A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 13, 19:30 UTC
[Hybrid allocator + kv connector] revert connector test changes related to hybrid allocator (#28011)
Signed-off-by: KuntaiDu <kuntai@uchicago.edu>
Kuntai Du committed
Nov 5, 2025 at 02:36 UTC
86dca07d9bfd9f892f820d2d9a45b102d1f4a096
8 files changed
-19
tests/v1/core/test_scheduler.py
-7
@@ -891,7 +891,6 @@ def test_kv_connector_basic():
891
scheduler = create_scheduler(
892
enable_prefix_caching=True,
893
use_kv_connector=True,
894
- disable_hybrid_kv_cache_manager=True,
894
)
895
NUM_TOTAL_BLOCKS = scheduler.kv_cache_manager.block_pool.get_num_free_blocks()
896
BLOCK_SIZE = scheduler.cache_config.block_size
@@ -1017,7 +1016,6 @@ def test_external_prefix_cache_metrics():
1016
scheduler = create_scheduler(
1017
enable_prefix_caching=False,
1018
use_kv_connector=True,
1020
- disable_hybrid_kv_cache_manager=True,
1019
)
1020
1021
# Mock connector to simulate a partial external cache hit
@@ -1082,7 +1080,6 @@ def test_kv_connector_unable_to_allocate():
1080
use_kv_connector=True,
1081
block_size=BLOCK_SIZE,
1082
num_blocks=NUM_BLOCKS,
1085
- disable_hybrid_kv_cache_manager=True,
1083
)
1084
NUM_MATCHED_NEW_TOKENS = BLOCK_SIZE * 2
1085
scheduler.connector.get_num_new_matched_tokens = Mock(name="method")
@@ -1166,7 +1163,6 @@ def test_kv_connector_handles_preemption():
1163
use_kv_connector=True,
1164
block_size=BLOCK_SIZE,
1165
num_blocks=NUM_BLOCKS,
1169
- disable_hybrid_kv_cache_manager=True,
1166
)
1167
1168
NUM_MATCHED_NEW_TOKENS = BLOCK_SIZE
@@ -1383,7 +1379,6 @@ def create_scheduler_with_priority(
1379
block_size: int = 16,
1380
max_model_len: int | None = None,
1381
num_speculative_tokens: int | None = None,
1386
- disable_hybrid_kv_cache_manager: bool = False,
1382
) -> Scheduler:
1383
"""Create scheduler with priority policy enabled.
1384
@@ -1408,7 +1403,6 @@ def create_scheduler_with_priority(
1403
disable_chunked_mm_input=disable_chunked_mm_input,
1404
enable_chunked_prefill=True,
1405
policy="priority", # Enable priority scheduling
1411
- disable_hybrid_kv_cache_manager=disable_hybrid_kv_cache_manager,
1406
)
1407
model_config = ModelConfig(
1408
model=model,
@@ -2015,7 +2009,6 @@ def test_priority_scheduling_preemption_and_resumption_when_out_of_kv():
2009
num_blocks=5, # Can hold 64 tokens (first block is null)
2010
block_size=16, # Standard block size
2011
use_kv_connector=True,
2018
- disable_hybrid_kv_cache_manager=True,
2012
)
2013
2014
# Create a request and schedule it
tests/v1/core/utils.py
-2
@@ -46,7 +46,6 @@ def create_scheduler(
46
num_speculative_tokens: int | None = None,
47
skip_tokenizer_init: bool = False,
48
async_scheduling: bool = False,
49
- disable_hybrid_kv_cache_manager: bool = False,
49
) -> Scheduler | AsyncScheduler:
50
"""Create scheduler under test.
51
@@ -71,7 +70,6 @@ def create_scheduler(
70
disable_chunked_mm_input=disable_chunked_mm_input,
71
enable_chunked_prefill=True,
72
async_scheduling=async_scheduling,
74
- disable_hybrid_kv_cache_manager=disable_hybrid_kv_cache_manager,
73
)
74
model_config = ModelConfig(
75
model=model,
tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh
-2
@@ -136,7 +136,6 @@ run_tests_for_model() {
136
vllm serve $model_name \
137
--port $PORT \
138
--enforce-eager \
139
- --disable-hybrid-kv-cache-manager \
139
--gpu-memory-utilization $GPU_MEMORY_UTILIZATION \
140
--tensor-parallel-size $PREFILLER_TP_SIZE \
141
--kv-transfer-config '$KV_CONFIG'"
@@ -179,7 +178,6 @@ run_tests_for_model() {
178
--port $PORT \
179
--enforce-eager \
180
--gpu-memory-utilization $GPU_MEMORY_UTILIZATION \
182
- --disable-hybrid-kv-cache-manager \
181
--kv-transfer-config '$KV_CONFIG'"
182
183
# DP-EP attention mode
tests/v1/kv_connector/nixl_integration/run_edge_case_test.sh
-2
@@ -85,7 +85,6 @@ run_tests_for_model() {
85
--port $PREFILL_PORT \
86
--enforce-eager \
87
--gpu-memory-utilization 0.2 \
88
- --disable-hybrid-kv-cache-manager \
88
--kv-transfer-config '$KV_CONFIG'"
89
90
if [ -n "$model_args" ]; then
@@ -104,7 +103,6 @@ run_tests_for_model() {
103
--port $DECODE_PORT \
104
--enforce-eager \
105
--gpu-memory-utilization 0.2 \
107
- --disable-hybrid-kv-cache-manager \
106
--kv-transfer-config '$KV_CONFIG'"
107
108
if [ -n "$model_args" ]; then
tests/v1/kv_connector/unit/test_multi_connector.py
-1
@@ -114,7 +114,6 @@ def test_multi_shared_storage_connector_consistency():
114
enforce_eager=True,
115
gpu_memory_utilization=0.5,
116
kv_transfer_config=kv_transfer_config,
117
- disable_hybrid_kv_cache_manager=True,
117
)
118
# Run generation - this should trigger saving KV cache
119
_ = llm.generate(PROMPTS, SAMPLING_PARAMS)
tests/v1/kv_connector/unit/test_nixl_connector.py
-1
@@ -1020,7 +1020,6 @@ def test_abort_timeout_on_prefiller(monkeypatch, distributed_executor_backend):
1020
"gpu_memory_utilization": 0.5,
1021
"kv_transfer_config": kv_transfer_config,
1022
"distributed_executor_backend": distributed_executor_backend,
1023
- "disable_hybrid_kv_cache_manager": True,
1023
}
1024
1025
timeout = 6
tests/v1/kv_connector/unit/test_shared_storage_connector.py
-1
@@ -132,7 +132,6 @@ def test_shared_storage_connector_hashes(tmp_path):
132
enforce_eager=True,
133
kv_transfer_config=kv_transfer_config,
134
limit_mm_per_prompt={"image": 2},
135
- disable_hybrid_kv_cache_manager=True,
135
)
136
137
# don't put this import at the top level
tests/v1/kv_connector/unit/utils.py
-3
@@ -91,9 +91,6 @@ def create_vllm_config(
91
max_num_batched_tokens=max_num_batched_tokens,
92
max_model_len=max_model_len,
93
enable_chunked_prefill=enable_chunked_prefill,
94
- # Disable hybrid KV cache manager for testing
95
- # Should be removed after we support hybrid KV cache manager-based testing.
96
- disable_hybrid_kv_cache_manager=True,
94
)
95
model_config = ModelConfig(
96
model=model,