[CI][AMD] Configure MI300 tests for native execution without DinD (#48387)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
@@ -31,7 +31,7 @@ def test_python_error():
|
||||
x = torch.empty(alloc_bytes, dtype=torch.uint8, device=DEVICE_TYPE)
|
||||
tensors.append(x)
|
||||
# release the memory
|
||||
allocator.sleep()
|
||||
allocator.sleep(offload_tags=())
|
||||
|
||||
# allocate more memory than the total memory
|
||||
y = torch.empty(alloc_bytes, dtype=torch.uint8, device=DEVICE_TYPE)
|
||||
|
||||
@@ -5,16 +5,16 @@
|
||||
set -e
|
||||
|
||||
# ROCm CI runs this script inside `run-amd-test.sh` where /vllm-workspace often has no .git
|
||||
# (wheel artifact layout). The wrapper passes VLLM_STANDALONE_MERGE_BASE from the agent checkout.
|
||||
# (wheel artifact layout). The wrapper passes CI_STANDALONE_MERGE_BASE from the agent checkout.
|
||||
merge_base_commit=""
|
||||
if [[ -n "${VLLM_STANDALONE_MERGE_BASE:-}" ]]; then
|
||||
merge_base_commit="${VLLM_STANDALONE_MERGE_BASE}"
|
||||
if [[ -n "${CI_STANDALONE_MERGE_BASE:-}" ]]; then
|
||||
merge_base_commit="${CI_STANDALONE_MERGE_BASE}"
|
||||
elif merge_base_commit="$(git -C /vllm-workspace merge-base HEAD origin/main 2>/dev/null)"; then
|
||||
:
|
||||
elif merge_base_commit="$(git merge-base HEAD origin/main 2>/dev/null)"; then
|
||||
:
|
||||
else
|
||||
echo "ERROR: need a git checkout or VLLM_STANDALONE_MERGE_BASE to resolve wheels.vllm.ai commit." >&2
|
||||
echo "ERROR: need a git checkout or CI_STANDALONE_MERGE_BASE to resolve wheels.vllm.ai commit." >&2
|
||||
exit 1
|
||||
fi
|
||||
|
||||
|
||||
@@ -20,7 +20,8 @@
|
||||
# SD_MODEL - drafter model path
|
||||
# MODEL_NAME - target model (default: meta-llama/Llama-3.1-8B-Instruct)
|
||||
# NUM_SPEC_TOKENS - number of speculative tokens (default: 3)
|
||||
# GPU_MEMORY_UTILIZATION - (default: 0.7)
|
||||
# GPU_MEMORY_UTILIZATION - used when KV_CACHE_MEMORY_BYTES is unset (default: 0.7)
|
||||
# KV_CACHE_MEMORY_BYTES - optional KV cache size per server
|
||||
# ATTENTION_BACKEND - attention backend to use
|
||||
# Default: TRITON_ATTN on ROCm, FLASH_ATTN on NVIDIA
|
||||
# ROCm options: TRITON_ATTN, ROCM_ATTN, ROCM_AITER_FA,
|
||||
@@ -52,10 +53,17 @@ NUM_DECODE_INSTANCES=${NUM_DECODE_INSTANCES:-1}
|
||||
PREFILLER_TP_SIZE=${PREFILLER_TP_SIZE:-1}
|
||||
DECODER_TP_SIZE=${DECODER_TP_SIZE:-1}
|
||||
GPU_MEMORY_UTILIZATION=${GPU_MEMORY_UTILIZATION:-0.7}
|
||||
KV_CACHE_MEMORY_BYTES=${KV_CACHE_MEMORY_BYTES:-}
|
||||
BLOCK_SIZE=${BLOCK_SIZE:-16}
|
||||
SERVER_HOST="${SERVER_HOST:-127.0.0.1}"
|
||||
NIXL_SIDE_CHANNEL_HOST="${NIXL_SIDE_CHANNEL_HOST:-$SERVER_HOST}"
|
||||
|
||||
if [[ -n "$KV_CACHE_MEMORY_BYTES" ]]; then
|
||||
KV_CACHE_ARGS=(--kv-cache-memory-bytes "$KV_CACHE_MEMORY_BYTES")
|
||||
else
|
||||
KV_CACHE_ARGS=(--gpu-memory-utilization "$GPU_MEMORY_UTILIZATION")
|
||||
fi
|
||||
|
||||
# Resolve the repository root from the script location instead of `.git`.
|
||||
SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd -P)"
|
||||
GIT_ROOT="${GIT_ROOT:-$(cd -- "${SCRIPT_DIR}/../../../.." && pwd -P)}"
|
||||
@@ -209,6 +217,7 @@ run_test_for_device() {
|
||||
echo "SD model: ${SD_MODEL}"
|
||||
echo "Spec tokens: ${NUM_SPEC_TOKENS}"
|
||||
echo "KV buffer device: ${kv_device}"
|
||||
echo "KV cache memory: ${KV_CACHE_MEMORY_BYTES:-auto-sized}"
|
||||
echo "Attention backend: ${ATTENTION_BACKEND}"
|
||||
echo "GPU platform: ${GPU_PLATFORM}"
|
||||
echo "Server host: ${SERVER_HOST}"
|
||||
@@ -248,7 +257,7 @@ run_test_for_device() {
|
||||
--enforce-eager \
|
||||
--max-model-len $MAX_MODEL_LEN \
|
||||
--block-size ${BLOCK_SIZE} \
|
||||
--gpu-memory-utilization $GPU_MEMORY_UTILIZATION \
|
||||
"${KV_CACHE_ARGS[@]}" \
|
||||
--tensor-parallel-size $PREFILLER_TP_SIZE \
|
||||
--kv-transfer-config "$kv_config_p" \
|
||||
--speculative-config "$PREFILL_SPEC_CONFIG" \
|
||||
@@ -287,7 +296,7 @@ run_test_for_device() {
|
||||
--enforce-eager \
|
||||
--max-model-len $MAX_MODEL_LEN \
|
||||
--block-size ${BLOCK_SIZE} \
|
||||
--gpu-memory-utilization $GPU_MEMORY_UTILIZATION \
|
||||
"${KV_CACHE_ARGS[@]}" \
|
||||
--tensor-parallel-size $DECODER_TP_SIZE \
|
||||
--kv-transfer-config "$kv_config_d" \
|
||||
--speculative-config "$DECODE_SPEC_CONFIG" \
|
||||
|
||||
@@ -97,14 +97,16 @@ def _build_worker_with_dict_store(vllm_config, kv_cache_config, store):
|
||||
with (
|
||||
patch.dict(sys.modules, {"mooncake.store": fake_mooncake_store}),
|
||||
patch.object(mooncake_store_worker, "MooncakeStoreConfig") as MCfg,
|
||||
patch.object(mooncake_store_worker, "LookupKeyServer"),
|
||||
):
|
||||
sc = MCfg.load_from_env.return_value
|
||||
sc = MCfg.load_from_config.return_value
|
||||
sc.metadata_server = ""
|
||||
sc.global_segment_size = 1 << 20
|
||||
sc.local_buffer_size = 1 << 20
|
||||
sc.protocol = "tcp"
|
||||
sc.device_name = ""
|
||||
sc.master_server_address = ""
|
||||
sc.mode = "embedded"
|
||||
sc.enable_offload = False
|
||||
with (
|
||||
patch(
|
||||
|
||||
Reference in New Issue
Block a user