Datasets:
Refresh verified six-GPU manifest: 26 tuples
Browse files- pipeline/SIXGPU.md +3 -0
- pipeline/archive_cache.py +18 -0
- pipeline/download_archives.py +52 -0
- pipeline/sixgpu_publish.py +1 -1
- pipeline/streaming.py +6 -1
- progress/sixgpu-status.json +10 -10
pipeline/SIXGPU.md
CHANGED
|
@@ -53,3 +53,6 @@ restarting. Root-disk state is ephemeral; verified media and the manifest are
|
|
| 53 |
uploaded to HF, while unuploaded media remain local for retries. Do not start
|
| 54 |
two supervisors, or fall back to the old single state
|
| 55 |
when resuming this deployment.
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
uploaded to HF, while unuploaded media remain local for retries. Do not start
|
| 54 |
two supervisors, or fall back to the old single state
|
| 55 |
when resuming this deployment.
|
| 56 |
+
|
| 57 |
+
|
| 58 |
+
Direct archive download: download_archives.py uses 16 concurrent 64 MiB HTTP ranges, resumable atomic chunks in /opt/styleedit-archive-cache (root disk). streaming.py reads completed chunks locally and otherwise falls back to HTTP. Archives are not fully extracted. Six GPU incremental pair filtering remains active, followed by depth, verified HF upload and 60-second cleanup. Archive chunks are shared by jobs and retained during processing; .part files are never read as complete data. Low disk space pauses new archive requests at 48 GiB free.
|
pipeline/archive_cache.py
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Root-disk, immutable range cache; only atomic completed chunks are readable."""
|
| 2 |
+
import os
|
| 3 |
+
from pathlib import Path
|
| 4 |
+
ROOT=Path('/opt/styleedit-archive-cache')
|
| 5 |
+
BLOCK=64*1024**2
|
| 6 |
+
|
| 7 |
+
def cached_range(category,start,length):
|
| 8 |
+
chunks=[]; pos=start; end=start+length
|
| 9 |
+
while pos<end:
|
| 10 |
+
base=pos//BLOCK*BLOCK
|
| 11 |
+
path=ROOT/category/f'{base:012d}.chunk'
|
| 12 |
+
try:
|
| 13 |
+
with path.open('rb') as f:
|
| 14 |
+
f.seek(pos-base); n=min(end-pos,BLOCK-(pos-base)); data=f.read(n)
|
| 15 |
+
except FileNotFoundError:return None
|
| 16 |
+
if len(data)!=n:return None
|
| 17 |
+
chunks.append(data);pos+=n
|
| 18 |
+
return b''.join(chunks)
|
pipeline/download_archives.py
ADDED
|
@@ -0,0 +1,52 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Resume 16 concurrent downloads on root disk, without extracting full archives."""
|
| 2 |
+
import concurrent.futures,fcntl,json,os,shutil,time,threading
|
| 3 |
+
from pathlib import Path
|
| 4 |
+
from archive_cache import ROOT,BLOCK
|
| 5 |
+
from streaming import archive_sizes,http_get_range
|
| 6 |
+
from common import atomic
|
| 7 |
+
STATUS=Path('/opt/styleedit-sixgpu-20261024/download-status.json')
|
| 8 |
+
|
| 9 |
+
def main():
|
| 10 |
+
ROOT.mkdir(parents=True,exist_ok=True)
|
| 11 |
+
lock=(ROOT/'download.lock').open('a');fcntl.flock(lock,fcntl.LOCK_EX|fcntl.LOCK_NB)
|
| 12 |
+
sizes=archive_sizes();done={c:0 for c in sizes};tasks=[];errors=[];guard=threading.Lock()
|
| 13 |
+
for c,size in sizes.items():
|
| 14 |
+
(ROOT/c).mkdir(exist_ok=True)
|
| 15 |
+
for start in range(0,size,BLOCK):
|
| 16 |
+
n=min(BLOCK,size-start);path=ROOT/c/f'{start:012d}.chunk'
|
| 17 |
+
if path.exists() and path.stat().st_size==n:done[c]+=n
|
| 18 |
+
else:tasks.append((start,c,n,path))
