-
It is essential to keep up to date with OpenAI Triton, to get the latest features, and reduce the difficulty to upstream our changes to OpenAI Triton.
This ticket is continuation of:
- #1448
- #1…
-
### Motivation.
Recently, the OpenAI Triton backend for AMD hardware [PR 3643](https://github.com/vllm-project/vllm/pull/3643) was merged, which is so far the only flash attention backend with the…
-
### Required prerequisites
- [X] Make sure you've read the [documentation](https://pybind11.readthedocs.io). Your issue may be addressed there.
- [X] Search the [issue tracker](https://github.com/pyb…
-
In the ops/flash_attention.py, K, V blocks are accessed through `make_block_ptr`. For example, I have a question:
The input tensors `q, k, v` are of size (Batch, n_head, seq_num, dim_per_head), but…
-
I have noticed that the README states Linux as the only compatible platform. https://github.com/openai/triton#compatibility
Some people in the past have managed to compile on Windows https://github…
-
The latest Triton refactoring removed the Intel Triton backend from the third-party, `llvm-target` branch is a fork of `openai/Triton` with in-tree modifications.
To upstream Intel XPU Triton backend…
-
Hi I want to run triton on cpu, i saw "That refers to the CPU support, which won't happen shortly, but you can offload triton to pytorch using the interpreter mode. https://github.com/openai/triton/c…
-
Dear authors, thanks so much for your fantastic work, I'm wondering if the IP adapter will be integrated in the Diffusion Backend someday? The IP Adapter pipeline should be suitable with SD 1.5, SDXL.…
-
I note this project use triton write kernel. It's cool, so can you share how learn triton ?
-
It seems that `triton` has recently switched to `mathlib` in lieu of `libdevice` which causes following errors in `nn.triton_based_modules`:
```bash
AttributeError: module 'triton.language' has no a…