Evo 2 模型检查点依赖 FP8 矩阵乘法精度，兼容性问题解析
Accuracy is not compatibility: two Evo 2 checkpoints hinge on one FP8 matrix multiplication
有人发现 Evo 2 的 DNA 模型换了新 GPU 就算不准了，还写了个模拟器救回来，做模型部署的可以看看
Evo 2 是一个 DNA 语言模型，其 1B 和 40B 检查点在文档中标注了 Hopper FP8 硬件要求。实测在两块 Blackwell（sm_120）GPU 上 1B 和 40B 均未通过前向测试，RTX 5070 Ti 上 40B 的 next-base 准确率仅 25.7%，接近随机猜测。作者开发了 H100 FP8 GEMM 模拟器 mmaemu，将问题定位到 1B 模型 21 个投影中的第一个。仅模拟该投影后，40B 在 RTX 5070 Ti 上 BRCA1 变异分类 AUROC 从 0.434 提升到 0.869，与 H100 的 0.868 基本一致。
Accuracy is not compatibility: two Evo 2 checkpoints hinge on one FP8 matrix multiplication
FP8 formats specify how values are stored, not how products are summed; this gap can break checkpoints. Evo 2, a DNA language model, documents a Hopper FP8 requirement for its 1B and 40B checkpoints. On two Blackwell (sm_120) GPUs both fail the developers' forward test (the 7B passes); on an RTX 5070 Ti the 40B's next-base accuracy is 25.7%, near guessing. Greater accuracy breaks them, as BioNeMo saw for the 40B on Hopper: natively, 99.996% of one projection's outputs are correctly rounded (H100: about 30%), and fp32 accumulation of the same operands fails. Our H100 FP8 GEMM emulator, mmaemu (bit-exact on captured 1B projections), localises the dependence to the 1B's first projection (of 21), where exact sums fail but pass when shrunk by the H100's 1.4% (measured in-sample), not 2.1%. The 40B fails with only its first projection native. Emulating only that projection, weights unchanged, raises 40B BRCA1 variant-classification AUROC from 0.434 to 0.869 (H100: 0.868) on the RTX 5070 Ti, despite a narrow forward-test miss, and 0.481 to 0.867 on an RTX PRO 6000. We tested one model family but not training origin; releases should specify reference arithmetic.