|
| 19 |
+
# All three archive fronts advance together; complete pairs can be processed early.
|
| 20 |
+
tasks.sort(key=lambda t:(t[0],list(sizes).index(t[1])))
|
| 21 |
+
initial=sum(done.values());begin=time.monotonic();last=begin
|
| 22 |
+
def fetch(task):
|
| 23 |
+
start,c,n,path=task
|
| 24 |
+
while shutil.disk_usage(ROOT).free < 48*1024**3:time.sleep(15)
|
| 25 |
+
body=http_get_range(c,start,n)
|
| 26 |
+
tmp=path.with_suffix('.part')
|
| 27 |
+
with tmp.open('wb') as f:f.write(body);f.flush();os.fdatasync(f.fileno())
|
| 28 |
+
os.replace(tmp,path)
|
| 29 |
+
with guard:done[c]+=n
|
| 30 |
+
def report(stage):
|
| 31 |
+
elapsed=time.monotonic()-begin;total=sum(done.values());rate=(total-initial)/max(elapsed,0.001)
|
| 32 |
+
atomic(STATUS,{'stage':stage,'utc':time.strftime('%Y-%m-%dT%H:%M:%SZ',time.gmtime()),'workers':16,'cache_root':str(ROOT),'total_bytes':sum(sizes.values()),'downloaded_bytes':total,'downloaded_this_run':total-initial,'elapsed_seconds':elapsed,'bytes_per_second':rate,'eta_seconds':(sum(sizes.values())-total)/rate if rate else None,'categories':{c:{'downloaded_bytes':done[c],'total_bytes':sizes[c]} for c in sizes},'errors':errors[-20:]})
|
| 33 |
+
report('downloading')
|
| 34 |
+
with concurrent.futures.ThreadPoolExecutor(16) as pool:
|
| 35 |
+
pending={};it=iter(tasks)
|
| 36 |
+
def fill():
|
| 37 |
+
while len(pending)<32:
|
| 38 |
+
task=next(it,None)
|
| 39 |
+
if task is None:break
|
| 40 |
+
pending[pool.submit(fetch,task)]=task
|
| 41 |
+
fill()
|
| 42 |
+
while pending:
|
| 43 |
+
finished,_=concurrent.futures.wait(pending,timeout=5,return_when=concurrent.futures.FIRST_COMPLETED)
|
| 44 |
+
for future in finished:
|
| 45 |
+
task=pending.pop(future)
|
| 46 |
+
try:future.result()
|
| 47 |
+
except Exception as e:errors.append({'category':task[1],'start':task[0],'error':str(e)[:250]})
|
| 48 |
+
fill()
|
| 49 |
+
if time.monotonic()-last>=10:report('downloading');last=time.monotonic();print(STATUS.read_text(),flush=True)
|
| 50 |
+
report('complete' if not errors else 'needs_retry')
|
| 51 |
+
if errors:raise RuntimeError(f'{len(errors)} chunks failed; rerun resumes only missing chunks')
|
| 52 |
+
if __name__=='__main__':main()
|
pipeline/sixgpu_publish.py
CHANGED
|
@@ -27,7 +27,7 @@ def publish():
|
|
| 27 |
ops=[CommitOperationAdd(path_in_repo='metadata.jsonl',path_or_fileobj=str(out))]
|
| 28 |
status=ROOT/'status.json'
|
| 29 |
if status.exists():ops.append(CommitOperationAdd(path_in_repo='progress/sixgpu-status.json',path_or_fileobj=str(status)))
|
| 30 |
-
for name in ['sixgpu_run.py','sixgpu_publish.py','workers.py','SIXGPU.md','pipeline.py','streaming.py','live_index.py','verify_live.py']:
|
| 31 |
ops.append(CommitOperationAdd(path_in_repo='pipeline/'+name,path_or_fileobj=str(APP/name)))
|
| 32 |
c=HfApi().create_commit(repo_id='faraway00/StyleEdit',repo_type='dataset',operations=ops,
|
| 33 |
commit_message=f'Refresh verified six-GPU manifest: {len(manifest)} tuples')
|
|
|
|
| 27 |
ops=[CommitOperationAdd(path_in_repo='metadata.jsonl',path_or_fileobj=str(out))]
|
| 28 |
status=ROOT/'status.json'
|
| 29 |
if status.exists():ops.append(CommitOperationAdd(path_in_repo='progress/sixgpu-status.json',path_or_fileobj=str(status)))
|
| 30 |
+
for name in ['sixgpu_run.py','sixgpu_publish.py','workers.py','SIXGPU.md','pipeline.py','streaming.py','live_index.py','verify_live.py','archive_cache.py','download_archives.py']:
|
| 31 |
ops.append(CommitOperationAdd(path_in_repo='pipeline/'+name,path_or_fileobj=str(APP/name)))
|
| 32 |
c=HfApi().create_commit(repo_id='faraway00/StyleEdit',repo_type='dataset',operations=ops,
|
| 33 |
commit_message=f'Refresh verified six-GPU manifest: {len(manifest)} tuples')
|
pipeline/streaming.py
CHANGED
|
@@ -4,7 +4,7 @@ from common import *
|
|
| 4 |
import requests
|
| 5 |
SPLIT=21474836480
|
| 6 |
|
| 7 |
-
def
|
| 8 |
from huggingface_hub import get_token
|
| 9 |
blocks=[];pos=start;remaining=length
|
| 10 |
while remaining:
|
|
@@ -25,6 +25,11 @@ def get_range(category,start,length):
|
|
| 25 |
remaining-=n;pos+=n
|
| 26 |
return b''.join(blocks)
|
| 27 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
class Remote(io.RawIOBase):
|
| 29 |
"""2 cached 8MiB blocks + 1 asynchronous prefetch. Fresh ranges on retries."""
