[XPU] Support ZE_AFFINITY_MASK passthrough in xpu_disagg_acc_test (#47105)
Signed-off-by: zhenwei-intel <zhenwei.liu@intel.com>
This commit is contained in:
@@ -103,6 +103,31 @@ steps:
|
|||||||
pytest -v -s v1/kv_offload &&
|
pytest -v -s v1/kv_offload &&
|
||||||
pytest -v -s v1/kv_connector/unit/test_offloading_connector.py'
|
pytest -v -s v1/kv_connector/unit/test_offloading_connector.py'
|
||||||
|
|
||||||
|
- label: NixlConnector PD accuracy (2 GPUs)
|
||||||
|
timeout_in_minutes: 60
|
||||||
|
num_devices: 2
|
||||||
|
device: intel_gpu
|
||||||
|
agent_tags:
|
||||||
|
label: production
|
||||||
|
gpu: 2+
|
||||||
|
mem: 24+
|
||||||
|
no_plugin: true
|
||||||
|
working_dir: "."
|
||||||
|
env:
|
||||||
|
REGISTRY: "public.ecr.aws/q9t5s3a7"
|
||||||
|
REPO: "vllm-ci-test-repo"
|
||||||
|
VLLM_TEST_DEVICE: "xpu"
|
||||||
|
source_file_dependencies:
|
||||||
|
- vllm/distributed/kv_transfer/kv_connector/v1/nixl/
|
||||||
|
- vllm/v1/worker/kv_connector_model_runner_mixin.py
|
||||||
|
- tests/v1/kv_connector/nixl_integration/
|
||||||
|
- vllm/platforms/xpu.py
|
||||||
|
commands:
|
||||||
|
- >-
|
||||||
|
bash .buildkite/scripts/hardware_ci/run-intel-test.sh
|
||||||
|
'cd tests &&
|
||||||
|
bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh'
|
||||||
|
|
||||||
- label: Regression
|
- label: Regression
|
||||||
key: regression
|
key: regression
|
||||||
timeout_in_minutes: 30
|
timeout_in_minutes: 30
|
||||||
|
|||||||
@@ -68,7 +68,6 @@ steps:
|
|||||||
- >-
|
- >-
|
||||||
bash .buildkite/scripts/hardware_ci/run-intel-test.sh
|
bash .buildkite/scripts/hardware_ci/run-intel-test.sh
|
||||||
'cd tests &&
|
'cd tests &&
|
||||||
bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh &&
|
|
||||||
pytest -v -s v1/core --ignore=v1/core/test_reset_prefix_cache_e2e.py --ignore=v1/core/test_scheduler_e2e.py &&
|
pytest -v -s v1/core --ignore=v1/core/test_reset_prefix_cache_e2e.py --ignore=v1/core/test_scheduler_e2e.py &&
|
||||||
pytest -v -s v1/engine --ignore=v1/engine/test_output_processor.py &&
|
pytest -v -s v1/engine --ignore=v1/engine/test_output_processor.py &&
|
||||||
pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py -k "not test_topk_only and not test_topp_only and not test_topk_and_topp" &&
|
pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py -k "not test_topk_only and not test_topp_only and not test_topk_and_topp" &&
