biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Feature] Prefill Context Parallel (PCP) basic support (#28718)

Browse Source 
Signed-off-by: QiuChunshuo <qiuchunshuo@huawei.com>
Signed-off-by: FENP <yuanyongjie.yyj@antgroup.com>
Signed-off-by: LookAround <lixushi@huawei.com>
Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com>
Signed-off-by: zhenwenqi2024 <zhenwenqi_2022@qq.com>
Co-authored-by: FENP <yuanyongjie.yyj@antgroup.com>
Co-authored-by: LookAround <lixushi@huawei.com>
Co-authored-by: Jingchun Gao <gaojingchun1@huawei.com>
Co-authored-by: zhenwenqi2024 <zhenwenqi_2022@qq.com>
Co-authored-by: Jingchun Gao <63247409+gjc0824@users.noreply.github.com>
This commit is contained in:
Qiu
2025-11-20 04:52:44 +08:00
committed by GitHub
parent 02f5903b84
commit 2fd893b4ce
27 changed files with 399 additions and 114 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
vllm/v1/kv_cache_interface.py
View File

@@ -95,10 +95,11 @@ class FullAttentionSpec(AttentionSpec):
def max_memory_usage_bytes(self, vllm_config: VllmConfig) -> int:
max_model_len = vllm_config.model_config.max_model_len
dcp_world_size = vllm_config.parallel_config.decode_context_parallel_size
pcp_world_size = vllm_config.parallel_config.prefill_context_parallel_size
# Note(hc): each dcp rank only need save
# (max_model_len//dcp_world_size) tokens locally.
if dcp_world_size > 1:
max_model_len = cdiv(max_model_len, dcp_world_size)
if dcp_world_size * pcp_world_size > 1:
max_model_len = cdiv(max_model_len, dcp_world_size * pcp_world_size)
return cdiv(max_model_len, self.block_size) * self.page_size_bytes
@classmethod