Turn the README accelerator list into a decision gate before ./build.sh. Prefill and Decode stay two phases of one engine.
Supported today
| Hardware | Software | Env | Status |
|---|---|---|---|
| NVIDIA SM100 | CUDA 13 · GCC ≥ 13 · Python 3.12 · Linux | DEEPINFER_ACCELERATOR=nvidia | Supported |
| Huawei Ascend 950 | CANN 9.2 · same host deps | DEEPINFER_ACCELERATOR=ascend | Supported |
910C & H20 — draft RFCs only
Not supported yet. Community RFCs propose expansion; do not install as if these SKUs were listed in the README.
| RFC | Target | Capacity / blocker | Link |
|---|---|---|---|
| #46 | Ascend 910C/A3 · CANN 9.1 · offline-first | Platform expansion beyond 950 + CANN 9.2 | PR #46 |
| #47 | SM90 · 4× H20-3E (141 GB) | FP4 fit ~499/564 GB; need FP8 grouped GEMM; FlashMLA Hopper drop 2026.09.30 | PR #47 |
Serving gaps
| Missing in OSS | Why it matters |
|---|---|
| KV cache system | Online multi-turn / high QPS |
| VE/Prefill/Decode disaggregation service | Full glue not released |
| DP load balancing | Replica balance for P and D |
Skip checklist
- Confirm SM100 or Ascend 950 (else skip / watch RFCs)
- Confirm CUDA 13 or CANN 9.2
- Plan offline V4.1-Flash unless you bring serving glue
- Agents → dshbase.com · protocols → recipe
Benchmark topology: Prefill 4× SM100 · Decode 64× SM100 (README). Next: install · hardware.