apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2026-04-23 00:17:25 +08:00
Code Issues Actions 21 Packages Projects Releases Wiki Activity

[Cherry-Pick][CI] Sync parallelism optimization from dev to 2.5 (#7085) (#7140)

Browse Source 
* [Cherry-Pick][CI] Sync parallelism optimization from dev to 2.5 (#7085)
This commit is contained in:
YuBaoku
2026-04-02 14:21:47 +08:00
committed by GitHub
parent 566699303c
commit 7648164f6e
17 changed files with 375 additions and 90 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/_base_test.yml
+4 -1
View File
@@ -182,7 +182,10 @@ jobs:
docker rm -f ${runner_name} || true
fi
docker run --rm --ipc=host --pid=host --net=host \
docker run --rm --net=host \
--shm-size=64g \
--sysctl kernel.msgmax=1048576 \
--sysctl kernel.msgmnb=268435456 \
--name ${runner_name} \
-v $(pwd):/workspace \
-w /workspace \
.github/workflows/_gpu_4cards_case_test.yml
+4 -1
View File
@@ -166,7 +166,10 @@ jobs:
docker rm -f ${runner_name} || true
fi
docker run --rm --ipc=host --net=host \
docker run --rm --net=host \
--shm-size=64g \
--sysctl kernel.msgmax=1048576 \
--sysctl kernel.msgmnb=268435456 \
--name ${runner_name} \
-v $(pwd):/workspace -w /workspace \
-v "${CACHE_DIR}/gitconfig:/etc/gitconfig:ro" \
.github/workflows/_logprob_test_linux.yml
+5 -1
View File
@@ -152,7 +152,11 @@ jobs:
echo "Removing stale container: ${runner_name}"
docker rm -f ${runner_name} || true
fi
docker run --rm --ipc=host --pid=host --net=host \
docker run --rm --net=host \
--shm-size=64g \
--sysctl kernel.msgmax=1048576 \
--sysctl kernel.msgmnb=268435456 \
--name ${runner_name} \
-v $(pwd):/workspace \
-w /workspace \
.github/workflows/_pre_ce_test.yml
+1
View File
@@ -163,6 +163,7 @@ jobs:
fi
docker run --rm --net=host \
--shm-size=64G \
--name ${runner_name} \
-v $(pwd):/workspace \
-w /workspace \
.github/workflows/_stable_test.yml
+1
View File
@@ -160,6 +160,7 @@ jobs:
fi
docker run --rm --net=host \
--shm-size=64G \
--name ${runner_name} \
-v $(pwd):/workspace \
-w /workspace \
.github/workflows/_unit_test_coverage.yml
+13
View File
@@ -47,6 +47,7 @@ jobs:
outputs:
all_cov_file_url: ${{ steps.cov_upload.outputs.all_cov_file_url }}
unittest_failed_url: ${{ steps.cov_upload.outputs.unittest_failed_url }}
unittest_logs_url: ${{ steps.cov_upload.outputs.unittest_logs_url }}
diff_cov_result_json_url: ${{ steps.cov_upload.outputs.diff_cov_result_json_url }}
steps:
- name: Code Prepare
@@ -173,6 +174,8 @@ jobs:
export RDMA_DEVICES=$(find /dev/infiniband/uverbs* -maxdepth 1 -not -type d | xargs -I{} echo '--device {}:{}')
docker run --rm --net=host \
--sysctl kernel.msgmax=1048576 \
--sysctl kernel.msgmnb=268435456 \
--name ${runner_name} \
--cap-add=SYS_PTRACE --cap-add=IPC_LOCK \
--shm-size=64G \
@@ -309,6 +312,15 @@ jobs:
echo "unittest_failed_url=${UNIT_TEST_RESULT_URL}" >> $GITHUB_ENV
fi
# Only upload logs when tests failed
unittest_logs_archive="unittest_logs.tar.gz"
if [ "$HAS_FAILED_TESTS" = true ]; then
python ${push_file} ${unittest_logs_archive} ${target_path}/UnitTestResult
