taeold / djev-run

# djev-run
Serve DiffusionGemma-Jev (`djev`) on Cloud Run
with an NVIDIA RTX PRO 6000 Blackwell GPU, built on
mmastrac/djev.
Built-in demo apps, inspired by:
- `/snake`: mizorewww/laya-coreml
- `/dino`: virajbhartiya/laya-vs-jev
- `/tetris`: trungdq88/jev-tetris
--------------------------------------------------------------------------------
## Deploy on Google Cloud Run
Follows
Cloud Run GPU best practices.
### Step 1: Upload Model to GCS
 [code] 
### Step 2: Deploy to Cloud Run
You have two options for deployment:
#### Option A: Pre-built `djev-run` Container (Recommended)
Deploy the pre-built image with `/dev/shm` staging and all `vllm serve` flags baked in.
 [code] 
#### Option B: Raw `vLLM` Nightly Container (Zero custom Dockerfile)
Directly deploy the upstream Docker container running pure OpenAI completions.
 [code] 
### Step 3: Query the Model
 [code] 
 [code] 
 [code] 
 [code] 
`ghcr.io/taeold/djev-run:latest` also includes support for the `/v1/systemone` endpoint:
 [code] 
 [code] 
--------------------------------------------------------------------------------
## Performance
- Cold start from zero instances: ~47.5s (down from 4m 05s baseline)
- Median response time: 61 ms on server, 117 ms end-to-end (`steps=1, samples=1`; 121 ms end-to-end with `samples="auto"`)
- Batch throughput: ~100-123 requests/sec at `concurrency=32`
JevBench v1.3.0 (`N = 231` public suite):
| Configuration | Overall Accuracy | Composite Score | Median Response Time |
| --- | --- | --- | --- |
| `djev-run (steps=1, samples=1)` | 81.4% | 73.4 | 117 ms |
| `djev-run (steps=1, samples="auto")` | 81.8% | 73.5 | 121 ms |
| `api.djev.dev` | 81.8% | 73.0 | 237 ms |
--------------------------------------------------------------------------------
## Pricing
1 NVIDIA RTX PRO 6000 GPU (20 vCPU, 80 GiB RAM) costs $3.19 per hour while
active and scales to $0 when idle with `--min-instances=0`.