[ROCm] mori: add InterNodeV1LL inter-node kernel selection via VLLM_MORI_INTERNODE_KERNEL (#41751)

Signed-off-by: jatseng-ai <jatseng@amd.com>
This commit is contained in:
jatseng-ai
2026-05-28 00:33:32 +08:00
committed by GitHub
parent 41688e2dc7
commit 05c50c721e
7 changed files with 36 additions and 14 deletions
+6 -3
View File
@@ -42,7 +42,8 @@ All2AllBackend = Literal[
"pplx",
"deepep_high_throughput",
"deepep_low_latency",
"mori",
"mori_high_throughput",
"mori_low_latency",
"nixl_ep",
"allgather_reducescatter",
"flashinfer_all2allv", # temporary alias for flashinfer_nvlink_two_sided
@@ -177,7 +178,8 @@ class ParallelConfig:
- "allgather_reducescatter": All2all based on allgather and reducescatter
- "deepep_high_throughput": Use deepep high-throughput kernels
- "deepep_low_latency": Use deepep low-latency kernels
- "mori": Use mori kernels
- "mori_high_throughput": MoRI EP with InterNodeV1 for multi-node
- "mori_low_latency": MoRI EP with InterNodeV1LL for multi-node
- "nixl_ep": Use nixl-ep kernels
- "flashinfer_nvlink_two_sided": Use flashinfer two-sided kernels for mnnvl
- "flashinfer_nvlink_one_sided": Use flashinfer high-throughput a2a kernels"""
@@ -621,7 +623,8 @@ class ParallelConfig:
"allgather_reducescatter",
"deepep_high_throughput",
"deepep_low_latency",
"mori",
"mori_high_throughput",
"mori_low_latency",
"nixl_ep",
)
and self.enable_expert_parallel