UNIT_TEST_LOGS_URL=https://paddle-github-action.bj.bcebos.com/${target_path_stripped}/UnitTestResult/${unittest_logs_archive}
echo "unittest_logs_url=${UNIT_TEST_LOGS_URL}" >> $GITHUB_OUTPUT
echo "unittest_logs_url=${UNIT_TEST_LOGS_URL}" >> $GITHUB_ENV
fi
if [[ "$IS_PR" != "true" ]]; then
full_cov_file="full_coverage_report.txt"
full_cov_csv="full_coverage_report.csv"
@@ -345,6 +357,7 @@ jobs:
if [ -f "${filename}" ];then
echo "Failed test cases:"
cat "${filename}"
echo "unittest_logs_url=${unittest_logs_url}"
fi
exit "$TEST_EXIT_CODE"
fi
scripts/coverage_run.sh
+311 -66
View File
@@ -7,73 +7,95 @@ run_path=$( realpath "$DIR/../")
export COVERAGE_FILE=${COVERAGE_FILE:-$DIR/../coveragedata/.coverage}
export COVERAGE_RCFILE=${COVERAGE_RCFILE:-$DIR/../scripts/.coveragerc}
# ============================================================
# Classify tests into one of the following categories
# - multi_gpu: requires multiple GPUs / ports (run sequentially)
# - single_gpu: independent tests (can run in parallel)
# ============================================================
classify_tests() {
local test_file=$1
# Rule 1: distributed tests (explicit multi-GPU launch)
if [[ "$test_file" =~ tests/distributed/.*test_.*\.py ]]; then
echo "multi_gpu"
return
fi
failed_tests_file="failed_tests.log"
> "$failed_tests_file"
# Rule 2: e2e tests (usually involve service / ports)
if [[ "$test_file" =~ tests/e2e/.*test_.*\.py ]]; then
echo "multi_gpu"
return
fi
# Rule 3: model loader tests (allocate multiple GPUs)
if [[ "$test_file" =~ tests/model_loader/.*test_.*\.py ]]; then
echo "multi_gpu"
return
fi
##################################
# Run pytest, one file at a time
# Use pytest's --collect-only output to extract the actual test file paths (e.g., tests/.../test_*.py).
# Note: pytest may output lines like "ERROR tests/xxx/test_xxx.py::test_xxx ..." on collection failure,
# to avoid treating prefixes like "ERROR"/"FAILED"/"collecting" as filenames,
# we only keep the "tests/.../test_*.py" portion and discard everything else.
TEST_FILES=$(
python -m pytest --collect-only -q -c "${PYTEST_INI}" "${tests_path}" --rootdir="${run_path}" --disable-warnings 2>&1 \
| grep -E 'tests/.+\/test_.*\.py' \
| sed -E 's@.*(tests/[^: ]*test_[^: ]*\.py).*@\1@' \
| sort -u
)
# Rule 4: check file content for tensor_parallel_size=[234] or --tensor-parallel-size [234]
# or CUDA_VISIBLE_DEVICES="0,1"
# or PORT environment variables
if [ -f "$test_file" ]; then
if grep -q '"tensor_parallel_size".*[1234]\|--tensor-parallel-size.*[1234]\|tensor_parallel_size.*=[1234]\|CUDA_VISIBLE_DEVICES.*0.*1\|paddle\.distributed\.launch.*--gpus.*0.*1\|FD_API_PORT\|FLASK_PORT\|FD_ENGINE_QUEUE_PORT\|FD_METRICS_PORT\|FD_CACHE_QUEUE_PORT\|FD_ROUTER_PORT\|FD_CONNECTOR_PORT\|FD_RDMA_PORT' "$test_file" 2>/dev/null; then
echo "multi_gpu"
return
fi
fi
# ========== Single-GPU tests (no port required, can run in parallel) ==========
echo "single_gpu"
}
failed_pytest=0
success_pytest=0
# ============================================================
# Run Test With Logging
# ============================================================
run_test_with_logging() {
local test_file=$1
