Files
vllm/csrc/torch_bindings.cpp
T
Alexander MatveevandAlexander Matveev e3b4fdaf5d perf: add push-based allreduce for small tensor reductions
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>
2026-06-15 16:23:12 -04:00

130 lines
5.1 KiB
C++

// Provides torch::Tensor for ops.h (previously included transitively via
// cache.h, which is no longer included here after cache ops moved to
// _C_stable_libtorch).
#include <torch/all.h>
#include "cuda_utils.h"
#include "ops.h"
#include "core/registration.h"
#include <torch/library.h>
#include <torch/version.h>
// Note on op signatures:
// The X_meta signatures are for the meta functions corresponding to op X.
// They must be kept in sync with the signature for X. Generally, only
// functions that return Tensors require a meta function.
//
// See the following links for detailed docs on op registration and function
// schemas.
// https://docs.google.com/document/d/1_W62p8WJOQQUzPsJYa7s701JXt0qf2OfLub2sbkHOaU/edit#heading=h.ptttacy8y1u9
// https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/README.md#annotations
TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
// vLLM custom ops
//
ops.def(
"persistent_masked_m_silu_mul_quant(Tensor input, Tensor counts, Tensor! "
"y_q, Tensor! y_s,"
"bool use_ue8m0) -> ()");
ops.impl("persistent_masked_m_silu_mul_quant", torch::kCUDA,
&persistent_masked_m_silu_mul_quant);
ops.def("weak_ref_tensor(Tensor input) -> Tensor");
ops.impl("weak_ref_tensor", torch::kCUDA, &weak_ref_tensor);
#ifdef USE_ROCM
// TODO: Remove this once we upgrade to torch 2.11.
// ROCm still uses torch 2.10,
// So we still need to use unstable torch ABI for now.
ops.def("get_cuda_view_from_cpu_tensor(Tensor cpu_tensor) -> Tensor");
ops.impl("get_cuda_view_from_cpu_tensor", torch::kCPU,
&get_cuda_view_from_cpu_tensor);
#endif
// Activation ops (quantized only — basic ops moved to _C_stable_libtorch)
ops.def(
"silu_and_mul_quant(Tensor! result, Tensor input, Tensor scale) -> ()");
ops.impl("silu_and_mul_quant", torch::kCUDA, &silu_and_mul_quant);
// Horizontally-fused DeepseekV4-MLA: per-head RMSNorm + GPT-J RoPE for Q, and
// GPT-J RoPE + UE8M0 FP8 quant + paged cache insert for KV, all in one
// kernel launch. Registered in _C_stable_libtorch (incl. the FlashInfer V4
// full-cache bf16/fp8 variants).
// Quantization ops
#ifndef USE_ROCM
// Note about marlin kernel 'workspace' arguments:
// Technically these should be mutable since they are modified by the kernel.
// But since they are set back to zero once the kernel is finished we can
// hand wave and say that they have no net effect.
//
// The reason to mark 'workspace' as immutable is so that they don't interfere
// with using ScalarType arguments in the ops. If they are marked as mutable,
// pytorch throws an assert in
// 'torch._higher_order_ops._register_effectful_op' that prevents these
// kernels from being torch.compile'd.
// See the following document for more info on custom types and ops that use
// custom types:
// https://docs.google.com/document/d/18fBMPuOJ0fY5ZQ6YyrHUppw9FA332CpNtgB6SOIgyuA
#endif
}
#ifdef USE_ROCM
TORCH_LIBRARY_FRAGMENT(CONCAT(TORCH_EXTENSION_NAME, _custom_ar), custom_ar) {
// Quick Reduce all-reduce kernels (ROCm-only; stays on legacy _C).
custom_ar.def(
"qr_all_reduce(int fa, Tensor inp, Tensor out, int quant_level, bool "
"cast_bf2half) -> ()");
custom_ar.impl("qr_all_reduce", torch::kCUDA, &qr_all_reduce);
custom_ar.def("init_custom_qr", &init_custom_qr);
custom_ar.def("qr_destroy", &qr_destroy);
custom_ar.def("qr_get_handle", &qr_get_handle);
custom_ar.def("qr_open_handles(int _fa, Tensor[](b!) handles) -> ()");
custom_ar.impl("qr_open_handles", torch::kCPU, &qr_open_handles);
custom_ar.def("qr_max_size", &qr_max_size);
}
// TODO: Remove this once ROCm upgrade to torch 2.11.
TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cuda_utils), cuda_utils) {
// Cuda utils
// Gets the specified device attribute.
cuda_utils.def("get_device_attribute(int attribute, int device_id) -> int");
cuda_utils.impl("get_device_attribute", &get_device_attribute);
// Gets the maximum shared memory per block device attribute.
cuda_utils.def(
"get_max_shared_memory_per_block_device_attribute(int device_id) -> int");
cuda_utils.impl("get_max_shared_memory_per_block_device_attribute",
&get_max_shared_memory_per_block_device_attribute);
}
#endif
// Push-based allreduce (ported from SGLang)
fptr_t init_push_ar(int64_t rank, int64_t world_size,
int64_t push_buffer_bytes, int64_t max_num_cta);
torch::Tensor get_push_ar_ipc_handle(fptr_t _mgr);
void post_init_push_ar(fptr_t _mgr, torch::Tensor all_handles);
void push_ar_all_reduce(fptr_t _mgr, torch::Tensor& inp, torch::Tensor& out);
void dispose_push_ar(fptr_t _mgr);
TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _push_ar), push_ar) {
push_ar.def("init_push_ar", &init_push_ar);
push_ar.def("get_push_ar_ipc_handle", &get_push_ar_ipc_handle);
push_ar.def("post_init_push_ar", &post_init_push_ar);
push_ar.def(
"push_ar_all_reduce(int mgr, Tensor inp, Tensor! out) -> ()");
push_ar.impl("push_ar_all_reduce", torch::kCUDA, &push_ar_all_reduce);
push_ar.def("dispose_push_ar", &dispose_push_ar);
}
REGISTER_EXTENSION(TORCH_EXTENSION_NAME)