starcoder
by bigcode
15.8B params · text-generation · 2.9k likes · 6.4k downloads
starcoder is a 15.8B parameter model. At Q4 quantization it requires 8GB of VRAM. It runs comfortably on GeForce RTX 4090 (82 tok/s), GeForce RTX 5090 (124 tok/s), M4 Max 128GB (30 tok/s).
GPU compatibility
| GPU | VRAM | Q4 Decode | Verdict |
|---|---|---|---|
| GeForce RTX 4090 | 24GB | 82 tok/s | comfortable |
| GeForce RTX 5090 | 32GB | 124 tok/s | comfortable |
| M4 Max 128GB | 128GB | 30 tok/s | comfortable |
| M4 Pro 48GB | 48GB | 15 tok/s | tight |
| M4 Pro 24GB | 24GB | 15 tok/s | tight |
| A100 PCIe 80 GB | 80GB | 152 tok/s | comfortable |
| H100 SXM5 80 GB | 80GB | 305 tok/s | comfortable |
| GeForce RTX 3090 | 24GB | 73 tok/s | comfortable |
| Radeon RX 7900 XTX | 24GB | 60 tok/s | comfortable |
| GeForce RTX 4080 | 16GB | 58 tok/s | comfortable |