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

12
tests/distributed/test_context_parallel.py
View File

@@ -31,7 +31,7 @@ class ParallelSetup(NamedTuple):
tp_size: int
pp_size: int
dcp_size: int
dcp_kv_cache_interleave_size: int
cp_kv_cache_interleave_size: int
eager_mode: bool
chunked_prefill: bool
@@ -55,7 +55,7 @@ class CPTestSettings:
tp_base: int = 4,
pp_base: int = 1,
dcp_base: int = 1,
dcp_kv_cache_interleave_size: int = 1,
cp_kv_cache_interleave_size: int = 1,
multi_node_only: bool = False,
runner: RunnerOption = "auto",
load_format: str | None = None,
@@ -71,7 +71,7 @@ class CPTestSettings:
tp_size=tp_base,
pp_size=pp_multiplier * pp_base,
dcp_size=int(dcp_multiplier * tp_base),
dcp_kv_cache_interleave_size=dcp_kv_cache_interleave_size,
cp_kv_cache_interleave_size=cp_kv_cache_interleave_size,
eager_mode=eager_mode_val,
chunked_prefill=chunked_prefill_val,
)
@@ -116,7 +116,7 @@ def _compare_cp_with_tp(
tp_size,
pp_size,
dcp_size,
dcp_kv_cache_interleave_size,
cp_kv_cache_interleave_size,
eager_mode,
chunked_prefill,
) = parallel_setup
@@ -197,7 +197,7 @@ def _compare_cp_with_tp(
"--decode-context-parallel-size",
str(dcp_size),
"--dcp-kv-cache-interleave-size",
str(dcp_kv_cache_interleave_size),
str(cp_kv_cache_interleave_size),
"--distributed-executor-backend",
distributed_backend,
]
@@ -227,7 +227,7 @@ CP_TEXT_GENERATION_MODELS = {
"deepseek-ai/DeepSeek-V2-Lite-Chat": [
CPTestSettings.detailed(),
CPTestSettings.detailed(tp_base=2),
CPTestSettings.detailed(tp_base=2, dcp_kv_cache_interleave_size=64),
CPTestSettings.detailed(tp_base=2, cp_kv_cache_interleave_size=64),
],
"bigcode/gpt_bigcode-santacoder": [
CPTestSettings.detailed(),

7
tests/kernels/moe/modular_kernel_tools/common.py
View File

@@ -15,7 +15,11 @@ from tests.kernels.quantization.nvfp4_utils import (
)
from tests.kernels.utils import torch_experts
from vllm.config import VllmConfig
from vllm.distributed import get_dp_group, get_tensor_model_parallel_world_size
from vllm.distributed import (
get_dp_group,
get_pcp_group,
get_tensor_model_parallel_world_size,
)
from vllm.forward_context import set_forward_context
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig,
@@ -561,6 +565,7 @@ def make_modular_kernel(
# make moe config
moe_parallel_config: FusedMoEParallelConfig = FusedMoEParallelConfig.make(
tp_size_=get_tensor_model_parallel_world_size(),
pcp_size_=get_pcp_group().world_size,
dp_size_=get_dp_group().world_size,
vllm_parallel_config=vllm_config.parallel_config,
)

4
tests/v1/worker/test_gpu_model_runner.py
View File

@@ -956,7 +956,7 @@ def test_hybrid_block_table_initialization():
max_num_reqs = 10
max_num_blocks_per_req = 20
max_num_batched_tokens = 512
dcp_kv_cache_interleave_size = 8
cp_kv_cache_interleave_size = 8
block_table = BlockTable(
block_size=block_size,
@@ -966,7 +966,7 @@ def test_hybrid_block_table_initialization():
pin_memory=False,
device=torch.device(DEVICE),
kernel_block_size=kernel_block_sizes[0],
dcp_kv_cache_interleave_size=dcp_kv_cache_interleave_size,
cp_kv_cache_interleave_size=cp_kv_cache_interleave_size,
)
# Verify hybrid block configuration