flexflow / FlexFlow

FlexFlow Serve: Low-Latency, High-Performance LLM Serving
https://flexflow.readthedocs.io
Apache License 2.0
1.6k stars 219 forks source link

Improve the design of the `compiler` library #1315

Open wmdi opened 4 months ago

wmdi commented 4 months ago

This issue is to track the remaining TODOs to improve the design of compiler.

Eventually it would be good to add infrastructure to include dot descriptions of these graphs in these testcases for readability accompanied with checks to make sure they don't get out of date. It might even be worth writing a quick parser and DSL for specifying them as otherwise it's hard to tell what these testcases are actually doing (at least for people who didn't write the testcases themselves 🙂 )

lockshaw commented 4 months ago

@wmdi When you divide up this issue can you make sure to make individual issues for (1) separating up the DP, and (2) separating up the MCMC search? I'm guessing it'll be easier to handle them individually rather than solving both at once

lockshaw commented 4 months ago

@wmdi Can you also make sure the rapidcheck issues you ran into get a corresponding issue created?

lockshaw commented 4 months ago

@wmdi Can you make sure that a more detailed issue gets created for https://reviewable.io/reviews/flexflow/FlexFlow/1229#-NpXEwmgBs4Q6-g0dSIh? This issue doesn't really mention it currently

lockshaw commented 4 months ago

@wmdi Can you make sure https://reviewable.io/reviews/flexflow/FlexFlow/1229#-NpXFKHVCMVgHr2GOPAK gets linked to in whatever issue(s) get created for graph hashing?

lockshaw commented 4 months ago

@wmdi Can you make sure that https://reviewable.io/reviews/flexflow/FlexFlow/1229#-NpXIiYi1DIya7jad3kQ gets linked in the issue that you create for handling variants in the graph library?