This is a bit much. If we want to do this, we should parse to an AST and left-factor on that instead.

Then again, that would take more lines but would offer more opportunities to optimize the AST.

I'll defer to @ggerganov if he is OK with this PR. I can run a few test scenarios to check for regressions.

Edit: On second look, it's probably close to the minimal implementation for left factoring. However, I do prefer these optimizations be done against an AST. In general, I think many of the outstanding security issues related to grammars can be handled by analyzing an AST.

But why? This is a 120-line-of-code optimization, mostly enclosed in one transformation function, with a huge performance gain. Sure, we can go full AST mode, but that would mean a complete rewrite of the grammar engine and meanwhile, people are having real problems with grammar efficiency and OpenClaw (see the attached issue). I don't see a problem with reworking the grammar overall, but I also don't see why we should take the "all-or-nothing" approach.

It's not a rewrite, it is changing the parser to emit an AST and then visit to emit the PDA rules.

Like I said, I can offer testing for regressions if this is desirable as-is.

Sounds like a rewrite to me but if you want to do it then I don't see why not :) but let's see what @ggerganov has to say.

Dear @pwilkin ,

I noticed your update, and I tried to use b8508 along with your modified source file and rebuild the code, then run the llama-server again, the gramma error still displayed.

Thanks,
Tang Laoya