Garrett Byrd
Garrett Byrd
### Description of errors There are some recent blog posts [1,2,3] that provide the following install instructions for `ROCm/bitsandbytes` for ROCm v6.x: ``` git clone --recurse https://github.com/ROCm/bitsandbytes.git cd bitsandbytes git...
I am currently reworking the flash attention benchmarking script provided [here](https://github.com/Dao-AILab/flash-attention/blob/main/benchmarks/benchmark_flash_attention.py). (This script was used to produce the data for the plots of the A100 and H100 benchmarks [provided in...
These benchmark plots and similar in the [readme of flash-attention](https://github.com/Dao-AILab/flash-attention?tab=readme-ov-file#speedup) are provided for A100 and H100. I am wanting to do some benchmarking on my own system and would like...
The install instructions for ROCm in the `README` recommend usage in a Docker container, but do not accurately list specific requirements otherwise. This impedes usage on non-containerized systems. Specific install...
### Suggestion Description Low-precision tensor operations are rising in usage as techniques such as LLM quantization gain traction in contemporary machine learning ecosystems. As such, it is appropriate to arm...
### Description of errors The support matrix discussed below needs to provide information on all ROCm libraries, not just the six provided. Currently the documentation for precision support ([here](https://rocm.docs.amd.com/en/latest/compatibility/precision-support.html#libraries-input-output-type-support)) aims...
There is a bit of conflicting information regarding the ownership of `bitsandbytes`. Recently, on GitHub, the ownership of the repository transferred from [Tim Dettmers](https://github.com/TimDettmers) to [bitsandbytes-foundation](https://github.com/bitsandbytes-foundation). Is the "bitsandbytes foundation"...
The README has some minor spots which could be cleaned up. Mainly utilizing `#>` instead of triple ticks for code blocks. E.g., #> make should be ```bash make ``` for...
### Problem Description Here is a quick cpp file that iterates over sockets and *should* list the names and devices on each socket. ```cpp #include #include #include int main() {...
### Problem Description If amdsmi is being marketed as "a successor to rocm_smi_lib and esmi_ib_library", why is E-SMI still required to get CPU information via amdsmi? Is this a wording...