Skip to content

[NVIDIA] update vllm b200 image. TODO: add logic for docker runner. #3

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
kimbochen merged 3 commits into from
Sep 3, 2025
Merged

[NVIDIA] update vllm b200 image. TODO: add logic for docker runner. #3

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
ec58c96
update vllm b200 image. TODO: add logic for docker runner.
kedarpotdar-nv Sep 2, 2025
9efe8ba
remove redundant flag
kedarpotdar-nv Sep 2, 2025
353351f
update dsr1 slurm
kedarpotdar-nv Sep 3, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion .github/workflows/70b-tmpl.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -73,7 +73,7 @@ jobs:
max-model-len: ${{ inputs.max-model-len }}
random-range-ratio: ${{ inputs.random-range-ratio }}
runner: b200
image: 'kedarpotdar147/vllm0.1:latest'
image: 'kedarpotdar147/vllm:05'
model: 'nvidia/Llama-3.3-70B-Instruct-FP8'
tp-list: '[1, 2, 4, 8]'
timeout: ${{ inputs.timeout }}
Expand Down
25 changes: 21 additions & 4 deletions benchmarks/70b_b200_slurm.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,12 +21,29 @@ hf download $MODEL
SERVER_LOG=$(mktemp /tmp/server-XXXXXX.log)
PORT=$(( 8888 + $PORT_OFFSET ))

#nccl update
pip uninstall -y nvidia-nccl-cu12
pip install nvidia-nccl-cu12==2.26.2.post1

pip uninstall -y flashinfer-python
git clone --recursive https://github.com/flashinfer-ai/flashinfer.git
git checkout 9720182476ede910698f8d783c29b2ec91cec023
cd flashinfer
pip install .

export VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB='{"2":32,"4":32,"8":8}'

FUSION_FLAG='{"pass_config":{"enable_fi_allreduce_fusion":true,"enable_attn_fusion":true,"enable_noop":true},"custom_ops":["+quant_fp8","+rms_norm"],"cudagraph_mode":"FULL_DECODE_ONLY","splitting_ops":[]}'

NO_PREFIX_CACHING_FLAG="--no-enable-prefix-caching"

@kimbochen kimbochen Sep 2, 2025

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This seems redundant



export TORCH_CUDA_ARCH_LIST="10.0"
vllm serve $MODEL --host 0.0.0.0 --port $PORT \
--trust-remote-code --quantization modelopt --kv-cache-dtype fp8 --gpu-memory-utilization 0.9 \
--pipeline-parallel-size 1 --tensor-parallel-size $TP --max-num-seqs $CONC --max-num-batched-tokens 8192 --max-model-len $MAX_MODEL_LEN \
--trust-remote-code --kv-cache-dtype fp8 --gpu-memory-utilization 0.9 \
--pipeline-parallel-size 1 --tensor-parallel-size $TP --max-num-seqs $CONC --max-num-batched-tokens 8192 --max-num-seqs 512 --max-model-len $MAX_MODEL_LEN \
--enable-chunked-prefill --async-scheduling --no-enable-prefix-caching \
--compilation-config '{"pass_config": {"enable_fi_allreduce_fusion": true}, "custom_ops": ["+rms_norm"], "level": 3}' \
--compilation-config ${FUSION_FLAG} \
--disable-log-requests > $SERVER_LOG 2>&1 &

set +x
Expand Down Expand Up @@ -54,4 +71,4 @@ python3 bench_serving/benchmark_serving.py \
--request-rate inf --ignore-eos \
--save-result --percentile-metrics 'ttft,tpot,itl,e2el' \
--result-dir /workspace/ \
--result-filename $RESULT_FILENAME.json
--result-filename $RESULT_FILENAME.json
2 changes: 1 addition & 1 deletion runners/launch_b200-nv.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,7 +5,7 @@ export PORT_OFFSET=${USER: -1}

MODEL_CODE="${1%%_*}"
PARTITION="dgx-b200"
SQUASH_FILE="/raid/image_${MODEL_CODE}_b200.sqsh"
SQUASH_FILE="/raid/image_${MODEL_CODE}_b200-2.sqsh"

salloc --partition=$PARTITION --gres=gpu:$TP --exclusive --time=180 --no-shell
JOB_ID=$(squeue -u $USER -h -o %A | head -n1)
Expand Down