Tong WU
Tong WU
Is there any evidence that can confirm such change is necessary?
Currently there's some problems with auto TMA stuff (in terms of completion mechanism, i guess) when implicit padding is needed in `T.copy`🤔 You can use `@tilelang.jit(pass_configs={"tl.disable_tma_lower": True})` as a workaround...
Hi @a1600012888, I cannot reproduce your error with the latest TileLang by directly running your script on Hopper. Could you please provide more detail about your environment configuration and how...
It's due to that TVM (`tilelang/3rdparty/tvm/src/tir/op/op.cc`) does not provide inf impl for bf16
could we directly contribute to upstream TVM? ---- Replied Message ---- | From | Lei ***@***.***> | | Date | 10/29/2025 12:34 | | To | tile-ai/tilelang ***@***.***> | |...
Could you provide your impl of `generate_qkv_1d` to help us reproduce?
Currently Tilelang wrapper has to determine all symbolic variables via given inputs' shapes at runtime, then create output tensors based on their specific values. However in this case, the compiler...
Even the first point, "determine all symbolic variables via given inputs' shapes at runtime", is quite not easy, i think, when inputs' shapes contain `PrimExpr`. Maybe we shall add some...
This is indeed a bug. Really appreciate your feedback, we'll fix this later
I've just refactored `inject_fence_proxy`. Could you please review the correctness of this pass? @LeiWang1999 @chengyupku