Skip to content

HIPBLAS / ROCm low prompt eval performance #7533

Closed
@etemiz

Description

@etemiz

I have two MI60's that don't perform well during prompt evaluation. What could be the reason?

Model Llama3-70B Q6:

llama_print_timings: prompt eval time = 3722.63 ms / 18 tokens ( 206.81 ms per token, 4.84 tokens per second)
llama_print_timings: eval time = 4274.60 ms / 35 runs ( 122.13 ms per token, 8.19 tokens per second)

compile:

HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" cmake -S . -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx906 -DCMAKE_BUILD_TYPE=Release && cmake --build build --config Release -- -j 16

ROCk module version 6.7.0

When using an 8B model Q8, it does this:

llama_print_timings: prompt eval time = 200.58 ms / 18 tokens ( 11.14 ms per token, 89.74 tokens per second)
llama_print_timings: eval time = 1819.74 ms / 94 runs ( 19.36 ms per token, 51.66 tokens per second)

I also did this hack #3772 (comment) which fixed the garbled output issue but I don't know if it is related.

Now I am wondering if it is a 6 bit quantization issue..

Thank you!

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions