🎙 LEHJA — build dashboard
Live voice translation + voice clone · deploying on 2× RTX A6000 with the winning translator · Updated 2026-06-14 16:49:01 UTC · auto-refresh 20s
OVERALL DEPLOYMENT 100%
Serving box: 2× RTX A6000 (96 GB) · GPU0: 6445/49140 MiB · GPU1: 29765/49140 MiB · ollama(35B) up
🔄 Deploying — Qwen3.6-35B-A3B (bench winner) + trained voices
✅ 1. Set up A6000 (torch, CosyVoice, whisper, ollama)100%
stage: SETUP_A6000_DONE
✅ 2. Download Qwen3.6-35B-A3B (the MT winner, ~23GB)100%
complete
✅ 3. Transfer trained Urdu + Arabic voice models100%
CosyVoice2-0.5B-ar, CosyVoice2-0.5B-ur
✅ 4. Launch pipeline service (STT GPU0 · MT GPU1)100%
up on :8770
✅ 5. Re-point lehja.app/test + quality test100%
LIVE — test it
Split: whisper(STT)+CosyVoice(cloned-voice TTS) on GPU 0, Qwen3.6-35B-A3B (MT) on GPU 1 → run in parallel, no VRAM contention. Each A6000 = 48GB.
✅ Already done
✅ Urdu voice model — trained, validated PASS (CER 0.065)
✅ Arabic Quranic voice v2 — trained, owner-validated 95% (canonical-Quran-aligned, full tashkeel)
✅ MT bench — Qwen3.6-35B-A3B WINNER (now deploying it) · Qwen3-14B baseline · gemma3:4b too weak
✅ First deploy on shared 24GB 4090 — worked but VRAM-capped MT quality (the reason we moved to A6000)
✅ H100 training box — done + destroyed (models backed up on VPS + transferred here)
⊘ QTM audio re-transcription — cancelled by owner (DB untouched)
🗺 Why the A6000 (vs prior boxes)
2× A6000 = 48GB/GPU → fits the 35B winner on one card with room; 650 GB/s (5× the GB10); 2 GPUs → real 2–4 user concurrency. Shared 24GB 4090 couldn't fit good MT (FORAN took half). H100 was faster but gone + 3× price.