AMDMIGraphX

AMDMIGraphX copied to clipboard

wait() failing for the default stream 0

3

This PR fixes these two issues when using the default stream on the GPU: i.e. via `MIGRAPHX_ENABLE_NULL_STREAM=1`: 1. Trying to `verify` on a model, (resnet50v2 here) gives this exception: `context.hpp:365:...

lakhinderwalia

GroupQueryAttention: concat query, key, and value inputs into packed qkv

1

Target is 6.4.1 May require an additional changes to support accuracy

causten

Enable BF16 Types for DLM Models in Examples

6

richagadgil

Use fill_argument for literals that have the same value

8

pfultz2

Avoid dynamic memory allocation in kernel launch

1

pfultz2

Investigate hip types BF16 and FP8

1

New headers hipamd/include/hip/amd_detail/amd_hip_bf16.h and hipamd/include/hip/amd_detail/amd_hip_fp8.h should be in the latest ROCm release? Will we be incorporating these?

CharlieL7

Motivation: https://github.com/ROCm/MIOpen/pull/3515

kahmed10

turneram

Llama2 7b model C++ example

5

Implemented an example for https://huggingface.co/amd/Llama-2-7b-chat-hf-awq-int4-asym-gs128-onnx/tree/main Llama2 7b model with MIGraphX. Details about the example and description for running is available in README (https://github.com/ROCm/AMDMIGraphX/tree/htec/mgx-llama2-7b-example/examples/transformers/mgx_llama2)

ototh-htec

Resize Onnx parsing

5

The old logic stepped through the args list in order, then assumed the first float-type input was the "Scales". This led to an error in an obscure case where the...

bpickrel