|
| 30 |
def __init__(self,category,size,block=8*1024**2):
|
|
|
|
| 4 |
import requests
|
| 5 |
SPLIT=21474836480
|
| 6 |
|
| 7 |
+
def http_get_range(category,start,length):
|
| 8 |
from huggingface_hub import get_token
|
| 9 |
blocks=[];pos=start;remaining=length
|
| 10 |
while remaining:
|
|
|
|
| 25 |
remaining-=n;pos+=n
|
| 26 |
return b''.join(blocks)
|
| 27 |
|
| 28 |
+
def get_range(category,start,length):
|
| 29 |
+
from archive_cache import cached_range
|
| 30 |
+
cached=cached_range(category,start,length)
|
| 31 |
+
return cached if cached is not None else http_get_range(category,start,length)
|
| 32 |
+
|
| 33 |
class Remote(io.RawIOBase):
|
| 34 |
"""2 cached 8MiB blocks + 1 asynchronous prefetch. Fresh ranges on retries."""
|
| 35 |
def __init__(self,category,size,block=8*1024**2):
|
progress/sixgpu-status.json
CHANGED
|
@@ -1,5 +1,5 @@
|
|
| 1 |
{
|
| 2 |
-
"utc": "2026-10-24T04:
|
| 3 |
"stage": "indexing_and_production",
|
| 4 |
"verified": 26,
|
| 5 |
"counts": {
|
|
@@ -37,39 +37,39 @@
|
|
| 37 |
"verified_pending_cleanup": 0,
|
| 38 |
"processes": {
|
| 39 |
"index-source": {
|
| 40 |
-
"pid":
|
| 41 |
"returncode": null
|
| 42 |
},
|
| 43 |
"index-global_style1": {
|
| 44 |
-
"pid":
|
| 45 |
"returncode": null
|
| 46 |
},
|
| 47 |
"index-global_style2": {
|
| 48 |
-
"pid":
|
| 49 |
"returncode": null
|
| 50 |
},
|
| 51 |
"gpu-0": {
|
| 52 |
-
"pid":
|
| 53 |
"returncode": null
|
| 54 |
},
|
| 55 |
"gpu-1": {
|
| 56 |
-
"pid":
|
| 57 |
"returncode": null
|
| 58 |
},
|
| 59 |
"gpu-2": {
|
| 60 |
-
"pid":
|
| 61 |
"returncode": null
|
| 62 |
},
|
| 63 |
"gpu-3": {
|
| 64 |
-
"pid":
|
| 65 |
"returncode": null
|
| 66 |
},
|
| 67 |
"gpu-4": {
|
| 68 |
-
"pid":
|
| 69 |
"returncode": null
|
| 70 |
},
|
| 71 |
"gpu-5": {
|
| 72 |
-
"pid":
|
| 73 |
"returncode": null
|
| 74 |
}
|
| 75 |
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"utc": "2026-10-24T04:17:20Z",
|
| 3 |
"stage": "indexing_and_production",
|
| 4 |
"verified": 26,
|
| 5 |
"counts": {
|
|
|
|
| 37 |
"verified_pending_cleanup": 0,
|
| 38 |
"processes": {
|
| 39 |
"index-source": {
|
| 40 |
+
"pid": 18737,
|
| 41 |
"returncode": null
|
| 42 |
},
|
| 43 |
"index-global_style1": {
|
| 44 |
+
"pid": 18738,
|
| 45 |
"returncode": null
|
| 46 |
},
|
| 47 |
"index-global_style2": {
|
| 48 |
+
"pid": 18739,
|
| 49 |
"returncode": null
|
| 50 |
},
|
| 51 |
"gpu-0": {
|
| 52 |
+
"pid": 18740,
|
| 53 |
"returncode": null
|
| 54 |
},
|
| 55 |
"gpu-1": {
|
| 56 |
+
"pid": 18741,
|
| 57 |
"returncode": null
|
| 58 |
},
|
| 59 |
"gpu-2": {
|
| 60 |
+
"pid": 18742,
|
| 61 |
"returncode": null
|
| 62 |
},
|
| 63 |
"gpu-3": {
|
| 64 |
+
"pid": 18743,
|
| 65 |
"returncode": null
|
| 66 |
},
|
| 67 |
"gpu-4": {
|
| 68 |
+
"pid": 18744,
|
| 69 |
"returncode": null
|
| 70 |
},
|
| 71 |
"gpu-5": {
|
| 72 |
+
"pid": 18745,
|
| 73 |
"returncode": null
|
| 74 |
}
|
| 75 |
}
|