[CI][PD] Add optional/nightly DSv4 Disaggregated eval (#42310)
Signed-off-by: NickLucche <nicolo.lucchesi@mistral.ai>
This commit is contained in:
@@ -203,3 +203,25 @@ steps:
|
|||||||
commands:
|
commands:
|
||||||
- bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
|
- bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
|
||||||
- bash v1/kv_connector/nixl_integration/run_multi_connector_edge_case_test.sh
|
- bash v1/kv_connector/nixl_integration/run_multi_connector_edge_case_test.sh
|
||||||
|
|
||||||
|
# P TP 4 - D DPEP 4 test case for DSv4-Flash
|
||||||
|
- label: DSv4-Flash Disaggregated DP EP
|
||||||
|
key: dsv4-flash-disaggregated
|
||||||
|
timeout_in_minutes: 60
|
||||||
|
device: h200
|
||||||
|
optional: true
|
||||||
|
working_dir: "/vllm-workspace/tests"
|
||||||
|
num_devices: 8
|
||||||
|
env:
|
||||||
|
ENABLE_HMA_FLAG: "1"
|
||||||
|
DP_EP: "1"
|
||||||
|
GPU_MEMORY_UTILIZATION: "0.85"
|
||||||
|
PREFILLER_TP_SIZE: "4"
|
||||||
|
DECODER_TP_SIZE: "4"
|
||||||
|
PREFILL_BLOCK_SIZE: "256"
|
||||||
|
DECODE_BLOCK_SIZE: "256"
|
||||||
|
MODEL_NAMES: "deepseek-ai/DeepSeek-V4-Flash"
|
||||||
|
VLLM_SERVE_EXTRA_ARGS: "--trust-remote-code,--kv-cache-dtype,fp8"
|
||||||
|
commands:
|
||||||
|
- bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
|
||||||
|
- bash v1/kv_connector/nixl_integration/run_accuracy_test.sh
|
||||||
|
|||||||
@@ -140,12 +140,13 @@ run_tests_for_model() {
|
|||||||
# Start prefill instances
|
# Start prefill instances
|
||||||
for i in $(seq 0 $((NUM_PREFILL_INSTANCES-1))); do
|
for i in $(seq 0 $((NUM_PREFILL_INSTANCES-1))); do
|
||||||
# Calculate GPU ID - we'll distribute across available GPUs
|
# Calculate GPU ID - we'll distribute across available GPUs
|
||||||
GPU_ID=$((i % $(get_num_gpus)))
|
GPU_START=$((i % $(get_num_gpus)))
|
||||||
NEXT_GPU=${GPU_ID}
|
GPU_ID=$GPU_START
|
||||||
|
NEXT_GPU=$GPU_START
|
||||||
# Reserve TP*PP GPUs for the prefiller (TP shards across PP stages).
|
# Reserve TP*PP GPUs for the prefiller (TP shards across PP stages).
|
||||||
PREFILLER_WORLD_SIZE=$((PREFILLER_TP_SIZE * PREFILLER_PP_SIZE))
|
PREFILLER_WORLD_SIZE=$((PREFILLER_TP_SIZE * PREFILLER_PP_SIZE))
|
||||||
for (( j=1; j < PREFILLER_WORLD_SIZE; j++ )); do
|
for (( j=1; j < PREFILLER_WORLD_SIZE; j++ )); do
|
||||||
NEXT_GPU=$(((GPU_ID + j) % $(get_num_gpus)))
|
NEXT_GPU=$(((GPU_START + j) % $(get_num_gpus)))
|
||||||
GPU_ID="${GPU_ID},${NEXT_GPU}"
|
GPU_ID="${GPU_ID},${NEXT_GPU}"
|
||||||
done
|
done
|
||||||
|
|
||||||
@@ -195,10 +196,12 @@ run_tests_for_model() {
|
|||||||
# Start decode instances
|
# Start decode instances
|
||||||
for i in $(seq 0 $((NUM_DECODE_INSTANCES-1))); do
|
for i in $(seq 0 $((NUM_DECODE_INSTANCES-1))); do
|
||||||
# Calculate GPU ID - we'll distribute across available GPUs, starting from after prefill GPUs
|
# Calculate GPU ID - we'll distribute across available GPUs, starting from after prefill GPUs
|
||||||
GPU_ID=$(((i + NEXT_GPU + 1) % $(get_num_gpus)))
|
DECODE_START=$(((i + NEXT_GPU + 1) % $(get_num_gpus)))
|
||||||
|
GPU_ID=$DECODE_START
|
||||||
|
NEXT_GPU=$DECODE_START
|
||||||
# If DECODER_TP_SIZE is more than 1
|
# If DECODER_TP_SIZE is more than 1
|
||||||
for (( j=1; j < DECODER_TP_SIZE; j++ )); do
|
for (( j=1; j < DECODER_TP_SIZE; j++ )); do
|
||||||
NEXT_GPU=$(((GPU_ID + j) % $(get_num_gpus)))
|
NEXT_GPU=$(((DECODE_START + j) % $(get_num_gpus)))
|
||||||
GPU_ID="${GPU_ID},${NEXT_GPU}"
|
GPU_ID="${GPU_ID},${NEXT_GPU}"
|
||||||
done
|
done
|
||||||
# Calculate port number (base port + instance number)
|
# Calculate port number (base port + instance number)
|
||||||
|
|||||||
@@ -26,6 +26,7 @@ EXPECTED_VALUES = {
|
|||||||
"Qwen/Qwen3.5-0.8B": 0.33,
|
"Qwen/Qwen3.5-0.8B": 0.33,
|
||||||
"google/gemma-4-E2B-it": 0.485,
|
"google/gemma-4-E2B-it": 0.485,
|
||||||
"ai21labs/AI21-Jamba2-3B": 0.74,
|
"ai21labs/AI21-Jamba2-3B": 0.74,
|
||||||
|
"deepseek-ai/DeepSeek-V4-Flash": 0.95,
|
||||||
}
|
}
|
||||||
|
|
||||||
SIMPLE_PROMPT = (
|
SIMPLE_PROMPT = (
|
||||||
|
|||||||
Reference in New Issue
Block a user