local log_prefix=$2
local status
# nullglob: if no match, the pattern expands to nothing
shopt -s nullglob
echo "Running pytest file: $test_file"
for file in $TEST_FILES; do
echo "Running pytest file: $file"
# Clean up previous logs
rm -rf "${run_path}"/log* || true
for f in "${run_path}"/*.log; do
[[ "$(basename "$f")" != "${failed_tests_file}" ]] && rm -f "$f"
done
# Create isolated log directory for this test to avoid race conditions
# Format: unittest_logs/<test_dir>/<test_file_base>/log
local test_rel_path="${test_file#tests/}"
local test_dir=$(dirname "$test_rel_path")
local test_name=$(basename "$test_file" .py)
local isolated_log_dir="${run_path}/unittest_logs/${test_dir}/${test_name}/log"
mkdir -p "$isolated_log_dir"
# Run pytest with coverage for the current file
# Set timeout to 600 seconds to avoid infinite loop
timeout 600 python -m coverage run -m pytest -c ${PYTEST_INI} "$file" -vv -s
# Set FD_LOG_DIR to isolate logs for each test
export FD_LOG_DIR="$isolated_log_dir"
# Run test
timeout 600 python -m coverage run -m pytest -c ${PYTEST_INI} "$test_file" -vv -s
status=$?
if [ "$status" -ne 0 ]; then
echo "$file" >> "$failed_tests_file"
failed_pytest=$((failed_pytest+1))
echo "$test_file" >> "$log_prefix"
echo ""
echo "==================== Dumping Logs ===================="
echo "==================== Test Failed: $test_file ===================="
for log_dir in "${run_path}"/log*; do
if [ -d "${log_dir}" ]; then
echo
echo ">>>> Processing log directory: ${log_dir}"
# Use isolated log directory for this test
if [ -d "$isolated_log_dir" ]; then
echo
echo ">>>> Processing log directory: ${isolated_log_dir}"
# print all workerlog.0
worker_logs=("${log_dir}"/workerlog.0)
if [ "${#worker_logs[@]}" -gt 0 ]; then
for worker_log in "${worker_logs[@]}"; do
if [ -f "${worker_log}" ]; then
echo "---------------- ${worker_log} (last 100 lines) ----------------"
tail -n 100 "${worker_log}" || true
echo "---------------------------------------------------------------"
fi
done
else
echo "No workerlog.0 found in ${log_dir}"
fi
# workerlog
worker_logs=("${isolated_log_dir}"/workerlog.0)
echo ">>> grep error in ${log_dir}"
grep -Rni --color=auto "error" "${log_dir}" || true
if [ -f "${worker_logs[0]}" ]; then
for worker_log in "${worker_logs[@]}"; do
[ -f "${worker_log}" ] || continue
echo "---------------- ${worker_log} (last 100 lines) ----------------"
tail -n 100 "${worker_log}" || true
echo "---------------------------------------------------------------"
done
fi
done
echo ">>> grep error in ${isolated_log_dir}"
grep -Rni --color=auto "error" "${isolated_log_dir}" || true
fi
# print all server logs
server_logs=("${run_path}"/*.log)
@@ -92,28 +114,251 @@ for file in $TEST_FILES; do
echo "No *.log files found"
fi
echo "======================================================"
else
success_pytest=$((success_pytest+1))
echo "======================================================="
fi
ps -ef | grep "${FD_CACHE_QUEUE_PORT}" | grep -v grep | awk '{print $2}' | xargs -r kill -9
ps -ef | grep "${FD_ENGINE_QUEUE_PORT}" | grep -v grep | awk '{print $2}' | xargs -r kill -9
