[faresobeid]

Adds a deep-ep-v2 wheel to the kernel release, built from deepseek-ai/DeepEP @ 93eb6eb (V2.5, DeepJIT submodule cba802d).

V2.5 is a rewrite on top of NCCL's device/GIN API (no NVSHMEM). In an EP8 trial in prime-rl (8k tokens/rank) it cut dispatch from 3.2 to 1.6 ms and combine from 5.8 to 3.7 ms vs V1. It ships as a separate package so it can sit next to V1 while we compare them:

• Dist deep-ep-v2, import deep_ep_v2, extension deep_ep_v2._C. The script renames the Python package and its deep_ep._C imports.
• The host extension is compiled with -Ddeep_ep=deep_ep_v2. pybind11 registers types by C++ type, and V1 already registers deep_ep::EventHandle etc., so without this import deep_ep, deep_ep_v2 fails with generic_type: type "EventHandle" is already registered!. The macro does not touch #include <deep_ep/...> or the JIT's kernel source strings (deep_ep::ep::...); deep_ep is only stringified in error messages.
• Unchanged: the header dir include/deep_ep/ and the namespaces in the JIT-compiled kernels. The JIT includes <deep_ep/...> relative to <package>/include.
• Version stays upstream: 2.5.0+93eb6eb.

Build quirks:

• The extension is host C++ only. All kernels are JIT-compiled at runtime by DeepJIT, so there is no TORCH_CUDA_ARCH_LIST and there are no cubins in the wheel. The host build works with the image's CUDA 13.0.
• NCCL comes from the pip package nvidia-nccl-cu13. V2.5 needs >= 2.32.3, but torch 2.13 pins 2.29.7, so the script upgrades it in the build venv.
• setup.py bakes in an absolute RPATH to the build venv's NCCL. The script repoints it to $ORIGIN/../nvidia/nccl/lib, like install_ep_kernels.sh does for NVSHMEM.
• DeepJIT needs elfutils/libdwfl.h (elfutils-devel added to the dnf line). libdw is dlopened at runtime and is optional (it is used for C++ stack traces).
• setup.py hard-codes /usr/local/cuda/include/cccl, so the script sets CPATH. The DeepJIT submodule URL is ssh, so it is rewritten to https.
• The wheel does not declare the NCCL requirement: declaring it would conflict with torch's pin. Consumers need override-dependencies with nvidia-nccl-cu13>=2.32.3. import deep_ep_v2 fails (check_nccl_so) if the NCCL that torch loads is not the one it links.
• Runtime also needs a CUDA >= 13.1 toolkit at CUDA_HOME for the JIT (e.g. pip nvidia-cuda-nvcc/crt/nvvm/runtime/cccl 13.x installed into a dedicated dir). Cubins from nvcc 13.3 run on the 13.0 driver.

Smoke test: import (this runs check_nccl_so + init_jit), check that the loaded libnccl is >= 2.32.3, and build the JIT runtime for the GPU (_C.get_jit()). Real dispatch/combine needs an EP group and a 13.1+ toolkit, so it is out of scope. The smoke job installs nccl 2.32.3 for this kernel.

Validated (built on the login node: Ubuntu 24.04, CUDA 13.0 toolkit, torch 2.13.0+cu130):

• The script builds the wheel (8.5 MB, one .so). RUNPATH is $ORIGIN/../nvidia/nccl/lib. NEEDED has no libcuda (the driver is loaded lazily).
• In a fresh venv (torch + nvidia-nccl-cu13 2.32.3 + wheel), import deep_ep_v2 passes, including check_nccl_so. /proc/self/maps shows one libnccl (pip 2.32.3) and one libcudart (torch's pip nvidia/cu13).
• Installs next to the V1 wheel deep_ep-1.2.1+29d31c0: no shared paths (deep_ep/, deep_ep_cpp*.so vs deep_ep_v2/).
• nm -DC: 145 exported deep_ep_v2:: symbols, 0 deep_ep::; the pybind type info is now N10deep_ep_v211EventHandleE.
• On an H200 node, in a prime-rl venv with V1 deep_ep 1.2.1 and nccl 2.32.3: import torch, deep_ep, deep_ep_v2 works, and scripts/smoke_test.py deep-ep-v2 passes (get_jit() builds the runtime even with CUDA_HOME pointing at a 12.9 toolkit).
• Earlier trial (same commit, built by hand): upstream EP tests and an EP8/EP16 prime-rl dispatcher bench pass on H200.

Validated since (the wheel from this branch, in prime-rl PR PrimeIntellect-ai/prime-rl#3999 on H200):

• The CI builds pass (x86_64 and aarch64, manylinux image).
• JIT kernels compile and run from this wheel, including with the system CUDA 13.0 toolkit and an empty cache. EP8 dispatch/combine match the torch dispatcher in eager, compile and full AC.
• GLM-4.5-Air EP8 training steps are ~5% faster than torch dispatch.

Not validated:

• The GPU smoke job: it is disabled in this repo's CI (if: false), so it was only run by hand on H200.
• sm_100 / sm_103: no Blackwell here. The JIT compiles for the running GPU, so the wheel itself has no per-arch code.
• Multi-node: V2.5 is slower than V1 at EP16 on our fabric, so prime-rl rejects deepep_v2 across nodes.

prime-rl will pin deep-ep-v2 after a release that includes this PR.