Gahow Wang
445e491123
third_party/vllm/ now tracked in git for direct patch management.
Based on vLLM v0.18.1 release with one patch applied:
vllm/v1/core/sched/scheduler.py:
Replace fatal assert with graceful skip when KV transfer callback
arrives for an already-aborted request during PD disaggregated serving.
Future vLLM modifications should be made directly in third_party/vllm/
and committed normally. The patches/ directory is kept as documentation
of what changed from upstream.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Expert parallel kernels
Large-scale cluster-level expert parallel, as described in the DeepSeek-V3 Technical Report, is an efficient way to deploy sparse MoE models with many experts. However, such deployment requires many components beyond a normal Python package, including system package support and system driver support. It is impossible to bundle all these components into a Python package.
Here we break down the requirements in 2 steps:
- Build and install the Python libraries (DeepEP), including necessary dependencies like NVSHMEM. This step does not require any privileged access. Any user can do this.
- Configure NVIDIA driver to enable IBGDA. This step requires root access, and must be done on the host machine.
Step 2 is necessary for multi-node deployment.
All scripts accept a positional argument as workspace path for staging the build, defaulting to $(pwd)/ep_kernels_workspace.
Usage
# for hopper
TORCH_CUDA_ARCH_LIST="9.0" bash install_python_libraries.sh
# for blackwell
TORCH_CUDA_ARCH_LIST="10.0" bash install_python_libraries.sh
Additional step for multi-node deployment:
sudo bash configure_system_drivers.sh # update-initramfs can take several minutes
sudo reboot # Reboot is required to load the new driver