done
shopt -u nullglob
##################################
# Summary
##################################
# Clean up port-related processes
if [ -n "$FD_CACHE_QUEUE_PORT" ]; then
ps -ef | grep "${FD_CACHE_QUEUE_PORT}" | grep -v grep | awk '{print $2}' | xargs -r kill -9 || true
fi
if [ -n "$FD_ENGINE_QUEUE_PORT" ]; then
ps -ef | grep "${FD_ENGINE_QUEUE_PORT}" | grep -v grep | awk '{print $2}' | xargs -r kill -9 || true
fi
# if passed, remove the isolated log directory and server logs
if [ "$status" -eq 0 ]; then
rm -rf "${isolated_log_dir}" || true
# Clean up server logs in run_path on pass
for f in "${run_path}"/*.log; do
[[ "$(basename "$f")" != "${failed_tests_file}" ]] && rm -f "$f" || true
done
fi
# Unset FD_LOG_DIR to avoid affecting next test
unset FD_LOG_DIR
return $status
}
# ============================================================
# Run a shard of tests on a dedicated GPU
# - one shard = one process = one GPU
# ============================================================
run_shard() {
local shard_name=$1
local gpu_id=$2
shift 2
local tests=("$@")
echo "===================================="
echo "Starting shard '${shard_name}' on GPU ${gpu_id}"
echo "Tests count: ${#tests[@]}"
echo "===================================="
# Set GPU
export CUDA_VISIBLE_DEVICES="$gpu_id"
export COVERAGE_FILE="${DIR}/../coveragedata/.coverage.${shard_name}"
# Failed log filename (no path, directly in project root)
local failed_log="${shard_name}_failed.txt"
rm -f "$failed_log"
> "$failed_log"
local success_count=0
local failed_count=0
for file in "${tests[@]}"; do
echo "[${shard_name}] Running: $file"
run_test_with_logging "$file" "$failed_log"
local status=$?
if [ "$status" -eq 0 ]; then
success_count=$((success_count + 1))
else
failed_count=$((failed_count + 1))
fi
done
unset COVERAGE_FILE
echo "===================================="
echo "Shard '${shard_name}' completed"
echo "Successful: $success_count"
echo "Failed: $failed_count"
echo "===================================="
unset CUDA_VISIBLE_DEVICES
return $failed_count
}
# ============================================================
# Main Flow
# ============================================================
failed_tests_file="failed_tests.log"
> "$failed_tests_file"
echo "===================================="
echo "Coverage Test Execution with Parallel Single-GPU Tests"
echo "===================================="
echo "Pytest total: $((failed_pytest + success_pytest))"
echo "Pytest successful: $success_pytest"
echo "Pytest failed: $failed_pytest"
# ============================================================
# Step 1: Collect & classify tests
# ============================================================
echo "Step 1: Collecting and classifying tests"
if [ "$failed_pytest" -ne 0 ]; then
ALL_TEST_FILES=$(
python -m pytest --collect-only -q -c "${PYTEST_INI}" "${tests_path}" --rootdir="${run_path}" --disable-warnings 2>&1 \
| grep -E 'tests/.+\/test_.*\.py' \
| sed -E 's@.*(tests/[^: ]*test_[^: ]*\.py).*@\1@' \
| sort -u
)
if [ -z "$ALL_TEST_FILES" ]; then
echo "ERROR: No test files found!"
