feat: make moe_ep (EP) part of the default install; drop nccl submodule - #3821

Merged

Anerudhan merged 13 commits into

flashinfer-ai:mainfrom

Anerudhan:feat/nvep-default

Jul 8, 2026

Commits on Jul 7, 2026

feat: make moe_ep (EP) part of the default install; drop nccl submodule

Anerudhan
and
claude
committed
153afef
fix: address review comments on EP-default build

Anerudhan
and
claude
committed
155b66b
test: skip nixl_ep mock tests on CUDA < 13 torch builds

Anerudhan
and
claude
committed
d12a6b5
fix: drop nvidia-nccl-cu13 floor from base deps; enforce at runtime

Anerudhan
and
claude
committed
e2bfadc
fix(moe_ep): clamp nccl_ep HT max_tokens_per_rank to MAX_SUPPORTED_TOKENS_PER_RANK

Anerudhan
and
claude
committed
127d539
perf(moe_ep): fleet-level host-path caches for the per-forward handle flow

Anerudhan
and
claude
committed
96ab962
docs(moe_ep): update runbook + results for the DP-EP perf iteration

Anerudhan
and
claude
committed
34512db
fix(moe_ep): declare _hot_cache on NcclEpFleet (mypy)

Anerudhan
and
claude
committed
0ce6980
fix(deps): relax cuda-python floor to >=12.0 (don't force a CUDA major)

Anerudhan
and
claude
committed
c9dcc80
docs(moe_ep): drop internal validation docs from the PR

Anerudhan
and
claude
committed
566e8db
docs(moe_ep): inline the content of the removed internal validation docs

Anerudhan
and
claude
committed
767f5df
ci: stop nvshmem4py-cu12 install from evicting CUDA torch; pin torch for job pips

Anerudhan
and
claude
committed
a68ab63
ci: strip the +cuXXX local tag from the torch pip constraint

Anerudhan
and
claude
committed
ae8d8f6