[bug][cubin] Binaries for fp8_blockscale_gemm_sm90 not present in flashinfer-cubin and flashinfer-jit-cache

[jasonlizhengjian]

Issue in vllm: vllm-project/vllm#33833

In the cu130 container for vllm, it seems that this kernel is being jit compiled despite latest flashinfer-cubin and flashinfer-jit-cache being installed (tested with both 0.6.3 and 0.6.2), this causes a error with cublasLt.h not found

[Edward-lyz]

The issue has been reproduced. The problem is that this Kernel is not included in aot.py, which requires JIT compilation, and your image lacks CUDA headers, causing the temporary compilation to fail. For this issue, it can be resolved by adding this Kernel to aot.py to generate the cache. I have already added the relevant code to my forked repository and would like to discuss whether it can be contributed to the upstream repository. @yzh119 @jasonlizhengjian

[jasonlizhengjian]

Thanks for tracing down the issue.
Do we know why this was not included in aot.py? Should we check if any other kernels which are missing?