exit 1
fi
MULTI_GPU_TESTS=()
SINGLE_GPU_TESTS=()
TOTAL_TESTS=0
for file in $ALL_TEST_FILES; do
TOTAL_TESTS=$((TOTAL_TESTS + 1))
test_type=$(classify_tests "$file")
case "$test_type" in
"multi_gpu")
MULTI_GPU_TESTS+=("$file")
;;
"single_gpu")
SINGLE_GPU_TESTS+=("$file")
;;
esac
done
echo "Multi-GPU tests: ${#MULTI_GPU_TESTS[@]}"
echo "Single-GPU tests: ${#SINGLE_GPU_TESTS[@]}"
echo "Total tests: $TOTAL_TESTS"
# ============================================================
# Step 2: Run multi-GPU tests (sequential)
# ============================================================
echo "Step 2: Running multi-GPU tests"
if [ ${#MULTI_GPU_TESTS[@]} -gt 0 ]; then
for file in "${MULTI_GPU_TESTS[@]}"; do
run_test_with_logging "$file" "$failed_tests_file"
done
else
echo "No multi-GPU tests to run."
fi
# ============================================================
# Step 3: Run single-GPU tests (parallel shards)
# ============================================================
echo "Step 3: Running single-GPU tests in parallel"
if [ ${#SINGLE_GPU_TESTS[@]} -gt 0 ]; then
# Split single-GPU tests into 2 shards (1 per GPU)
TOTAL=${#SINGLE_GPU_TESTS[@]}
HALF=$(( TOTAL / 2 ))
SHARD_1=("${SINGLE_GPU_TESTS[@]:0:$HALF}")
SHARD_2=("${SINGLE_GPU_TESTS[@]:$HALF}")
echo "Shard 1: ${#SHARD_1[@]} tests on GPU 0"
echo "Shard 2: ${#SHARD_2[@]} tests on GPU 1"
# Run in parallel (1 process per GPU)
run_shard "shard1" 0 "${SHARD_1[@]}" &
PID1=$!
run_shard "shard2" 1 "${SHARD_2[@]}" &
PID2=$!
# Wait for all shards to complete
wait $PID1
EXIT_CODE1=$?
wait $PID2
EXIT_CODE2=$?
# Merge shard failed logs to main failed log
for shard in shard1 shard2; do
if [ -f "${shard}_failed.txt" ]; then
cat "${shard}_failed.txt" >> "$failed_tests_file"
rm -f "${shard}_failed.txt"
fi
done
echo ""
echo "===================================="
echo "Parallel execution completed"
echo "Shard 1 exit code: $EXIT_CODE1"
echo "Shard 2 exit code: $EXIT_CODE2"
echo "===================================="
else
echo "No single-GPU tests to run."
fi
# ============================================================
# Step 4: Summary
# ============================================================
echo "Step 4: Summary"
# Count failed tests
if [ -f "$failed_tests_file" ]; then
failed_count=$(wc -l < "$failed_tests_file" | tr -d ' ')
else
failed_count=0
fi
success_count=$((TOTAL_TESTS - failed_count))
echo "Pytest total: $TOTAL_TESTS"
echo "Pytest successful: $success_count"
echo "Pytest failed: $failed_count"
echo "===================================="
# Exit with error and package logs if there were failures
if [ "$failed_count" -ne 0 ]; then
echo "Failed test cases are listed in $failed_tests_file"
cat "$failed_tests_file"
# clean the empty directories
if [ -d "${run_path}/unittest_logs" ]; then
echo "Cleaning empty directories..."
# remove console_error.log files (cleanup logs from stopped processes)
find "${run_path}/unittest_logs" -name "console_error.log*" -delete || true
# perform multi-round clean until no more empty directories are found
while true; do
before=$(find "${run_path}/unittest_logs" -type d | wc -l)
find "${run_path}/unittest_logs" -mindepth 1 -type d -empty -delete || true
after=$(find "${run_path}/unittest_logs" -type d | wc -l)
[ "$before" -eq "$after" ] && break
done
fi
# Only package logs when there are failures
echo "===================================="
echo "Step 5: Packaging logs (only on failure)"
echo "===================================="
if [ -d "${run_path}/unittest_logs" ]; then
tar -czf "${run_path}/unittest_logs.tar.gz" -C "${run_path}" unittest_logs
echo "Logs packaged to: ${run_path}/unittest_logs.tar.gz"
ls -lh "${run_path}/unittest_logs.tar.gz"
else
echo "No unittest_logs directory found."
