# Six A40 deployment, 2026-10-24 Pod: `69.30.85.145:22115`. Runtime: `/opt/styleedit-streaming-20261024`. Root-disk control state: `/opt/styleedit-sixgpu-20261024`. All new media, indexes, databases, receipts, caches, and logs stay on the 1 TB root disk. The existing network volume is read only for restoring old assets and the prior verified state; no task data is moved onto it. `sixgpu_run.py` owns 84 styles in six disjoint shards (14 styles/GPU). Each worker inherits one `CUDA_VISIBLE_DEVICES` and its own `STYLEEDIT_GPU_LOCK`. The deployed workers.py reads that variable with the old single-GPU lock as fallback. Each GPU runs its own long-lived DINO and DA3 worker, preserving the existing within-GPU overlap and checks. The prior verified tuples and rejects are imported without modifying the old state directory. Three independent exact archive indexes build while all six production shards run with `--incremental-index`. Each shard polls committed catalogue entries every 15 seconds and processes a pair as soon as both original and styled clips are available. It never waits for complete indexes. Archive ordering can still leave GPUs idle until pairs are found. The archive scans cover about 534.67 GiB and retain metadata only. Tar scanning is sequential, including correct absolute offsets when resuming; a completed prefetch is kept for consumption instead of re-requesting the same block. The target is 100 verified distinct source clips per style. Candidate exhaustion or errors appear in per-GPU summaries and quota reports; thresholds are never relaxed. Downloads: 4/GPU (24 total). Uploads: 1/GPU (6 total). Bounded window: 6/GPU (36 total). Scratch cap: 60 GiB/GPU (360 total), 40 GiB disk reserve. Videos stay under `/opt/styleedit-sixgpu-scratch`. Cleanup every 60 seconds and on shard exit removes verified media only. Failed uploads retain media. Immutable-commit size/SHA256 verification is required before counting or deleting a tuple. Source/style bytes stay unchanged; DA3 uses the GPU depth/NVENC path; no NPZ. `sixgpu_publish.py` refreshes HF `metadata.jsonl` and `progress/sixgpu-status.json` every five minutes from verified receipts. Failures retry without deleting data. The manifest uses immutable video URLs. `verify_live.py` tests delayed pair readiness, completed quotas, duplicate-free prefetch, exact archive member offsets, and resume at a committed tar boundary. Its in-memory CPU benchmark is not an end-to-end performance claim. Read `status.json`, `supervisor.log`, `index-*.log`, `gpu*-pilot.log`, `gpu*-production.log`, `gpu*/summary.json`, and `manifest-local.log` under the root-disk control directory (`supervisor-local.log` is the active supervisor log). The supervisor and processing jobs are detached from SSH. No automatic Pod shutdown is configured. HF credentials are read from a restricted root-only credential file and are absent from scripts and logs. Python environments and restored model weights are on ephemeral `/opt`; a replacement Pod must restore them before restarting. Root-disk state is ephemeral; verified media and the manifest are uploaded to HF, while unuploaded media remain local for retries. Do not start two supervisors, or fall back to the old single state when resuming this deployment. 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.