[DGX Station][Inference] GB300 Mamba SSU falls back to the default configuration
- Dominant language
- TypeScript
- Stars
- 22.5k
- Forks
- 3.1k
- Avg merge
- 1d 1h
- Merged PRs (30d)
- 715
Description
## Description
The v0.0.97 dual DGX Station Nemotron Ultra service cannot find the GB300-specific Triton Mamba selective-state-update configuration. Both pipeline workers fall back to the default configuration and warn that performance may be sub-optimal.
Platform scope: Reproduced on dual DGX Station only; other platforms not tested
Regression: Unknown — earlier versions not tested
## Environment
```text
Device: Dual DGX Station (GB300)
OS: Ubuntu 24.04.4 LTS
Architecture: aarch64
Node.js: v22.23.2
npm: 10.9.8
Docker: 29.2.1
OpenShell CLI: 0.0.85
NemoClaw: v0.0.97
OpenClaw: 2026.7.1 (2d2ddc4)
vLLM: 0.25.1
Model: nemotron-ultra
```
## Steps to Reproduce
1. Prepare two GB300 DGX Stations using the supported dual-Station fabric setup.
2. From Station A, install v0.0.97 with `NEMOCLAW_DGX_STATION_PEER` set to Station B.
3. Allow the distributed Nemotron Ultra vLLM service to start.
4. Inspect the vLLM startup logs on both pipeline workers.
## Expected Result
The pinned vLLM runtime should include and select a tuned Mamba selective-state-update configuration for `NVIDIA_GB300` with `headdim=64`, `dstate=128`, and `cache_dtype=float32`.
## Actual Result
Both pipeline workers select the Triton Mamba SSU backend but cannot find the GB300-specific configuration file. They fall back to the default Mamba SSU configuration and warn that performance may be sub-optimal.
## Logs
```text
Using triton Mamba SSU backend.
Using default Mamba SSU config. Performance might be sub-optimal! Config file not found at {vllm-package}/selective_state_update/headdim=64,dstate=128,device_name=NVIDIA_GB300,cache_dtype=float32.json
```
Contributor guide
Assessment
This issue has not been assessed yet.