Contributing: Are these TODOs still on the roadmap?

[Bias92]

Hi, I'm interested in contributing to FlashInfer. I found a few TODOs/FIXMEs in the codebase and wanted to confirm they're still relevant before starting work:

1. include/flashinfer/norm.cuh — // TODO(kaixih): add support for fp8 quantization if needed
2. include/flashinfer/attention/persistent.cuh — // FIXME: fix the offset calculation
3. include/flashinfer/sampling.cuh — // TODO: compare the speed of log2 and log

For context, I have experience with FP8 RMSNorm CUDA kernel development (merged PRs to vLLM) and would be happy to take on any of these if they're still needed.

Would appreciate any guidance on which of these would be most valuable to the project. Thanks!

[Bias92]

I also found a potential correctness issue in RMSNormQuantKernel and FusedAddRMSNormQuantKernel (include/flashinfer/norm.cuh).

Both kernels hardcode the FP8 E4M3 max value for clamping:

output_vec[j] = fmaxf(-448.0f, fminf(output_vec[j], 448.0f));

However, the output type O is dispatched via DISPATCH_DLPACK_DTYPE_TO_CTYPE_FP8 in csrc/norm.cu, which includes both E4M3 (max=448) and E5M2 (max=57344). When the output dtype is E5M2, this clamp silently truncates ~99% of the representable range.

I have a fix ready — will submit a PR shortly.