Following the steps here to compile 405b with direct codegen causes the benchmark to hang.
https://github.com/nod-ai/iree-model-benchmark/tree/main/llama3

This issue only occurs with direct codegen, and compiling with the asm kernel works fine.

Need to triage and identify which commit introduced the regression, as this setup used to work correctly.

CC: @MaheshRavishankar