[Core][Hybrid allocator + kv connector 1/n] Enable hybrid allocator + KV cache connector (#25712)

Signed-off-by: KuntaiDu <kuntai@uchicago.edu>
Signed-off-by: Kuntai Du <kuntai@uchicago.edu>
This commit is contained in:
Kuntai Du
2025-10-24 23:34:18 -07:00
committed by GitHub
parent 56ed7609a9
commit b853540388
15 changed files with 113 additions and 18 deletions
+7
View File
@@ -899,6 +899,7 @@ def test_kv_connector_basic():
scheduler = create_scheduler(
enable_prefix_caching=True,
use_kv_connector=True,
disable_hybrid_kv_cache_manager=True,
)
NUM_TOTAL_BLOCKS = scheduler.kv_cache_manager.block_pool.get_num_free_blocks()
BLOCK_SIZE = scheduler.cache_config.block_size
@@ -1024,6 +1025,7 @@ def test_external_prefix_cache_metrics():
scheduler = create_scheduler(
enable_prefix_caching=False,
use_kv_connector=True,
disable_hybrid_kv_cache_manager=True,
)
# Mock connector to simulate a partial external cache hit
@@ -1088,6 +1090,7 @@ def test_kv_connector_unable_to_allocate():
use_kv_connector=True,
block_size=BLOCK_SIZE,
num_blocks=NUM_BLOCKS,
disable_hybrid_kv_cache_manager=True,
)
NUM_MATCHED_NEW_TOKENS = BLOCK_SIZE * 2
scheduler.connector.get_num_new_matched_tokens = Mock(name="method")
@@ -1171,6 +1174,7 @@ def test_kv_connector_handles_preemption():
use_kv_connector=True,
block_size=BLOCK_SIZE,
num_blocks=NUM_BLOCKS,
disable_hybrid_kv_cache_manager=True,
)
NUM_MATCHED_NEW_TOKENS = BLOCK_SIZE
@@ -1387,6 +1391,7 @@ def create_scheduler_with_priority(
block_size: int = 16,
max_model_len: int | None = None,
num_speculative_tokens: int | None = None,
disable_hybrid_kv_cache_manager: bool = False,
) -> Scheduler:
"""Create scheduler with priority policy enabled.
@@ -1411,6 +1416,7 @@ def create_scheduler_with_priority(
disable_chunked_mm_input=disable_chunked_mm_input,
enable_chunked_prefill=True,
policy="priority", # Enable priority scheduling
disable_hybrid_kv_cache_manager=disable_hybrid_kv_cache_manager,
)
model_config = ModelConfig(
model=model,
@@ -2018,6 +2024,7 @@ def test_priority_scheduling_preemption_and_resumption_when_out_of_kv():
num_blocks=5, # Can hold 64 tokens (first block is null)
block_size=16, # Standard block size
use_kv_connector=True,
disable_hybrid_kv_cache_manager=True,
)
# Create a request and schedule it
+2
View File
@@ -46,6 +46,7 @@ def create_scheduler(
num_speculative_tokens: int | None = None,
skip_tokenizer_init: bool = False,
async_scheduling: bool = False,
disable_hybrid_kv_cache_manager: bool = False,
) -> Scheduler | AsyncScheduler:
"""Create scheduler under test.
@@ -70,6 +71,7 @@ def create_scheduler(
disable_chunked_mm_input=disable_chunked_mm_input,
enable_chunked_prefill=True,
async_scheduling=async_scheduling,
disable_hybrid_kv_cache_manager=disable_hybrid_kv_cache_manager,
)
model_config = ModelConfig(
model=model,
@@ -136,6 +136,7 @@ run_tests_for_model() {
vllm serve $model_name \
--port $PORT \
--enforce-eager \
--disable-hybrid-kv-cache-manager \
--gpu-memory-utilization $GPU_MEMORY_UTILIZATION \
--tensor-parallel-size $PREFILLER_TP_SIZE \
--kv-transfer-config '$KV_CONFIG'"
@@ -178,6 +179,7 @@ run_tests_for_model() {
--port $PORT \
--enforce-eager \
--gpu-memory-utilization $GPU_MEMORY_UTILIZATION \
--disable-hybrid-kv-cache-manager \
--kv-transfer-config '$KV_CONFIG'"
# DP-EP attention mode
@@ -85,6 +85,7 @@ run_tests_for_model() {
--port $PREFILL_PORT \
--enforce-eager \
--gpu-memory-utilization 0.2 \
--disable-hybrid-kv-cache-manager \
--kv-transfer-config '$KV_CONFIG'"
if [ -n "$model_args" ]; then
@@ -103,6 +104,7 @@ run_tests_for_model() {
--port $DECODE_PORT \
--enforce-eager \
--gpu-memory-utilization 0.2 \
--disable-hybrid-kv-cache-manager \
--kv-transfer-config '$KV_CONFIG'"
if [ -n "$model_args" ]; then
@@ -114,6 +114,7 @@ def test_multi_shared_storage_connector_consistency():
enforce_eager=True,
gpu_memory_utilization=0.5,
kv_transfer_config=kv_transfer_config,
disable_hybrid_kv_cache_manager=True,
)
# Run generation - this should trigger saving KV cache
_ = llm.generate(PROMPTS, SAMPLING_PARAMS)
@@ -932,6 +932,7 @@ def test_abort_timeout_on_prefiller(monkeypatch, distributed_executor_backend):
"gpu_memory_utilization": 0.5,
"kv_transfer_config": kv_transfer_config,
"distributed_executor_backend": distributed_executor_backend,
"disable_hybrid_kv_cache_manager": True,
}
timeout = 6
@@ -132,6 +132,7 @@ def test_shared_storage_connector_hashes(tmp_path):
enforce_eager=True,
kv_transfer_config=kv_transfer_config,
limit_mm_per_prompt={"image": 2},
disable_hybrid_kv_cache_manager=True,
)
# don't put this import at the top level
+3
View File
@@ -91,6 +91,9 @@ def create_vllm_config(
max_num_batched_tokens=max_num_batched_tokens,
max_model_len=max_model_len,
enable_chunked_prefill=enable_chunked_prefill,
# Disable hybrid KV cache manager for testing
# Should be removed after we support hybrid KV cache manager-based testing.
disable_hybrid_kv_cache_manager=True,
)
model_config = ModelConfig(
model=model,
@@ -27,6 +27,7 @@ def test_cpu_offloading(cpu_block_size: int) -> None:
model="meta-llama/Llama-3.2-1B-Instruct",
gpu_memory_utilization=0.5,
kv_transfer_config=kv_transfer_config,
disable_hybrid_kv_cache_manager=True,
)
prompts = ["Hi " * 100]