Releases · RL-Align/RL-Kernel

Highlights

RL-Kernel v0.1.0 is our first public release for kernel-level train–inference consistency in RL post-training.

This release includes 174 merged pull requests from 28 GitHub contributors, covering end-to-end Qwen3-8B consistency across vime, vLLM rollout and Megatron-LM training, with CUDA, ROCm, Ascend(Partial) and MUSA(In progress) support.

End-to-end train–inference consistency: Qwen3-8B VIME integration and strict train/rollout LogP validation on CUDA and ROCm (#377, #388, #400).
Canonical multi-topology execution: 14 H100 training/rollout TP/CP configurations with 28,672 active LogPs and zero byte mismatches (#432).
ROCm deterministic execution: strict Attention, distributed FFN, vocab-parallel LogP, deterministic RCCL, direct-paged execution, MFMA GEMM and Triton Attention (#319, #325, #328, #356, #390, #396, #437).
Full-model validation: Qwen3-8B Dense WS1 train/inference closeout across CUDA and Triton-on-CUDA (#315).
Reproducibility: profile-driven CUDA/ROCm workflows through the unified rlk reproduction path (#426, #430).
Ascend: deterministic kernel coverage and Qwen3-8B Dense cross-configuration validation (#405, #406, #417).
MUSA: Moore Threads platform support, Triton-backed operator dispatch and native deterministic kernel coverage (#346, #375).

Train–Inference Consistency

fix: restore native logp API parity (#56) by @inaniloquentee
feat(alignment): add cross-configuration framework (#230) by @CyberSecurityErial
[WS1] Close embedding and LM-head invariance coverage (#243) by @inaniloquentee
[ws1]: WS1 Full Qwen3-8B Dense Train-Inference Closeout (#315) by @maxiaosong1124
fix(integrations): restore strict Qwen3 bitwise parity (#339) by @inaniloquentee
Add VIME Qwen3-8B TP4/CP2 consistency experiment and results (#377) by @inaniloquentee
feat(rocm): enable bitwise Qwen3 Dense VIME alignment (#388) by @inaniloquentee

Hardware & Platforms

[FEAT][platforms]: implement hardware-aware context for ROCm and CUDA (#1) by @Flink-ddd
[fea][Multiplatform]: add multi-architecture build support (#316) by @Flink-ddd

CUDA

[FEAT][kernels]: implement TMA-accelerated Online Softmax Fused LogP for SM90+ (#26) by @Flink-ddd
[FEAT][kernels] Fix CUDA extension build on non-Hopper SM>=90 (Blackkwell SM120) (#91) by @KJLdefeated
Add batch-invariant deterministic CUDA logp (#98) by @inaniloquentee
[FEAT][kernels] Fused linear log-prob without materializing logits — SM90 TMA + tensor-core forward, chunked backward (#122) by @KJLdefeated
[WS1][kernels] Batch-invariant logprob (CUDA) (#204) by @KJLdefeated
ci: build & test the CUDA extension on the real GPU arch, fail fast i… (#210) by @Liuzin521
feat: add SM90 linear logp backward fast paths (#211) by @inaniloquentee
[CI][SM90] fix sm90 run GPU-CI error (#221) by @Flink-ddd
[WS1][kernels] RoPE (Triton, CUDA on SM90) (#228) by @zhangj1an
[WS1][kernels] Batch-invariant attention (standard softmax--cuda version) (#240) by @maxiaosong1124
[WS1] Add CUDA and Triton SiLU/SwiGLU activation kernels (#280) by @maxiaosong1124
feat(linear-logp): add strict bitwise SM90 contract (#336) by @inaniloquentee
feat(integrations): add H100 Megatron/vLLM runtime adapters (#338) by @inaniloquentee
bench(ffn): add H100 cuBLAS and vLLM comparisons (#348) by @frank-2077
feat(cuda): promote deterministic cross-config runtime and kernel validation to main (#361) by @inaniloquentee
Enable exact-batch CUDA Graph for strict rollout (#367) by @inaniloquentee
Perf/cuda compact top p replay (#424) by @inaniloquentee

ROCm

[FEAT][kernels]: add ROCm FlashAttention backend (#104) by @FED4
feat(attention): add strict bitwise ROCm path (#319) by @inaniloquentee
feat(ffn): add deterministic distributed Triton FFN for ROCm (#325) by @frank-2077
feat(logprob): add deterministic ROCm vocab-parallel path (#328) by @hihaluemen
[ROCm] Deterministic fused linear logp in Triton (#347) by @KJLdefeated
[FEAT][distributed]: add deterministic ROCm-RCCL transport collectives (#356) by @Flink-ddd
[Bugfix] Fix ROCm build for deterministic collectives (#380) by @Flink-ddd
[feat][rocm]: add end-to-end Vime rollout-training attention ablation matrix (#385) by @Flink-ddd
[perf][rocm]: accelerate strict VIME rollout with direct paged CK and graph-safe execution (#390) by @Flink-ddd
Update H100 and MI300X distributed benchmark data (#391) by @frank-2077
[perf][rocm]: restore full Graph and cache strict Attention schedules (#393) by @inaniloquentee
fix(rocm): stabilize direct paged CK attention for bitwise R/R (#394) by @inaniloquentee
perf(rocm): MFMA batch-invariant GEMM and chunked Triton attention for strict R/R (#396) by @Flink-ddd
docs(rocm): publish G10/G11 200-step RL results (#400) by @inaniloquentee
[FEAT][ROCm]: promote deterministic cross-config runtime and kernel validation to main (#401) by @Flink-ddd
perf(rocm): restore VIME TP4 decode throughput (#403) by @inaniloquentee
perf(rocm): reduce strict no-reuse rollout overhead (#437) by @inaniloquentee