Fix: support TileLang 0.1.9

[benglewis]

Summary

• update the exact TileLang pin from 0.1.8 to 0.1.9
• cap apache-tvm-ffi below 0.1.10, matching the working compatibility ceiling identified in apache-tvm-ffi<=0.1.12 (after #984) is incompatible with pinned tilelang==0.1.8 — measured working ceiling is 0.1.9 #986

This allows Mamba to resolve alongside vLLM releases that require TileLang 0.1.9.

Validation

• verified pyproject.toml and setup.py carry the same dependency pair
• git diff --check

A CUDA Mamba/MIMO smoke test is still required before publishing a rebuilt wheel.

[tridao]

cc @berlinchen7
I did try 0.1.9 at some point very briefly and saw some perf regression on some other kernel, so we should check that the MIMO kernels don't regress after updating tilelang.

[Chessing234]

looks reasonable. maybe pin an upper bound too so the next tilelang break doesn't silently install?

[benglewis]

@Chessing234 I have added the pin on the upper bound instead of the fixed version. @tridao How can I test the MIMO kernels? What GPU or similar would I need to test and how?

[bvolpato]

Could we allow apache-tvm-ffi==0.1.10 here? vLLM v0.26 pins it alongside tilelang==0.1.9, so the current <0.1.10 bound still makes that combination unsatisfiable. I tested TileLang 0.1.9 with tvm-ffi 0.1.10 and sampled MIMO forward/backward cases passed.

Could we also run a broader MIMO perf pass before raising the floor? On SM120 I measured one N64 forward shape about 5.4% slower with TileLang 0.1.9, 0.6369 ms versus 0.6042 ms across three medians, while the smaller N16 case was flat. This looks shape-specific, but seems worth checking.