|
||||||
|
|||||||
@@ -21,27 +21,28 @@ DECODER_TP_SIZE=${DECODER_TP_SIZE:-1}
|
|||||||
KV_BUFFER_DEVICE=${KV_BUFFER_DEVICE:-"xpu"}
|
KV_BUFFER_DEVICE=${KV_BUFFER_DEVICE:-"xpu"}
|
||||||
GPU_MEMORY_UTILIZATION=${GPU_MEMORY_UTILIZATION:-0.8}
|
GPU_MEMORY_UTILIZATION=${GPU_MEMORY_UTILIZATION:-0.8}
|
||||||
|
|
||||||
generate_affinity_mask() {
|
# Parse available devices from ZE_AFFINITY_MASK (e.g. "2,3") or default to "0,1,...".
|
||||||
local count=$1
|
# compute_gpu_id indexes into this list to assign devices per-instance.
|
||||||
local start=${2:-0}
|
if [[ -n "${ZE_AFFINITY_MASK:-}" ]]; then
|
||||||
local mask=""
|
IFS=',' read -r -a AVAILABLE_DEVICES <<< "${ZE_AFFINITY_MASK}"
|
||||||
local i
|
else
|
||||||
|
# Default: devices 0 .. (PREFILLER_TP_SIZE + DECODER_TP_SIZE - 1)
|
||||||
for ((i=0; i<count; i++)); do
|
AVAILABLE_DEVICES=()
|
||||||
local device=$((start + i))
|
for ((i=0; i<PREFILLER_TP_SIZE + DECODER_TP_SIZE; i++)); do
|
||||||
if [[ -z "${mask}" ]]; then
|
AVAILABLE_DEVICES+=("$i")
|
||||||
mask="${device}"
|
|
||||||
else
|
|
||||||
mask="${mask},${device}"
|
|
||||||
fi
|
|
||||||
done
|
done
|
||||||
|
fi
|
||||||
|
|
||||||
echo "${mask}"
|
compute_gpu_id() {
|
||||||
|
local start=$1
|
||||||
|
local tp_size=$2
|
||||||
|
local gpu_id="${AVAILABLE_DEVICES[$start]}"
|
||||||
|
for (( j=1; j<tp_size; j++ )); do
|
||||||
|
gpu_id="${gpu_id},${AVAILABLE_DEVICES[$((start + j))]}"
|
||||||
|
done
|
||||||
|
echo "${gpu_id}"
|
||||||
}
|
}
|
||||||
|
|
||||||
PREFILLER_ZE_AFFINITY_MASK=${PREFILLER_ZE_AFFINITY_MASK:-$(generate_affinity_mask "${PREFILLER_TP_SIZE}" 0)}
|
|
||||||
DECODER_ZE_AFFINITY_MASK=${DECODER_ZE_AFFINITY_MASK:-$(generate_affinity_mask "${DECODER_TP_SIZE}" "${PREFILLER_TP_SIZE}")}
|
|
||||||
|
|
||||||
|
|
||||||
# execution env
|
# execution env
|
||||||
SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd -P)"
|
SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd -P)"
|
||||||
@@ -69,8 +70,11 @@ wait_for_server() {
|
|||||||
}
|
}
|
||||||
|
|
||||||
launch_baseline() {
|
launch_baseline() {
|
||||||
|
local BASELINE_GPU_ID
|
||||||
|
BASELINE_GPU_ID=$(compute_gpu_id 0 1)
|
||||||
|
|
||||||
BASELINE_BASE_CMD="
|
BASELINE_BASE_CMD="
|
||||||
ZE_AFFINITY_MASK=0 \
|
ZE_AFFINITY_MASK=$BASELINE_GPU_ID \
|
||||||
VLLM_WORKER_MULTIPROC_METHOD=spawn \
|
VLLM_WORKER_MULTIPROC_METHOD=spawn \
|
||||||
VLLM_ENABLE_V1_MULTIPROCESSING=1 vllm serve $MODEL_NAME \
|
VLLM_ENABLE_V1_MULTIPROCESSING=1 vllm serve $MODEL_NAME \
|
||||||
--host ${BASELINE_HOST} \
|
--host ${BASELINE_HOST} \
|
||||||
@@ -89,9 +93,14 @@ launch_baseline() {
|
|||||||
}
|
}
|
||||||
|
|
||||||
launch_pd() {
|
launch_pd() {
|
||||||
|
local PREFILL_GPU_ID
|
||||||
|
PREFILL_GPU_ID=$(compute_gpu_id 0 "${PREFILLER_TP_SIZE}")
|
||||||
|
local DECODE_GPU_ID
|
||||||
|
DECODE_GPU_ID=$(compute_gpu_id "${PREFILLER_TP_SIZE}" "${DECODER_TP_SIZE}")
|
||||||
|
|
||||||
PREFILL_BASE_CMD="
|
PREFILL_BASE_CMD="
|
||||||
ZE_AFFINITY_MASK=${PREFILLER_ZE_AFFINITY_MASK} \
|
ZE_AFFINITY_MASK=$PREFILL_GPU_ID \
|
||||||
VLLM_MULTIPROC_EXECUTE_MODEL_TIMEOUT_S=200 \
|
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=200 \
|
||||||
VLLM_NIXL_SIDE_CHANNEL_HOST=${PREFILL_HOST} \
|
VLLM_NIXL_SIDE_CHANNEL_HOST=${PREFILL_HOST} \
|
||||||
VLLM_NIXL_SIDE_CHANNEL_PORT=${PREFILL_NIXL_SIDE_PORT} \
|
VLLM_NIXL_SIDE_CHANNEL_PORT=${PREFILL_NIXL_SIDE_PORT} \
|
||||||
VLLM_WORKER_MULTIPROC_METHOD=spawn \
|
VLLM_WORKER_MULTIPROC_METHOD=spawn \
|
||||||
@@ -109,8 +118,8 @@ launch_pd() {
|
|||||||
|
|
||||||
|
|
||||||
DECODE_BASE_CMD="
|
DECODE_BASE_CMD="
|
||||||
ZE_AFFINITY_MASK=${DECODER_ZE_AFFINITY_MASK} \
|
ZE_AFFINITY_MASK=$DECODE_GPU_ID \
|
||||||
VLLM_MULTIPROC_EXECUTE_MODEL_TIMEOUT_S=200 \
|
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=200 \
|
||||||
VLLM_WORKER_MULTIPROC_METHOD=spawn \
|
VLLM_WORKER_MULTIPROC_METHOD=spawn \
|
||||||
VLLM_ENABLE_V1_MULTIPROCESSING=1 vllm serve $MODEL_NAME \
|
VLLM_ENABLE_V1_MULTIPROCESSING=1 vllm serve $MODEL_NAME \
|
||||||
--host ${DECODE_HOST} \
|
--host ${DECODE_HOST} \
|
||||||
|
|||||||
Reference in New Issue
Block a user