info, ENDPOINT)
problems: list[str] = []
table: list[str] = []
if generate is None:
problems.append(f"the Space publishes no {ENDPOINT}")
else:
problems, table = generate_contract_problems(generate.parameters, input_types=generate.input_types,
plan_modes=[c.mode for c in calls])
route = _lora_route(config, info)
problems += lora_contract_problems(route, calls)
out_dir.mkdir(parents=True, exist_ok=True)
(out_dir / "contract.json").write_text(_redacted_json(
{"space": args.space, "space_url": base_url, "space_sha": sha, "problems": problems, "table": table,
"info": info.get("named_endpoints", {}).get(ENDPOINT)}))
if problems:
out(f"\nCONTRACT MISMATCH: the Space at {sha or '?'} does not take what this checkout's field spec builds. "
"Nothing was spent.")
for line in table:
out(line)
for problem in problems:
out(f" ✗ {problem}")
return EXIT_ILLEGAL
out(f"\ncontract {ENDPOINT} matches the field spec ({len(generate.parameters)} parameters) · Space "
f"commit {sha or '?'} · {base_url}")
invocation = uuid.uuid4().hex[:12]
ledger.append({"kind": "invocation", "utc_day": utc_day(), "at": _utcnow().isoformat(), "invocation": invocation,
"space": args.space, "space_sha": sha, "calls": [c.label for c in calls],
"skipped": [label for label, _state in skipped], "out": str(out_dir)})
ctx = Session(app=app, generate=generate, lora=route, ledger=ledger, out=out_dir, log=out,
invocation=invocation, ceiling=args.quota_ceiling, floor_seconds=args.reservation_floor_seconds,
multiplier=multiplier, call_timeout=args.call_timeout, read_timeout=args.read_timeout,
sampling_rate=args.sampling_rate, space_sha=sha, produced=dict(produced or {}))
for label, state in skipped:
out(f" {label:<32} skipped: it has an outcome in {out_dir} ({state})")
outcomes: list[Outcome] = []
code = EXIT_OK
stop = ""
for call in calls:
folder = out_dir / call.label
if folder.exists() and any(folder.iterdir()):
# --resume: an earlier attempt nothing of which reached the Space; kept, moved aside.
aside = folder.with_name(f"{folder.name}.unsent-{invocation}")
folder.rename(aside)
out(f" {call.label}: its earlier attempt was never sent; its folder is now {aside.name}")
try:
outcome = run_call(ctx, call)
except CeilingReached as exc:
stop = str(exc)
code = EXIT_CEILING
out(f"\n*** STOPPED AT THE CEILING ***\n {exc}")
break
except QuotaLedgerUnreadable as exc:
stop = str(exc)
code = EXIT_ILLEGAL
out(f"\n*** STOPPED: the ledger became unreadable ***\n {exc}")
break
outcomes.append(outcome)
out(f" -> {outcome.status.upper()} billed <= {outcome.billed_conservative:.1f} GPU-s (conservative)"
+ (f" [{outcome.category}]" if outcome.category else "")
+ (f" {outcome.detail}" if outcome.detail else "")
+ (f" flags: {', '.join(outcome.flags)}" if outcome.flags else ""))
if outcome.audio:
out(f" audio {outcome.audio} (measurements, no verdict)")
if outcome.status == "quota":
stop = outcome.detail
code = EXIT_QUOTA
out("\n*** THE PLATFORM REFUSED ON QUOTA: stopping now. Nothing is retried. ***")
break
if outcome.failed:
code = EXIT_CALL_FAILED
if not args.keep_going:
stop = f"{call.label} {outcome.status}"
out(" stopping: a call failed (pass --keep-going to run the rest)")
break
day = utc_day()
try:
spent_today = ledger.spent(day)
except QuotaLedgerUnreadable:
spent_today = float("nan")
this_run = sum(o.billed_conservative for o in outcomes)
summary = {
"invocation": invocation,
"space_sha": sha,
"exit_code": code,
"stopped": stop or None,
"skipped": [{"label": label, "outcome": state} for label, state in skipped],
"calls": [{"label": o.label, "status": o.status, "category": o.category, "detail": o.detail,
"reserved_seconds": o.reserved, "billed_conservative": round(o.billed_conservative, 3),
"flags": o.flags, "audio": o.audio or None} for o in outcomes],
"not_run": [c.label for c in calls[len(outcomes):]],
"this_invocation_billed_conservative": round(this_run, 3),
"utc_day": day,
"spent_today_conservative": round(spent_today, 3),
"ceiling": args.quota_ceiling,
}
(out_dir / "summary.json").write_text(_redacted_json(summary))
(out_dir / f"summary-{invocation}.json").write_text(_redacted_json(summary))
out("\n=== SUMMARY ===")
for o in outcomes:
out(f" {o.label:<32} {o.status:<9} <= {o.billed_conservative:7.1f} GPU-s"
+ (f" [{'; '.join(o.flags)}]" if o.flags else ""))
for label in summary["not_run"]:
out(f" {label:<32} not run")
out(f" this invocation <= {this_run:.1f} GPU-s · today (UTC {day}) <= {spent_today:.1f} of {args.quota_ceiling:.0f}"
" · UPPER BOUNDS from wall clock, not measured billing")
out(f" evidence {out_dir} · ledger {ledger.path}")
if summary["not_run"] and code == EXIT_CEILING and resume_command:
out(f" the rest runs after the UTC day rolls over, with:\n {resume_command}")
return code
def _subcommand(name: str, argv: list[str], out) -> int:
args = _sub_parser(name).parse_args(argv)
http = Http(read_token())
try:
if name == "status":
payload = Hub(args.hub_url, http).space(args.space)
runtime = payload.get("runtime") or {}
