digital-garbage
1e2ac5ca67
Learned tile-sparse attention (Veda, arXiv 2605.30325): tiles, tile plans, predictor scoring and selection in plain torch (tested on CPU, including a real tiny H3 with a stand-in kernel), an INT8 Triton kernel derived from SageAttention (untested on a GPU), installed through the same override and per-step wrapper as SLA. Runs at the predictor's trained budget; LoRA, MXFP8, Two stages, dense steps and the attention backend compose. Settings > Models offers the official Hugging Face download of the predictor on a click, with its licence beside it; taggers and this share one downloader (core/fetch.py).
75 lines
3.7 KiB
Python
75 lines
3.7 KiB
Python
"""The Veda predictor: where it lives, and an optional download on the person's own click.
|
|
|
|
FunUI never hosts, bundles or mirrors the file. Its licence (the MiniMax H3 Community License) has conditions
|
|
on distribution, so the only route offered is a fetch straight from the official Hugging Face repository,
|
|
started by a button that sits next to the licence name and a link to its text. Streaming, the `.part` name and
|
|
the size and checksum checks are core/fetch.py's.
|
|
"""
|
|
import threading
|
|
from pathlib import Path
|
|
|
|
from .._core import fetch, log
|
|
|
|
SOURCE = "FunUI Veda"
|
|
HOST = "https://huggingface.co"
|
|
LICENSE = {"name": "MiniMax H3 Community License",
|
|
"url": f"{HOST}/Veda-Sparse/Minimax-H3-T2VA-Veda-8NFE-600Step-Preview/blob/main/LICENSE"}
|
|
# Sizes and checksums are the repositories' own listing (the Git LFS record of each file).
|
|
MODELS = (
|
|
{"id": "t2va", "title": "T2VA / FL2VA", "hint": "Text, or first and last frames, to audio and video.",
|
|
"repo": "Veda-Sparse/Minimax-H3-T2VA-Veda-8NFE-600Step-Preview", "file": "minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors",
|
|
"size": 275415648, "sha256": "2a8d8845c5342756a2781e8e69563940e4bb573c9a40ebb534915ff8fd76573a"},
|
|
{"id": "r2va", "title": "R2VA", "hint": "Reference images or videos to audio and video.",
|
|
"repo": "Veda-Sparse/Minimax-H3-R2VA-Veda-Preview", "file": "minimax_h3_r2va_veda_preview_fp8.safetensors",
|
|
"size": 275415752, "sha256": "23edace948693772be33d5cc945d302d2b037a29bd44bd57071bfbdd8b02517d"},
|
|
)
|
|
_state, _lock = {}, threading.Lock()
|
|
|
|
|
|
def folder() -> Path:
|
|
import folder_paths
|
|
return Path(folder_paths.models_dir) / "veda"
|
|
|
|
|
|
def files() -> list:
|
|
"""The predictor files in models/veda, by name. A `.part` is not one."""
|
|
return sorted(p.name for p in folder().glob("*.safetensors")) if folder().is_dir() else []
|
|
|
|
|
|
def listing() -> dict:
|
|
present = files()
|
|
rows = []
|
|
for m in MODELS:
|
|
s = _state.get(m["id"]) or {}
|
|
row = {k: m[k] for k in ("id", "title", "hint", "file", "size")}
|
|
row["url"] = f"{HOST}/{m['repo']}/resolve/main/{m['file']}"
|
|
row["state"] = "downloading" if s.get("state") == "downloading" else "ready" if m["file"] in present else "error" if s.get("state") == "error" else "absent"
|
|
row.update(done=s.get("done", 0), error=s.get("error", ""))
|
|
rows.append(row)
|
|
return {"files": present, "models": rows, "license": LICENSE}
|
|
|
|
|
|
def download(model_id: str, get=None) -> bool:
|
|
"""Start the download in a thread. False when the id is unknown, the file is there already, or it is already running."""
|
|
spec = next((m for m in MODELS if m["id"] == model_id), None)
|
|
get = get or fetch.to_file
|
|
with _lock:
|
|
if spec is None or (_state.get(model_id) or {}).get("state") == "downloading" or spec["file"] in files():
|
|
return False
|
|
_state[model_id] = {"state": "downloading", "done": 0, "error": ""}
|
|
|
|
def work():
|
|
try:
|
|
target = folder()
|
|
target.mkdir(parents=True, exist_ok=True)
|
|
get(f"{HOST}/{spec['repo']}/resolve/main/{spec['file']}", target / spec["file"],
|
|
lambda done, total: _state[model_id].update(done=done), size=spec["size"], sha256=spec["sha256"])
|
|
_state[model_id] = {"state": "ready", "done": spec["size"], "error": ""}
|
|
log.info(SOURCE, f"Veda predictor {spec['title']} downloaded to {target / spec['file']}")
|
|
except Exception as exc: # noqa: BLE001 -- shown beside the button, and in the Log
|
|
_state[model_id] = {"state": "error", "done": 0, "error": f"{type(exc).__name__}: {exc}"}
|
|
log.warning(SOURCE, f"Veda predictor {spec['title']} could not be downloaded: {exc}")
|
|
|
|
threading.Thread(target=work, name=f"funui-veda-{model_id}", daemon=True).start()
|
|
return True
|