[WideEP] Integrate DeepEP v2 (#41183)

Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Tyler Michael Smith
2026-06-08 18:07:29 -07:00
committed by GitHub
co-authored by Claude Opus 4.6
parent 05cb606cad
commit e2f993dc41
22 changed files with 1409 additions and 17 deletions
+46 -3
View File
@@ -15,7 +15,7 @@ from torch.distributed import ProcessGroup
from torch.multiprocessing import spawn # pyright: ignore[reportPrivateImportUsage]
from typing_extensions import ParamSpec
from vllm.utils.import_utils import has_deep_ep
from vllm.utils.import_utils import has_deep_ep, has_deep_ep_v2
from vllm.utils.network_utils import get_open_port
if has_deep_ep():
@@ -26,6 +26,11 @@ if has_deep_ep():
DeepEPLLPrepareAndFinalize,
)
if has_deep_ep_v2():
from vllm.model_executor.layers.fused_moe.prepare_finalize.deepep_v2 import (
DeepEPV2PrepareAndFinalize,
)
## Parallel Processes Utils
P = ParamSpec("P")
@@ -55,11 +60,10 @@ def _worker_parallel_launch(
torch.accelerator.set_device_index(local_rank)
device = torch.device("cuda", local_rank)
torch.distributed.init_process_group(
backend="cpu:gloo,cuda:nccl",
backend="nccl",
init_method=init_method,
rank=rank,
world_size=world_size,
device_id=device,
)
barrier = torch.tensor([rank], device=device)
torch.distributed.all_reduce(barrier)
@@ -200,3 +204,42 @@ def make_deepep_a2a(
assert deepep_ll_args is not None
return make_deepep_ll_a2a(pg, pgi, deepep_ll_args, q_dtype, block_shape)
@dataclasses.dataclass
class DeepEPV2Args:
num_local_experts: int
num_experts: int
num_topk: int
hidden_size: int
max_tokens_per_rank: int
use_fp8_dispatch: bool
def make_deepep_v2_a2a(
pg: ProcessGroup,
pgi: ProcessGroupInfo,
dp_size: int,
v2_args: DeepEPV2Args,
use_cudagraph: bool = False,
):
import deep_ep
buffer = deep_ep.ElasticBuffer(
group=pg,
num_max_tokens_per_rank=v2_args.max_tokens_per_rank,
hidden=v2_args.hidden_size,
num_topk=v2_args.num_topk,
use_fp8_dispatch=v2_args.use_fp8_dispatch,
explicitly_destroy=True,
)
return DeepEPV2PrepareAndFinalize(
buffer=buffer,
num_dispatchers=pgi.world_size,
dp_size=dp_size,
rank_expert_offset=pgi.rank * v2_args.num_local_experts,
num_experts=v2_args.num_experts,
num_topk=v2_args.num_topk,
use_fp8_dispatch=v2_args.use_fp8_dispatch,
use_cudagraph=use_cudagraph,
)