A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 12, 16:00 UTC
[Build/CI][DP/EP] Add QWen/Qwen3-30B-A3B-FP8 + EPLB tests to Nightly H100 and B200 (#29195)
Signed-off-by: Varun Sundar Rabindranath <vsundarr@redhat.com> Co-authored-by: Varun Sundar Rabindranath <vsundarr@redhat.com>
Varun Sundar Rabindranath committed
Nov 24, 2025 at 11:06 UTC
e924bbb4f4ac3258a71a18ac4c753c8056bc059f
3 files changed
+19
-6
.buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh
renamed
+7
-3
@@ -1,10 +1,12 @@
1
#!/usr/bin/env bash
2
set -euxo pipefail
3
4
-# args: [THRESHOLD] [NUM_QUESTIONS] [START_PORT]
4
+# args: [THRESHOLD] [NUM_QUESTIONS] [START_PORT] [DATA_PARALLEL_SIZE] [TENSOR_PARALLEL_SIZE]
5
THRESHOLD=${1:-0.8}
6
NUM_Q=${2:-1319}
7
PORT=${3:-8020}
8
+DATA_PARALLEL_SIZE=${4:-2}
9
+TENSOR_PARALLEL_SIZE=${5:-2}
10
OUT_DIR=${OUT_DIR:-/tmp/vllm-scheduled}
11
mkdir -p "${OUT_DIR}"
12
@@ -45,8 +47,10 @@ for BACK in "${BACKENDS[@]}"; do
47
VLLM_ALL2ALL_BACKEND=$BACK \
48
vllm serve "$MODEL" \
49
--enforce-eager \
48
- --tensor-parallel-size 2 \
49
- --data-parallel-size 2 \
50
+ --enable-eplb \
51
+ --eplb-config '{"window_size":10, "step_interval":100, "num_redundant_experts":0, "log_balancedness":true}' \
52
+ --tensor-parallel-size ${TENSOR_PARALLEL_SIZE} \
53
+ --data-parallel-size ${DATA_PARALLEL_SIZE} \
54
--enable-expert-parallel \
55
--trust-remote-code \
56
--max-model-len 2048 \
.buildkite/test-amd.yaml
+1
-1
@@ -1486,4 +1486,4 @@ steps:
1486
num_gpus: 4
1487
working_dir: "/vllm-workspace"
1488
commands:
1489
- - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep.sh 0.8 200 8020
1489
+ - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020
.buildkite/test-pipeline.yaml
+11
-2
@@ -1340,11 +1340,20 @@ steps:
1340
commands:
1341
- bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh 0.25 200 8010
1342
1343
-- label: Qwen3-30B-A3B-FP8-block Accuracy
1343
+- label: Qwen3-30B-A3B-FP8-block Accuracy (H100)
1344
timeout_in_minutes: 60
1345
gpu: h100
1346
optional: true
1347
num_gpus: 4
1348
working_dir: "/vllm-workspace"
1349
commands:
1350
- - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep.sh 0.8 200 8020
1350
+ - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020
1351
+
1352
+- label: Qwen3-30B-A3B-FP8-block Accuracy (B200)
1353
+ timeout_in_minutes: 60
1354
+ gpu: b200
1355
+ optional: true
1356
+ num_gpus: 2
1357
+ working_dir: "/vllm-workspace"
1358
+ commands:
1359
+ - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020 2 1
\ No newline at end of file