DCP supports hybrid attention (#40996)

Signed-off-by: YanXu <yancey.yx@alibaba-inc.com>
Signed-off-by: Jingyi Yang <girasoleyang@gmail.com>
Co-authored-by: Jingyi Yang <girasoleyang@gmail.com>
This commit is contained in:
Yan Xu
2026-07-09 21:34:45 -07:00
committed by GitHub
co-authored by Jingyi Yang
parent 2d814a0082
commit 95ed0feaa5
26 changed files with 783 additions and 125 deletions
@@ -33,6 +33,7 @@ CP_TEST_MODELS = [
# [LANGUAGE GENERATION]
"deepseek-ai/DeepSeek-V2-Lite-Chat",
"Qwen/Qwen2.5-1.5B-Instruct",
"Qwen/Qwen3.5-0.8B", # hybrid attention model
]
# GSM8K eval configuration
@@ -46,6 +47,7 @@ MIN_ACCURACY = {
"deepseek-ai/DeepSeek-V2-Lite-Chat": 0.64,
# .buildkite/lm-eval-harness/configs/Qwen2.5-1.5B-Instruct.yaml
"Qwen/Qwen2.5-1.5B-Instruct": 0.52,
"Qwen/Qwen3.5-0.8B": 0.33,
}
@@ -151,6 +153,12 @@ else:
cp_kv_cache_interleave_size=16, attn_backend="FLASHINFER"
),
],
"Qwen/Qwen3.5-0.8B": [
CPTestSettings.detailed(
cp_kv_cache_interleave_size=16,
attn_backend="FLASH_ATTN",
),
],
}
+47
View File
@@ -16,6 +16,7 @@ from vllm.distributed.device_communicators.pynccl import PyNcclCommunicator
from vllm.distributed.device_communicators.pynccl_wrapper import NCCLLibrary
from vllm.distributed.parallel_state import (
ensure_model_parallel_initialized,
get_tp_group,
get_world_group,
graph_capture,
init_distributed_environment,
@@ -199,6 +200,52 @@ def test_pynccl_all_gather():
distributed_run(all_gather_worker_fn, 2)
@worker_fn_wrapper
def cuda_communicator_all_gather_dim_worker_fn():
with ensure_current_vllm_config():
ensure_model_parallel_initialized(2, 1)
tp_group = get_tp_group()
comm = tp_group.device_communicator
assert comm is not None
rank = tp_group.rank_in_group
world_size = tp_group.world_size
device = tp_group.device
shape = (2, 3, 4)
num_elems = 1
for size in shape:
num_elems *= size
for dim in (1, -1):
tensor = (
torch.arange(num_elems, dtype=torch.float32, device=device).reshape(shape)
+ rank * num_elems
)
expected = torch.cat(
[
torch.arange(num_elems, dtype=torch.float32, device=device).reshape(
shape
)
+ r * num_elems
for r in range(world_size)
],
dim=dim,
)
result = comm.all_gather(tensor, dim=dim)
torch.accelerator.synchronize()
torch.testing.assert_close(result, expected, rtol=1e-5, atol=1e-8)
@pytest.mark.skipif(
torch.accelerator.device_count() < 2, reason="Need at least 2 GPUs to run the test."
)
def test_cuda_communicator_all_gather_dim_not_zero():
distributed_run(cuda_communicator_all_gather_dim_worker_fn, 2)
@worker_fn_wrapper
def all_gatherv_worker_fn():
pynccl_comm = PyNcclCommunicator(