Completed experiments
Results and failure analysis ↗
IEOR, IIT Bombay
passpoli research server
| Hardware | 2 NVIDIA RTX A5000 GPUs, 24 GB VRAM each |
|---|---|
| Models run | Granite 4.2 3B and public Param2 17B A2.4B Thinking |
| Experiments | Synthetic smoke tests, expanded protocol checks, and ten ITBench Lite SRE snapshots |
| Serving | Granite through Ollama; Param2 through Transformers |
| Status | Reported experiments complete; interrupted runs labeled separately |
Current preparation
Current investigation ↗
BharatGen on AWS
SageMaker HyperPod, Mumbai region
| Cluster | bgen-cluster |
|---|---|
| Worker type | ml.p5.48xlarge: 8 NVIDIA H100 GPUs, 80 GB each |
| Model target | Internal Param17B step_20000 SFT checkpoint |
| Storage and runtime | FSx project storage; Docker experiments on compute workers |
| Verified | An idle worker candidate and cached-container startup |
| Pending | Model runtime, native tool protocol, benchmark runs, and scores |
Environment boundaries
The institute smoke environment used an isolated Python environment. The AWS workflow uses Docker on compute workers with project data on FSx. The login node is a gateway. The earlier server scripts require adaptation before use on AWS.