fi
echo "===================================="
exit 8
fi
echo "All tests passed!"
exit 0
tests/e2e/test_EB_Lite_serving.py
+2 -1
View File
@@ -1389,7 +1389,8 @@ def test_streaming_chat_finish_reason(openai_client):
def test_profile_reset_block_num():
"""测试profile reset_block_num功能，与baseline diff不能超过5%"""
log_file = "./log/config.log"
log_dir = os.getenv("FD_LOG_DIR", "log")
log_file = os.path.join(log_dir, "config.log")
baseline = 31446
if not os.path.exists(log_file):
tests/e2e/test_EB_VL_Lite_serving.py
+2 -1
View File
@@ -734,7 +734,8 @@ def test_chat_with_response_max_tokens(openai_client):
def test_profile_reset_block_num():
"""测试profile reset_block_num功能，与baseline diff不能超过5%"""
log_file = "./log/config.log"
log_dir = os.getenv("FD_LOG_DIR", "log")
log_file = os.path.join(log_dir, "config.log")
baseline = 40000
if not os.path.exists(log_file):
tests/e2e/test_Qwen2-7B-Instruct_serving.py
+2 -1
View File
@@ -612,7 +612,8 @@ def test_streaming(openai_client, capsys):
def test_profile_reset_block_num():
"""测试profile reset_block_num功能，与baseline diff不能超过5%"""
log_file = "./log/config.log"
log_dir = os.getenv("FD_LOG_DIR", "log")
log_file = os.path.join(log_dir, "config.log")
baseline = 32562
if not os.path.exists(log_file):
tests/e2e/test_Qwen2_5_VL_serving.py
+2 -1
View File
@@ -430,7 +430,8 @@ def test_streaming_chat_with_return_token_ids(openai_client, capsys):
def test_profile_reset_block_num():
"""测试profile reset_block_num功能，与baseline diff不能超过15%"""
log_file = "./log/config.log"
log_dir = os.getenv("FD_LOG_DIR", "log")
log_file = os.path.join(log_dir, "config.log")
baseline = 30000
if not os.path.exists(log_file):
tests/e2e/test_ernie_03b_pd_router_v0_ipc.py
+5 -3
View File
@@ -81,10 +81,12 @@ def setup_and_run_server():
model_path = "baidu/ERNIE-4.5-0.3B-Paddle"
print(f"model_path: {model_path}")
base_log_dir = os.getenv("FD_LOG_DIR", "log")
# router
print("start router...")
env_router = os.environ.copy()
env_router["FD_LOG_DIR"] = "log_router"
env_router["FD_LOG_DIR"] = os.path.join(base_log_dir, "log_router")
router_log_path = "router.log"
router_cmd = [
@@ -110,7 +112,7 @@ def setup_and_run_server():
env_prefill = os.environ.copy()
env_prefill["CUDA_VISIBLE_DEVICES"] = "0"
env_prefill["ENABLE_V1_KVCACHE_SCHEDULER"] = "0"
env_prefill["FD_LOG_DIR"] = "log_prefill"
env_prefill["FD_LOG_DIR"] = os.path.join(base_log_dir, "log_prefill")
prefill_log_path = "server_prefill.log"
prefill_cmd = [
sys.executable,
@@ -160,7 +162,7 @@ def setup_and_run_server():
env_decode = os.environ.copy()
env_decode["CUDA_VISIBLE_DEVICES"] = "1"
env_decode["ENABLE_V1_KVCACHE_SCHEDULER"] = "0"
env_decode["FD_LOG_DIR"] = "log_decode"
env_decode["FD_LOG_DIR"] = os.path.join(base_log_dir, "log_decode")
decode_log_path = "server_decode.log"
decode_cmd = [
sys.executable,
tests/e2e/test_ernie_03b_pd_router_v1_ipc.py
+5 -3
View File
@@ -81,10 +81,12 @@ def setup_and_run_server():
model_path = "baidu/ERNIE-4.5-0.3B-Paddle"
print(f"model_path: {model_path}")
base_log_dir = os.getenv("FD_LOG_DIR", "log")
# router
print("start router...")
