Motivation.

Almost all PassConfig field names have enable_ in the name, which is unnecessarily verbose. They are also pretty long, and sometimes not descriptive enough. Finally, enable_fusion should be split into rmsnorm+quant and activation+quant flags as we want to control these flags separately.

Proposed Change.

We should rename the flags:

• enable_async_tp -> fuse_gemm_comms
• enable_attn_fusion -> fuse_attn_quant
• enable_fi_allreduce_fusion -> fuse_allreduce_rms
• enable_fusion -> fuse_norm_quant, fuse_act_quant
• enable_noop -> eliminate_noops
• enable_sequence_parallelism -> enable_sp

For future RoPE-based fusion passes, the flags will look like:

• enable_qknorm_rope_fusion -> fuse_qknorm_rope
• enable_rope_cache_fusion -> fuse_rope_cache
• ...

We can deprecate the original flags in the next release and map them to the new ones, and remove them 1 or even 2 releases later (shouldn't be hard to support). These flags will be used less commonly after -O optimization levels land anyway.

Feedback Period.

1 week, 11/3 - 11/7

CC List.

@zou3519 @youkaichao @mgoin @ilmarkov @nvpohanh @pavanimajety

Any Other Things.

With passes following a common construction convention, we can also add a full_pass_pipeline arg where users can control the exact order of the passes if necessary, but that is less likely to be needed urgently and can be added later.

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.