- Add PushAllReduce to benchmark_device_communicators.py for
comparing against other allreduce implementations
- Add PUSH_AR to _log_all_reduce_backend_selection in
cuda_communicator.py for visibility in dispatch logging
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com>
Replace direct os.environ.get(_DISABLE_ENV_VAR) == "1" check with
envs.VLLM_DISABLE_PUSH_ALLREDUCE to use the centrally registered
env var from envs.py, which provides validation and caching.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com>
Register the push allreduce feature toggle env var in the central
envs.py registry so it is validated on startup and follows the
standard vllm env var pattern. Default is False (push allreduce
enabled); set to 1 to disable.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com>
The push threshold map was labeled as sm100-specific but applied
unconditionally to all architectures. Now:
- PUSH_THRESHOLD_SM100 is only used on Blackwell (compute capability 10.x)
- PUSH_THRESHOLD_DEFAULT provides conservative 512 KB thresholds for
architectures without tuned values
- _THRESHOLD_BY_ARCH maps GPU major compute capability to threshold tables
- A log message is emitted when falling back to conservative defaults
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com>
- Add PushAllReduce | None type annotation on push_ar_comm to be
consistent with other communicator fields (ca_comm, qr_comm, etc.)
- Add push_ar_comm.close() + None assignment in destroy() method
to match the cleanup pattern for other communicators
- Add lazy import of PushAllReduce alongside other communicator imports
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com>
Fix CodeQL security warning by binding test helper sockets to
"localhost" instead of "" (all interfaces). These sockets are only
used for finding a free port for torch distributed init in tests.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com>
- Add PUSH_AR_CUDACHECK macro wrapping all CUDA API calls (cudaGetDevice,
cudaDeviceGetAttribute, cudaMalloc, cudaMemset, cudaIpcGetMemHandle,
cudaIpcOpenMemHandle) to match the CUDACHECK pattern in custom_all_reduce.cuh
- Replace assert(input_bytes <= push_buffer_bytes_) with a runtime
std::runtime_error check that is not compiled out under -DNDEBUG
- Add #include <stdexcept> and #include <string> for the runtime check
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com>
Port SGLang's push-based 2-buffer allreduce protocol into vLLM as a new
communicator backend for small-message reductions. The push protocol
eliminates the two explicit cross-GPU NVLink barrier round-trips used by
the existing barrier-based CustomAllreduce, replacing them with a
sentinel-based data arrival detection mechanism and double-buffered epoch
alternation.
Key advantages over the barrier-based approach:
- Zero barriers: data arrival IS the synchronization (positive-zero sentinel)
- Single NVLink round-trip instead of two barrier exchanges + remote reads
- All SMs active (SM_count CTAs vs 2 CTAs) for higher NVLink bandwidth
- No cudaMemcpy to IPC staging buffer in eager mode
- PDL (griddepcontrol) support for kernel overlap on sm_90+
The new PushAllReduce is inserted in the CudaCommunicator dispatch chain
above the existing CustomAllreduce for messages below a size threshold
(~720 KB at TP=8). Larger messages continue to use the barrier-based
path. The existing CustomAllreduce code is not modified.
Measured results on DeepSeek-V4-Pro (61 layers, TP=8, 8x NVIDIA B200,
BS=1, decode with ISL=4, OSL=33024):
- Throughput: +2.14% (84.06 vs 82.30 tokens/s)
- TPOT: -2.09% (11.90 vs 12.15 ms/token)
Correctness verified via lm_eval gsm8k 5-shot with no regression
(exact_match delta within statistical noise).
The feature can be disabled at runtime via VLLM_DISABLE_PUSH_ALLREDUCE=1
to fall back to the barrier-based path.
Signed-off-by: Alexander Matveev <amatveev@redhat.com>
Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com>
