refactor: import deep_ep where it is used

SM120 has no DeepEP feature left (no deepep backend, Engram in host memory,
no context parallelism, no NVLink domain), yet five modules imported deep_ep
at the top, so every SM120 process needed a DeepEP build, one compiled
against an NCCL newer than the one the process loads.

- mlp, engram, cp, userbuffer: deep_ep names used only in annotations move
  under TYPE_CHECKING; EPBuffer, EngramBuffer, BucketBuffer and the CP
  alignment helpers are imported in the functions that build them.
- parallel_state: is_within_nvl_domain imports deep_ep after the SM120 and
  single-rank returns.

No behavior change where DeepEP is installed: the same code imports it,
later. tests/test_sm120_without_deep_ep.py imports each module in a fresh
interpreter that cannot import deep_ep, and runs the SM120 paths (nccl EP
buffer, no CP buffer, no NVLink domain) with deep_ep blocked; the nccl
backend allocation test blocks deep_ep too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
shunyadandClaude Opus 5.5 committed 2026-11-18 23:43:19 -08:00
1 parent a2d45e4f22
commit f009131253
7 files changed
+128 -9

No files matched your search

@@ -4,12 +4,10 @@ from collections.abc import Generator
from contextlib import contextmanager
from typing import TYPE_CHECKING, NamedTuple
import deep_ep
import deep_gemm
import tile_kernels
import torch
import torch.distributed as dist
from deep_ep import EngramBuffer, EPBuffer
from deepinfer import PLATFORM, envs, sm120_support
from deepinfer.logger import init_logger
@@ -17,7 +15,10 @@ from deepinfer.model_executor.parallel_utils.cp import CPManager
from deepinfer.model_executor.parallel_utils.host_engram import HostEngramBuffer
from deepinfer.model_executor.parallel_utils.nccl_ep import NcclEPBuffer
# deep_ep is imported where it is used: SM120 builds do not install DeepEP (see README.sm120.md)
if TYPE_CHECKING:
from deep_ep import EngramBuffer, EPBuffer
from deepinfer.model_executor.layers.cake_moe import CakeMoE
logger = init_logger(__name__)
@@ -129,6 +130,8 @@ class _UserBufferManager:
logger.info(f"Created NCCL EP buffer, (EP {ep_group.rank()}/{ep_group.size()}), {ep_kwargs=}")
assert isinstance(self.ep, NcclEPBuffer) and self.ep.group is ep_group
return self.ep
from deep_ep import EPBuffer
if self.ep is not None:
assert self.ep.num_bytes >= EPBuffer.get_buffer_size_hint(**ep_kwargs)
return self.ep
@@ -158,6 +161,8 @@ class _UserBufferManager:
if cp_group.size() <= 1:
assert self.cp is None, "cp buffer was already allocated for a multi-rank CP group"
return None
import deep_ep
max_ag_bytes_per_session = max_num_tokens * max_ag_bytes_per_token * cp_group.size()
max_ag_bytes_per_session += max_num_ags_per_session * deep_ep.get_num_rdma_alignment()
alignment = deep_ep.get_num_allocation_alignment()
@@ -204,6 +209,8 @@ class _UserBufferManager:
f"({len(num_entries_per_layer)} layers, Engram {engram_group.rank()}/{engram_group.size()})"
)
return self.engram
from deep_ep import EngramBuffer
num_gpu_bytes, num_rdma_storage_size = EngramBuffer.get_storage_size_hint(
engram_group,
num_entries_per_layer,