[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
+15
View File
@@ -246,6 +246,9 @@ if TYPE_CHECKING:
VLLM_DEEPEP_BUFFER_SIZE_MB: int = 1024
VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE: bool = False
VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL: bool = False
VLLM_DEEPEP_V2_ALLOW_HYBRID_MODE: bool = True
VLLM_DEEPEP_V2_PREFER_OVERLAP: bool = False
VLLM_DEEPEP_V2_ALLOW_MULTIPLE_REDUCTION: bool = False
VLLM_DBO_COMM_SMS: int = 20
VLLM_PATTERN_MATCH_DEBUG: str | None = None
VLLM_DEBUG_DUMP_PATH: str | None = None
@@ -1840,6 +1843,18 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL": lambda: bool(
int(os.getenv("VLLM_DEEPEP_LOW_LATENCY_USE_MNNVL", "0"))
),
# DeepEP v2: enable two-tier NVLink+RDMA hybrid mode
"VLLM_DEEPEP_V2_ALLOW_HYBRID_MODE": lambda: bool(
int(os.getenv("VLLM_DEEPEP_V2_ALLOW_HYBRID_MODE", "0"))
),
# DeepEP v2: use fewer SMs at slight throughput cost
"VLLM_DEEPEP_V2_PREFER_OVERLAP": lambda: bool(
int(os.getenv("VLLM_DEEPEP_V2_PREFER_OVERLAP", "0"))
),
# DeepEP v2: trade precision for transfer size in combine
"VLLM_DEEPEP_V2_ALLOW_MULTIPLE_REDUCTION": lambda: bool(
int(os.getenv("VLLM_DEEPEP_V2_ALLOW_MULTIPLE_REDUCTION", "0"))
),
# The number of SMs/CUs to allocate for communication kernels when
# running DBO; the rest will be allocated to compute.
# Default: 20 on CUDA (SMs), 64 on ROCm (CUs).