ByteDance Seed fixes full-pipeline FP8 RL instability with Calibrated Clipping across 8B-32B models
ByteDance-Seed · hf · 2026-09-23
ByteDance Seed's new research tackles a previously overlooked failure mode in full-pipeline FP8 reinforcement learning for LLMs: compounded quantization noise distorts the importance ratio, pushing negative-advantage tokens outside the trust region and erroneously zeroing their gradients, so pathological outputs accumulate as entropy surges and garbled text.
- Prior fixes (e.g. TIS) addressed train-inference mismatch, but instability persisted across the full FP8 pipeline.
- The proposed Calibrated Clipping aligns FP8 clipping bounds with high-precision BF16 distributions by matching the lower-bound clipping quantile and rebalancing the upper bound.
- Experiments span GRPO and DAPO, model scales from 8B to 32B, and multiple FP8 scaling granularities, eliminating entropy spikes and matching BF16 baseline performance.
More from Infra
- fmgo: call Apple's on-device Foundation Models from Go with no CGO and no Swift — Super_Run_8466 · 2026-09-23
- Huawei unveils Peerium architecture: nested BSP unifies million processors into one computer — Dr_Singularity · 2026-09-23
- Grok explains why DeepSeek picked DualPipe + ZeRO-1 over ZeRO-3 on 2048 H800s — TheZachMueller · 2026-09-23
- AI costs fall 47% per quarter, 4x faster than DNA sequencing: Epoch AI — daveholtz · 2026-09-23
- M5 Ultra LLM test: 4x faster prompt processing, but double the power draw — DigitalguyCH · 2026-09-23
- $500 of Dell OptiPlexes become a diskless netboot lab where AI agents can't brick the hardware — colinmcnamara · 2026-09-23