cloudhan
cloudhan
@zjing14 Chat with @ltqin on Teams. In the original issue, I'd like either: 1. ck provides enough client api instances or 2. ck make the following code to be customizable,...
Ping me if you have any question. This following is also some very important things you should take care of. https://github.com/microsoft/onnxruntime/blob/7b31bcda2e9cd45c709e3fe31a544297db37ea3c/cmake/CMakeLists.txt#L255-L262 `-mllvm -amdgpu-early-inline-all=true` is propagated from some underlying targets, you...
For more reference, seqlen 32768 ``` void vllm::paged_attention_v2_kernel(float *, float *, T1 *, const T1 *, const T1 *, const T1 *, int, float, const int *, const int *,...
@yzh119 Thanks for the quick reply. It is a little bit clearer now. Is there any branch that host the vllm benchmark part of code? Current main only has flashinfer...
The flag is simply inherited from tensorflow tfruntime toolchain, you can access it from the orphan branch https://github.com/bazel-contrib/rules_cuda/tree/subtree-tfruntime. The main branch is a rewrote version with native starlark implementation. As...
I am interested in the rule based config. Currently, rules_cuda use a [feature config DSL interpreter](https://github.com/bazel-contrib/rules_cuda/blob/1b36a264f61f4fb811d403428cc999f0d2bf3eb1/cuda/private/toolchain_config_lib.bzl) implemented in starlark, with all [test cases](https://github.com/bazel-contrib/rules_cuda/blob/1b36a264f61f4fb811d403428cc999f0d2bf3eb1/tests/toolchain_config_lib/toolchain_config_lib_test.bzl) extracted from bazel java impl. And then...
To get into bazel contrib, @alexeagle may know the process.
It is not C++ actually. I implemented bazel cc rules feature config (which is implemented in java and not exposed) in pure starlark. Then configured cuda toolchains as if we...
All 0.4 builds have some correctness problems, that is, many unit tests failed with weird error. So you better not use them. Aside from that, `DNN library initialization failed.` typically...
@lukasoyen Do you know anyway to query the info from cc_toolchain directly. This creates an unused args on each action run. And is quite wasteful... Better way should fuse it...