LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget
Merged summary
TL;DR - LongStraw is an architecture-aware execution stack that makes million-token RL post-training (via GRPO) feasible on a fixed GPU budget, closing the gap between long-context inference (~1M tokens) and post-training that typically stays at ≤256K tokens—especially relevant for long-trajectory AI agents.
- Reduces the live training graph by evaluating the shared prompt without autograd, retaining only model state needed by later tokens, and replaying short response branches one at a time (trading extra replay time for memory).
- On 8 H20 GPUs, it completes grouped Qwen3.6-27B scoring and response backward at 2.1M positions (groups of 2 and 8); larger group size adds only 0.21 GB peak memory, and a stress test reaches 4.46M positions.
- On 32 H20 GPUs, it validates the end-to-end path for a 2.1M-token prompt across all 78 layers of the compressed-attention MoE GLM-5.2.
- Authors explicitly frame results as demonstrating execution capacity, not training correctness—prompt state is detached and some distributed forward/gradient composition paths remain incomplete.
Sources (1)
LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget
TL;DR - LongStraw is an architecture-aware execution stack that makes million-token RL post-training (via GRPO) feasible on a fixed GPU budget, closing the gap between long-context inference (~1M tokens) and post-training that typically stays at ≤256K tokens—especially relevant for long-trajectory AI agents.
- Reduces the live training graph by evaluating the shared prompt without autograd, retaining only model state needed by later tokens, and replaying short response branches one at a time (trading extra replay time for memory).
- On 8 H20 GPUs, it completes grouped Qwen3.6-27B scoring and response backward at 2.1M positions (groups of 2 and 8); larger group size adds only 0.21 GB peak memory, and a stress test reaches 4.46M positions.
- On 32 H20 GPUs, it validates the end-to-end path for a 2.1M-token prompt across all 78 layers of the compressed-attention MoE GLM-5.2.
- Authors explicitly frame results as demonstrating execution capacity, not training correctness—prompt state is detached and some distributed forward/gradient composition paths remain incomplete.