You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[docs]: write comprehensive onboarding tutorials for NVIDIA and AMD users #23
Compose step-by-step technical documentation detailing environment isolation, kernel compilation instructions, driver verification, and full GRPO alignment script invocation on both NVIDIA and AMD environments.
Hi @Flink-ddd , I’ve made some contributions toward the NVIDIA/B200 onboarding part of this issue in PR #547. The PR has been rebased onto main, and I’d appreciate your review and suggestions.
The contributions include:
A B200 onboarding guide covering environment isolation, driver/compiler checks, SM100 kernel compilation, backend verification, and a reproducible single-GPU synthetic GRPO example.
Deterministic CUDA LogP training support, including explicit FP64, indexed/out interfaces, recursive differentiation, and torch.func transforms, reusing the existing deterministic CUDA kernels.
Validation reports and figures with actual backend traces, numerical errors, and reproduction commands.
On a single B200, the main acceptance suite recorded 701 passes and 30 skips—15 SM90-only and 15 Ascend-only cases, counted separately from passes. Additional FP64 and layout checks passed separately. Derivative validation covers orders 1–8 across four precisions, with additional FP64 reverse-mode/mixed checks through order 10; this is finite validation.
This contributes to the NVIDIA/B200 portion; AMD onboarding remains open. Feedback on the setup instructions and validation scope would be especially helpful. Thank you for your time!
Activity
feat(docs): [RL-Align#23] add A100 validation materials
feat(docs): [RL-Align#23] Add A100 validation documentation
feat(docs): [RL-Align#23] Add A100 validation documentation
feat(docs): [RL-Align#23] Add A100 validation documentation
fusheng-ji commentedon Oct 11, 2026
Hi @Flink-ddd , I’ve made some contributions toward the NVIDIA/B200 onboarding part of this issue in PR #547. The PR has been rebased onto main, and I’d appreciate your review and suggestions.
The contributions include:
On a single B200, the main acceptance suite recorded 701 passes and 30 skips—15 SM90-only and 15 Ascend-only cases, counted separately from passes. Additional FP64 and layout checks passed separately. Derivative validation covers orders 1–8 across four precisions, with additional FP64 reverse-mode/mixed checks through order 10; this is finite validation.
This contributes to the NVIDIA/B200 portion; AMD onboarding remains open. Feedback on the setup instructions and validation scope would be especially helpful. Thank you for your time!