hardware = runtime.get("hardware") or {}
out(f"space {payload.get('id', args.space)}{' (private)' if payload.get('private') else ''}")
out(f"stage {runtime.get('stage')}")
out(f"hardware {hardware.get('current')} (requested {hardware.get('requested')})")
out(f"commit {payload.get('sha')} (repo head) · running {runtime.get('sha')}")
out(f"host {payload.get('host') or space_host(args.space)}")
return EXIT_OK
if name == "logs":
lines = Hub(args.hub_url, http).logs(args.space, kind=args.kind, max_seconds=args.max_seconds)
target = Path(args.out).expanduser()
target.parent.mkdir(parents=True, exist_ok=True)
target.write_text(redact("\n".join(lines) + ("\n" if lines else "")))
out(f"wrote {len(lines)} {args.kind}-log line(s) to {target}")
return EXIT_OK
app = SpaceApp(args.space_url or space_host(args.space), http)
info = app.info()
published = (info.get("named_endpoints") or {}).get(ENDPOINT)
if published is None:
out(f"the Space publishes no {ENDPOINT}")
return EXIT_ILLEGAL
problems, table = generate_contract_problems(list(published.get("parameters") or []))
for line in table:
out(line)
for problem in problems:
out(f" ✗ {problem}")
out("MATCHES the local field spec" if not problems else "MISMATCH: a --live run would refuse")
return EXIT_OK if not problems else EXIT_ILLEGAL
except (HttpError, TransportError, ValueError) as exc:
out(f"{name}: {exc}")
return 1
def _entry() -> int:
try:
return main()
except KeyboardInterrupt:
print("interrupted", file=sys.stderr)
return 130
except Exception as exc: # noqa: BLE001 -- the traceback is formatted here, so it is redacted here
print(redact("".join(traceback.format_exception(exc))), file=sys.stderr)
return 1
if __name__ == "__main__":
raise SystemExit(_entry())
#!/usr/bin/env python
"""Run a plan of calls against the private LTX-2.5 Studio Space under a hard daily GPU budget.
# DRY RUN (the default): resolve, validate and price every call; contacts nothing
python scripts/space_matrix.py --plan scripts/space_plan.example.json \\
--assets-dir ~/assets --quota-ledger ~/space-ledger.jsonl
# the same plan, executed; when it stops at the ceiling (exit 3), the same command with
# --resume on the next UTC day runs what is left (calls with an outcome in --out are skipped)
python scripts/space_matrix.py --plan plan.json --assets-dir ~/assets \\
--quota-ledger ~/space-ledger.jsonl --out ~/space-runs/matrix --live [--resume]
# one call from flags; the historical 11-mode loop is --plan default
python scripts/space_matrix.py --mode text_to_video --width 704 --height 480 --num-frames 49
# read-only, one GET each, no GPU time
python scripts/space_matrix.py status
python scripts/space_matrix.py info
python scripts/space_matrix.py logs --out run.log
Exit codes: 0 plan complete · 2 illegal plan or contract mismatch (nothing spent) · 3 stopped at
the ceiling · 4 the platform refused on quota · 5 a call failed (with --keep-going the rest of the
plan still runs, and the exit is still 5).
Cost figures are BOUNDS. There is no readable ZeroGPU quota endpoint, so billed seconds cannot be
measured, only bounded: the conservative figure is the wall clock from the moment the Space starts
the event to its completion, times the size multiplier. The ceiling check before each call adds
that call's full reservation (``zerogpu.estimate_duration``, times the multiplier) to today's
conservative total.
The token is read when a command needs it (``HF_TOKEN``, else the Hub's cached token file), is
sent only as the ``Authorization`` header to the Space and the Hub API, and is redacted from
every line, file and error this script writes.
"""
from __future__ import annotations
import argparse
import hashlib
import html
import json
import os
import re
import shutil
import sys
import time
import traceback
import urllib.error
import urllib.parse
import urllib.request
import uuid
from contextlib import contextmanager
from dataclasses import dataclass, field, replace
from dataclasses import fields as dataclass_fields
from datetime import UTC, datetime, timedelta
from pathlib import Path
from typing import Any
REPO = Path(__file__).resolve().parent.parent
# `tests/` for `verify_streams`, the registry-driven stream checker.
for _path in (REPO, REPO / "tests"):
if str(_path) not in sys.path:
sys.path.insert(0, str(_path))
import verify_streams # noqa: E402
from ltx25_studio import constraints, modes, outputs, zerogpu # noqa: E402
from ltx25_studio.ui import fields # noqa: E402
__all__ = [
"API_FIELD_KEYS",
"DEFAULT_MODE_ORDER",
"DEFAULT_QUOTA_CEILING_SECONDS",
"ENDPOINT",
"EXIT_CALL_FAILED",
"EXIT_CEILING",
"EXIT_ILLEGAL",
"EXIT_OK",
"EXIT_QUOTA",
"GATE_ACCEPTED",
"GATE_DEFERRED",
"LORA_ENDPOINT",
"OUTPUT_SLOTS",
"UNBILLED_CATEGORIES",
"Call",
"Ledger",
"PlanError",
"QuotaLedgerUnreadable",
"Secret",
"app_gate",
"as_the_space",
"classify",
"count_onsets",
"default_plan",
"generate_contract_problems",
"load_plan",
"main",
"measure_audio",
"measure_video",
"platform_refusal",