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 14, 05:30 UTC
[bugfix][DCP] fix block_size of hash in DCP prefix caching (#26296)
Signed-off-by: Chen Zhang <zhangch99@outlook.com>
Chen Zhang committed
Oct 10, 2025 at 18:02 UTC
606b00e80f32698544b117f6c8fbf9265eae61dc
5 files changed
+12
-10
tests/v1/core/test_scheduler.py
+1
@@ -1411,6 +1411,7 @@ def create_scheduler_with_priority(
1411
kv_cache_config=kv_cache_config,
1412
log_stats=True,
1413
structured_output_manager=StructuredOutputManager(vllm_config),
1414
+ block_size=block_size,
1415
)
1416
1417
tests/v1/core/utils.py
+1
@@ -129,6 +129,7 @@ def create_scheduler(
129
return scheduler_cls(
130
vllm_config=vllm_config,
131
kv_cache_config=kv_cache_config,
132
+ block_size=block_size,
133
log_stats=True,
134
structured_output_manager=StructuredOutputManager(vllm_config),
135
)
tests/v1/kv_connector/unit/utils.py
+1
@@ -138,6 +138,7 @@ def create_scheduler(
138
kv_cache_config=kv_cache_config,
139
log_stats=True,
140
structured_output_manager=StructuredOutputManager(vllm_config),
141
+ block_size=block_size,
142
)
143
144
vllm/v1/core/sched/scheduler.py
+2
-8
@@ -45,6 +45,7 @@ class Scheduler(SchedulerInterface):
45
vllm_config: VllmConfig,
46
kv_cache_config: KVCacheConfig,
47
structured_output_manager: StructuredOutputManager,
48
+ block_size: int,
49
mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
50
include_finished_set: bool = False,
51
log_stats: bool = False,
@@ -101,15 +102,8 @@ class Scheduler(SchedulerInterface):
102
num_gpu_blocks = self.cache_config.num_gpu_blocks
103
assert num_gpu_blocks is not None and num_gpu_blocks > 0
104
104
- self.block_size = self.cache_config.block_size
105
-
105
+ self.block_size = block_size
106
self.dcp_world_size = vllm_config.parallel_config.decode_context_parallel_size
107
- # Note(hc): The scheduler’s block_size must be multiplied
108
- # by dcp_world_size, since block hashes are computed on the
109
- # original full token sequence at a granularity of
110
- # original_block_size × dcp_world_size.
111
- if self.dcp_world_size > 1:
112
- self.block_size *= self.dcp_world_size
107
108
# req_id -> Request
109
self.requests: dict[str, Request] = {}
vllm/v1/engine/core.py
+7
-2
@@ -142,12 +142,18 @@ class EngineCore:
142
logger.info("Disabling chunked prefill for model without KVCache")
143
vllm_config.scheduler_config.chunked_prefill_enabled = False
144
145
+ scheduler_block_size = (
146
+ vllm_config.cache_config.block_size
147
+ * vllm_config.parallel_config.decode_context_parallel_size
148
+ )
149
+
150
self.scheduler: SchedulerInterface = Scheduler(
151
vllm_config=vllm_config,
152
kv_cache_config=kv_cache_config,
153
structured_output_manager=self.structured_output_manager,
154
include_finished_set=vllm_config.parallel_config.data_parallel_size > 1,
155
log_stats=self.log_stats,
156
+ block_size=scheduler_block_size,
157
)
158
self.use_spec_decode = vllm_config.speculative_config is not None
159
if self.scheduler.connector is not None: # type: ignore
@@ -177,14 +183,13 @@ class EngineCore:
183
self.vllm_config.cache_config.enable_prefix_caching
184
or self.scheduler.get_kv_connector() is not None
185
):
180
- block_size = vllm_config.cache_config.block_size
186
caching_hash_fn = get_hash_fn_by_name(
187
vllm_config.cache_config.prefix_caching_hash_algo
188
)
189
init_none_hash(caching_hash_fn)
190
191
self.request_block_hasher = get_request_block_hasher(
187
- block_size, caching_hash_fn
192
+ scheduler_block_size, caching_hash_fn
193
)
194
195
self.step_fn = (