SkyRL's IsoExec brings bitwise logprob parity between vLLM rollouts and Megatron training
vllm_project · x · 2026-08-21
In RL training, a vLLM rollout engine and a Megatron trainer can run the same policy yet disagree on token logprobs due to floating-point non-associativity.
SkyRL's IsoExec combines an execution contract with a unified model, aligning rounding-sensitive execution choices across rollout and training. Bitwise parity holds across different TP/EP/SP layouts; for Gated DeltaNet, a chunkwise-parallel recurrent algorithm makes parallel training/prefill bitwise identical to recurrent decode.
Results: Qwen3.5-35B-A3B + DAPO, 8×H100, 50 steps — logprob diff drops from 1.6e-2 to 6.7e-7, with 25.3% full-step overhead. vLLM's scheduler and CUDA graphs still apply.
More from Infra
- Starcloud seeks FCC nod for 88,000 satellites with 25 staff — HaktanSuren · 2026-08-22
- Canonical Launches Groundwork to Track Actual Data Center Permits — ai · 2026-08-22
- Ultrafast Qwen3-TTS: 34ms TTFA & 10 RPS on H100 — Forsaken_Goal3692 · 2026-08-22
- Data center jobs pay premium: facility managers +64%, network techs +42% — a16z · 2026-08-22
- TRiP: Transformer inference and training in plain C (Gemma, Llama2, PaliGemma) — RelevantShape3963 · 2026-08-21
- App SaaS industry's biggest risk: running cloud infra without knowing it — matt_slotnick · 2026-08-21