EvSpark: Lossless Speculative Decoding for Hybrid DNA Foundation Models

DNA foundation models such as Evo2 7B adopt hybrid Hyena/attention architectures (StripedHyena2) whose single-stream autoregressive decoding is bounded by weight bandwidth at ~45 tok/s. Speculative decoding on such hybrids faces a systems problem that prior SSM work solves only partially: after a draft is verified, rolling back model state cannot be reduced to truncating a KV cache, because the inference state mixes fixed-length FIR sliding windows, infinite-memory IIR recurrences, and append-only KV caches. We present EvSpark. (i) A block-verification forward pass together with a per-position state-slicing rollback protocol handles all three state classes jointly with zero recomputation, reducing speculative overhead to 1.05-1.16x under an identity drafter and eliminating the second per-round forward pass of snapshot-and-replay schemes. (ii) A DSpark-style distilled neural drafter is ported to this hybrid architecture (gamma-parallel trunk + target hidden-state prefix + full-matrix Markov head), with an fp32-scale distillation codec required by the ~10^11 residual magnitudes of the final blocks. (iii) Under a 48-prompt (43 real + 5 synthetic) x 1024-token x 3-training-seed protocol, EvSpark delivers 3.27x on the full suite and 2.96x on real sequence at its deployment configuration (gamma=12, 150M supervised positions), holding from 1k to 262k context (1.84x-2.43x on two complete bacterial genomes at 262k) and across 32k tokens of generation depth. In exact arithmetic the scheme preserves the target distribution; empirically, greedy decoding reproduces native outputs token-for-token (zero non-tie divergences across 48 prompts x 6 checkpoints); the sampling path sits at the matched-N native noise floor for about half of 24 KL prompts per checkpoint, with the remaining deviations bounded, bf16-level, and concentrated in low-complexity repeat regions (worst-case unigram TVD 8.2%, <4% elsewhere). The budget sweet spot adds 1.06 GPU-hours of training and reaches 3.15x.