env_router = os.environ.copy()
env_router["FD_LOG_DIR"] = "log_router"
env_router["FD_LOG_DIR"] = os.path.join(base_log_dir, "log_router")
router_log_path = "router.log"
router_cmd = [
@@ -110,7 +112,7 @@ def setup_and_run_server():
env_prefill = os.environ.copy()
env_prefill["CUDA_VISIBLE_DEVICES"] = "0"
env_prefill["ENABLE_V1_KVCACHE_SCHEDULER"] = "1"
env_prefill["FD_LOG_DIR"] = "log_prefill"
env_prefill["FD_LOG_DIR"] = os.path.join(base_log_dir, "log_prefill")
prefill_log_path = "prefill.log"
prefill_cmd = [
sys.executable,
@@ -160,7 +162,7 @@ def setup_and_run_server():
env_decode = os.environ.copy()
env_decode["CUDA_VISIBLE_DEVICES"] = "1"
env_decode["ENABLE_V1_KVCACHE_SCHEDULER"] = "1"
env_decode["FD_LOG_DIR"] = "log_decode"
env_decode["FD_LOG_DIR"] = os.path.join(base_log_dir, "log_decode")
decode_log_path = "decode.log"
decode_cmd = [
sys.executable,
tests/e2e/test_ernie_03b_pd_router_v1_rdma_tp1.py
+5 -3
View File
@@ -84,6 +84,8 @@ def setup_and_run_server():
model_path = "baidu/ERNIE-4.5-0.3B-Paddle"
print(f"model_path: {model_path}")
base_log_dir = os.getenv("FD_LOG_DIR", "log")
# get rdma nics
current_dir = os.path.dirname(os.path.abspath(__file__))
shell_path = os.path.join(current_dir, "utils/get_rdma_nics.sh")
@@ -94,7 +96,7 @@ def setup_and_run_server():
# router
print("start router...")
env_router = os.environ.copy()
env_router["FD_LOG_DIR"] = "log_router"
env_router["FD_LOG_DIR"] = os.path.join(base_log_dir, "log_router")
router_log_path = "router.log"
router_cmd = [
@@ -119,7 +121,7 @@ def setup_and_run_server():
print("start prefill...")
env_prefill = os.environ.copy()
env_prefill["CUDA_VISIBLE_DEVICES"] = "0"
env_prefill["FD_LOG_DIR"] = "log_prefill"
env_prefill["FD_LOG_DIR"] = os.path.join(base_log_dir, "log_prefill")
env_prefill["KVCACHE_RDMA_NICS"] = rdma_nics
prefill_log_path = "prefill.log"
@@ -166,7 +168,7 @@ def setup_and_run_server():
print("start decode...")
env_decode = os.environ.copy()
env_decode["CUDA_VISIBLE_DEVICES"] = "1"
env_decode["FD_LOG_DIR"] = "log_decode"
env_decode["FD_LOG_DIR"] = os.path.join(base_log_dir, "log_decode")
env_decode["KVCACHE_RDMA_NICS"] = rdma_nics
decode_log_path = "decode.log"
tests/e2e/test_ernie_03b_pd_router_v1_rdma_tp2.py
+5 -3
View File
@@ -86,6 +86,8 @@ def setup_and_run_server():
model_path = "baidu/ERNIE-4.5-0.3B-Paddle"
print(f"model_path: {model_path}")
base_log_dir = os.getenv("FD_LOG_DIR", "log")
# get rdma nics
current_dir = os.path.dirname(os.path.abspath(__file__))
shell_path = os.path.join(current_dir, "utils/get_rdma_nics.sh")
@@ -96,7 +98,7 @@ def setup_and_run_server():
# router
print("start router...")
