djev-run Deploys DiffusionGemma on Cloud Run's RTX PRO 6000 Blackwell for ~$3/hr
bodonoghue85 · x · 2026-09-21
A new open-source project, djev-run, lets anyone stand up a Jev API-compatible endpoint for DiffusionGemma-Jev on Google Cloud Run with a single gcloud command, running on an NVIDIA RTX PRO 6000 Blackwell GPU at roughly $3/hr while active and $0 when idle.
The two-step deployment:
- Create a GCS bucket in a supported region (us-central1, europe-west4, asia-southeast1, asia-south2), pull nvidia/diffusiongemma-26B-A4B-it-NVFP4 weights with hf download, and upload them
- Deploy with gcloud beta run deploy specifying 1 Blackwell GPU, 20 vCPU, 80Gi memory, concurrency=32 and min-instances=0, with weights mounted via a Cloud Storage volume
Built on mmastrac's djev-spark following Cloud Run GPU best practices, the repo (taeold/djev-run) is live on GitHub — a cheap way to try diffusion-based LLM serving without buying local Blackwell hardware.
More from Infra
- Qdrant experiments: 10→500 candidate depth lifts best-possible nDCG by 0.28 but real score by ≤0.01 — qdrant_engine · 2026-09-21
- gemini-cli PR Fixes Process Hang on Exit via stdin and MCP Child Process Cleanup — Pcmhacker-piro · 2026-09-21
- Daniel Lemire Tests Whether CPUs Can Take More Than One Branch Per Cycle — lemire · 2026-09-21
- Qwen3.8-27B in native 8-bit hits 37-55 tok/s on Apple Silicon, avoiding the 4-bit reasoning cliff — SnooPredictions515 · 2026-09-21
- Google's Orphaned VMs Patches Keep VMs Running While Host Kernel Goes Offline — jedisct1 · 2026-09-21
- 16GB (often 12GB) is the realistic VRAM ceiling for most people running local AI — ECrispy · 2026-09-21