Base-2 vs. base-e log-sum-exp mismatch silently corrupts attention in SGLang

A base-2 vs. base-e log-sum-exp mismatch in SGLang's merge v2 API silently corrupts attention outputs for models using MLA with the FlashInfer backend when radix cache hits exceed 8k, affecting DeepSeek V2/V3/R1/V3.1/V3.2, bailing, glm4moe_lite, and Sarvam. The issue arises because the merge operation assumes natural logs, but FlashInfer uses base-2 logs, leading to incorrect outputs. A fix has been proposed in a FlashInfer pull request.

If you serve models with mla on sglang with --attention-backend flashinfer and you get radix cache hits above 8k your outputs are wrong right now. DeepSeek V2/V3/R1/V3.1/V3.2, bailing, glm4moe lite, Sarvam affected. Fix: Konstantin on X: "If you serve models with mla on sglang with --attention-backend flashinfer and you get radix cache hits above 8k your outputs are wrong right now. DeepSeek V2/V3/R1/V3.1/V3.2, bailing, glm4moe lite, Sarvam affected. Fix: https://t.co/IRxPRx9Ojf" - flashinfer api proposal for clarity: github.com/flashinfer-ai/… https://github.com/flashinfer-ai/flashinfer/pull/4547 some context: there is merge v2 api in sglang and it does: S = log exp s a + exp s b v = v a·exp s a − S + v b·exp s b − S which is true if and only if s a and s b are natural logs. but this is not true for