env_router = os.environ.copy()
env_router["FD_LOG_DIR"] = "log_router"
env_router["FD_LOG_DIR"] = os.path.join(base_log_dir, "log_router")
router_log_path = "router.log"
router_cmd = [
@@ -121,7 +123,7 @@ def setup_and_run_server():
print("start prefill...")
env_prefill = os.environ.copy()
env_prefill["CUDA_VISIBLE_DEVICES"] = "0,1"
env_prefill["FD_LOG_DIR"] = "log_prefill"
env_prefill["FD_LOG_DIR"] = os.path.join(base_log_dir, "log_prefill")
env_prefill["KVCACHE_RDMA_NICS"] = rdma_nics
prefill_log_path = "prefill.log"
@@ -170,7 +172,7 @@ def setup_and_run_server():
print("start decode...")
env_decode = os.environ.copy()
env_decode["CUDA_VISIBLE_DEVICES"] = "1"
env_decode["FD_LOG_DIR"] = "log_decode"
env_decode["FD_LOG_DIR"] = os.path.join(base_log_dir, "log_decode")
env_decode["KVCACHE_RDMA_NICS"] = rdma_nics
decode_log_path = "decode.log"
tests/e2e/test_ernie_03b_router.py
+6 -4
View File
@@ -97,10 +97,12 @@ def setup_and_run_server():
model_path = "baidu/ERNIE-4.5-0.3B-Paddle"
print(f"model_path: {model_path}")
base_log_dir = os.getenv("FD_LOG_DIR", "log")
# router
print("start router...")
env_router = os.environ.copy()
env_router["FD_LOG_DIR"] = "log_router"
env_router["FD_LOG_DIR"] = os.path.join(base_log_dir, "log_router")
router_log_path = "router.log"
router_cmd = [
@@ -121,11 +123,11 @@ def setup_and_run_server():
)
# server0
print("start server0...")
print("start server 0...")
env_server_0 = os.environ.copy()
env_server_0["CUDA_VISIBLE_DEVICES"] = "0"
env_server_0["ENABLE_V1_KVCACHE_SCHEDULER"] = "0"
env_server_0["FD_LOG_DIR"] = "log_server_0"
env_server_0["FD_LOG_DIR"] = os.path.join(base_log_dir, "log_server_0")
env_server_0["INFERENCE_MSG_QUEUE_ID"] = str(FD_API_PORT)
log_path = "server_0.log"
cmd = [
@@ -171,7 +173,7 @@ def setup_and_run_server():
env_server_1["CUDA_VISIBLE_DEVICES"] = "1"
env_server_1["ENABLE_V1_KVCACHE_SCHEDULER"] = "0"
env_server_1["INFERENCE_MSG_QUEUE_ID"] = str(FD_API_PORT + 1)
env_server_1["FD_LOG_DIR"] = "log_server_1"
env_server_1["FD_LOG_DIR"] = os.path.join(base_log_dir, "log_server_1")
log_path = "server_1.log"
cmd = [
sys.executable,
tests/entrypoints/openai/test_multi_api_server.py
+2 -1
View File
@@ -99,7 +99,8 @@ class TestMultiApiServer(unittest.TestCase):
# Verify environment variables are set correctly
first_call_kwargs = mock_popen.call_args_list[0][1]
self.assertIn("env", first_call_kwargs)
self.assertEqual(first_call_kwargs["env"]["FD_LOG_DIR"], "log/log_0")
log_dir = os.getenv("FD_LOG_DIR", "log")
self.assertEqual(first_call_kwargs["env"]["FD_LOG_DIR"], os.path.join(log_dir, "log_0"))
@patch("fastdeploy.entrypoints.openai.multi_api_server.is_port_available")
def test_check_param_success(self, mock_is_port_available):