Faster tokens.
None of them wrong.
Attention, RoPE, and KV-append are exactly the kernels where a fast rewrite can be silently, structurally wrong — and a demo-case test will never catch it. Kernalin optimizes your serving path and proves each speedup holds before it reaches production.
Injected bug: max never subtracted from the exponent
A single demo-case probe
how most tools checkWith kv_len = 1, softmax over one element is exactly 1.0 — the output is v₀ regardless of the entire score path.
Kernalin · structural verification
how we checkA structural check sees the missing subtraction immediately — it doesn't need an input that happens to trigger the overflow.
Why LLM kernels are the dangerous ones to optimize
The bug hides at position 0
A single decode probe runs at kv_len=1 with an empty cache — where softmax is 1.0 and a rotation is the identity. Whole classes of attention and RoPE bugs are invisible there.
Corruption surfaces far from the cause
A stray KV-cache write self-heals for a step, then shows up as garbage tokens many steps later — or in another head entirely. Impossible to bisect by eye.
A wrong kernel is worse than a slow one
Ship a fast-but-wrong attention kernel and you degrade every response quietly. No stack trace, no crash — just worse output at scale.
Speed on the decode path, proven safe.
Attention & flash-decode tuning
Online-softmax tiling, KV-cache layout, and fusion across the decode step — measured on your shapes, not a demo case.
Structural verification of ATTENTION, ROPE, KV_APPEND (research preview)
For megakernels, the three ops a pos-0 gate cannot reach are checked by PTX-symbolic methods that catch the bug classes a pos-0 probe misses: 5 of 6, 3 of 5 and 3 of 5.
Bayesian benchmark on your GPUs
A verdict over a bootstrap posterior vs. your hardware’s noise floor — so a real tokens/sec win is never confused with jitter.
Lower cost per token
Faster verified kernels mean fewer GPUs for the same throughput — the savings flow straight to cost-per-token.
A tokens/sec win you can take to production.
99.4% of the posterior mass sits below the ROPE — convincingly beyond what this GPU's jitter could fake. This one earns the number.
Ship faster tokens, not faster bugs.
Bring verification-first optimization to your LLM serving stack. Join the early-access program.