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 4, 06:15 UTC
[Bugfix][KVConnector] Support DCP/PCP in OffloadingConnector (#41549)
Signed-off-by: Itay Etelis <itay.etelis@ibm.com> Co-authored-by: Itay Etelis <itay.etelis@ibm.com>
Itay Etelis committed
May 5, 2026 at 14:54 UTC
98661fe012c5c467252d4df8411d2f46190e9268
1 file changed
+10
-2
vllm/v1/kv_offload/base.py
+10
-2
@@ -331,12 +331,20 @@ class OffloadingSpec(ABC):
331
assert kv_transfer_config is not None
332
self.extra_config = kv_transfer_config.kv_connector_extra_config
333
334
+ parallel_config = vllm_config.parallel_config
335
+ context_parallel_factor = (
336
+ parallel_config.decode_context_parallel_size
337
+ * parallel_config.prefill_context_parallel_size
338
+ )
339
+
340
# block size used by vLLM for hashing request tokens for the sake
341
# of enabling prefix caching
336
- self.hash_block_size = vllm_config.cache_config.block_size
342
+ self.hash_block_size = (
343
+ vllm_config.cache_config.block_size * context_parallel_factor
344
+ )
345
# gpu block size per group
346
self.gpu_block_size: tuple[int, ...] = tuple(
339
- kv_cache_group.kv_cache_spec.block_size
347
+ kv_cache_group.kv_cache_spec.block_size * context_parallel_factor
348
for kv_cache_group in kv_cache_config.kv_cache_groups
349
)
350