From da7510c1778fe8188ea2504185c8d1dfddf6ce95 Mon Sep 17 00:00:00 2001 From: arakowsk-amd <182798202+arakowsk-amd@users.noreply.github.com> Date: Mon, 16 Dec 2024 13:41:17 -0800 Subject: [PATCH 1/4] adding `--enable-chunked-prefill false` --- docs/dev-docker/README.md | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-) diff --git a/docs/dev-docker/README.md b/docs/dev-docker/README.md index 11c0ef04fd8f7..8f39102ad280d 100644 --- a/docs/dev-docker/README.md +++ b/docs/dev-docker/README.md @@ -261,7 +261,8 @@ Benchmark Meta-Llama-3.1-405B FP8 with input 128 tokens, output 128 tokens and t --num-scheduler-steps 10 \ --tensor-parallel-size 8 \ --input-len 128 \ - --output-len 128 + --output-len 128 \ + --enable-chunked-prefill false If you want to run Meta-Llama-3.1-405B FP16, please run @@ -278,9 +279,7 @@ If you want to run Meta-Llama-3.1-405B FP16, please run --swap-space 16 \ --max-model-len 8192 \ --max-num-batched-tokens 65536 \ - --swap-space - --max-model-len - --gpu-memory-utilization 0.99 + --enable-chunked-prefill false For fp8 quantized Llama3.18B/70B models: From 08818103523a106b2766d8194fbe86829cc35e5a Mon Sep 17 00:00:00 2001 From: arakowsk-amd <182798202+arakowsk-amd@users.noreply.github.com> Date: Mon, 16 Dec 2024 13:48:49 -0800 Subject: [PATCH 2/4] Update README.md Co-authored-by: Reed Slobodin --- docs/dev-docker/README.md | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/docs/dev-docker/README.md b/docs/dev-docker/README.md index 8f39102ad280d..b3a295e914f3a 100644 --- a/docs/dev-docker/README.md +++ b/docs/dev-docker/README.md @@ -24,12 +24,12 @@ The performance data below was measured on a server with MI300X accelerators wit |---|---| | BKC | 24.13 | | ROCm | version ROCm 6.3 | -| amdgpu | build 2009461 | +| amdgpu | build 2041575 | | OS | Ubuntu 22.04 | -| Linux Kernel | 5.15.0-117-generic | -| BMCVersion | C2789.BC.0809.00 | -| BiosVersion | C2789.5.BS.1C11.AG.1 | -| CpldVersion | 02.02.00 | +| Linux Kernel | 5.15.0-70-generic | +| BMCVersion | C2789.BC.0811.00 | +| BiosVersion | C2789.5.BS.1C17.GN.1 | +| CpldVersion | 02.03.00 | | DCSCMCpldVersion | 02.02.00 | | CX7 | FW 28.40.1000 | | RAM | 1 TB | From c75a4c446c33f16244cb6346d3001c6c107042c1 Mon Sep 17 00:00:00 2001 From: arakowsk-amd <182798202+arakowsk-amd@users.noreply.github.com> Date: Tue, 17 Dec 2024 13:47:31 -0800 Subject: [PATCH 3/4] removing --enable-chunked-prefill --- docs/dev-docker/README.md | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-) diff --git a/docs/dev-docker/README.md b/docs/dev-docker/README.md index b3a295e914f3a..d8e1c9f68a07b 100644 --- a/docs/dev-docker/README.md +++ b/docs/dev-docker/README.md @@ -261,8 +261,7 @@ Benchmark Meta-Llama-3.1-405B FP8 with input 128 tokens, output 128 tokens and t --num-scheduler-steps 10 \ --tensor-parallel-size 8 \ --input-len 128 \ - --output-len 128 \ - --enable-chunked-prefill false + --output-len 128 If you want to run Meta-Llama-3.1-405B FP16, please run @@ -278,8 +277,7 @@ If you want to run Meta-Llama-3.1-405B FP16, please run --output-len 128 \ --swap-space 16 \ --max-model-len 8192 \ - --max-num-batched-tokens 65536 \ - --enable-chunked-prefill false + --max-num-batched-tokens 65536 For fp8 quantized Llama3.18B/70B models: From a1bd23fe663806555d0c867d5e9aa6ec0f0cbc0a Mon Sep 17 00:00:00 2001 From: arakowsk-amd <182798202+arakowsk-amd@users.noreply.github.com> Date: Tue, 17 Dec 2024 19:56:51 -0800 Subject: [PATCH 4/4] Updating Docker Manifest --- docs/dev-docker/README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/docs/dev-docker/README.md b/docs/dev-docker/README.md index d8e1c9f68a07b..7b17f93afc3b2 100644 --- a/docs/dev-docker/README.md +++ b/docs/dev-docker/README.md @@ -417,6 +417,6 @@ To reproduce the release docker: ``` git clone https://github.com/ROCm/vllm.git cd vllm -git checkout 2c60adc83981ada77a77b2adda78ef109d2e2e2b -docker build -f Dockerfile.rocm -t --build-arg BUILD_HIPBLASLT=1 --build-arg USE_CYTHON=1 . +git checkout 14b11f5ad74e04aee9f85f5f591a1f85fad6025d +docker build -f Dockerfile.rocm -t . ```