Without this patch, a debug version of python complains that:
```
Fatal Python error: Python memory allocator called without holding the GIL
Python runtime state: initialized
```
Based on the discussion in #700, this PR enables downloading pybind11 in
`setup.py` without `git submodule` instead of copy-pasting pybind11
code. The downloaded pybind11 will be in `~/.triton/pybind` (like
`llvm`).
I suspect this was the cause of the "new compiles even on a warm cache"
behavior I was seeing, though haven't 100% confirmed it.
Python `set()` iteration order is nondeterministic when you create a new
process. So the same args could produce different `instance_descriptor`s
and have false cache misses.
This PR changes the `pybind11` source code management from copy-paste to
a package controlled by git-submodule.
See the discussion in #694 for details.
This revives #671 , removing the static functions that may unnecessarily hold a reference to the grid and the JITFunction object
Co-authored-by: Jason Ansel <jansel@jansel.net>
@ngimel figured this one out.
The errors we were seeing from cudagraphs capture were coming from
`cuStreamGetCtx` which is not allowed while a stream is capturing.
It appears the result of `cuStreamGetCtx()` isn't even used, so I
believe it can just be removed.
Reverts openai/triton#671
It seems like for some reason this caused out-of-memory errors on some
of our internal workloads. I'm reverting this so that HEAD can be used
in production at OpenAI, and I will work on digging into this issue
asynchronously.
This PR completely rewrites the runtime of Triton to be more lean and
clearly separate the compilation step from the just-in-time caching logic.
This should substantially reduce launch overhead.
Redo of #651 against master. Fixes#525 by catching CUDA error when we
check pytorch tensor size and rethrowing a more informative error that
says why we failed.
This PR adds several optimization capabilities in the compiler backend:
- Now using inline PTX for `tl.store`, making it possible to use things like evict_last
- For A100, mma layout can be directly converted to shared memory
- For A100, an additional "transpose" argument in `dot` allows tensors to be loaded once and used both row- and col- major.
- Fixed liveness analysis; this was broken.
- Now can load/store directly mma layout without converting. Useful for when tl.dot accumulator is initialized with DRAM data inside of an inner loop.
- `tl.dot` can now take LHS inputs in registers when it comes from a previous `tl.dot` instruction. Useful for e.g. fused attention.
UserWarning: __floordiv__ is deprecated, and its behavior will change in a future version of pytorch. It currently rounds toward 0 (like the 'trunc' function NOT 'floor'). This results in incorrect rounding for negative values. To keep the current behavior, use torch.div(a, b, rounding_mode='trunc').
