Sync from v0.13

2026-01-19 10:38:50 +08:00
parent b2ef04d792
commit 5aef6c175a
3714 changed files with 854317 additions and 89342 deletions
--- a/tools/ep_kernels/elastic_ep/eep_nvshmem.patch
+++ b/tools/ep_kernels/elastic_ep/eep_nvshmem.patch
@@ -0,0 +1,92 @@
+From 18c0599c2f07ec965132efa25961dc8179c2dda3 Mon Sep 17 00:00:00 2001
+From: Yongji Wu <wuyongji317@gmail.com>
+Date: Tue, 20 May 2025 13:41:12 -0700
+Subject: [PATCH] fix reinit issues due to states not cleaned up
+
+fix double free
+---
+ src/host/init/init.cu                             | 10 ++++++++++
+ .../internal/host/nvshmemi_mem_transport.hpp      | 15 +++++++++++++++
+ src/modules/bootstrap/uid/bootstrap_uid.cpp       |  5 +++++
+ 3 files changed, 30 insertions(+)
+
+diff --git a/src/host/init/init.cu b/src/host/init/init.cu
+index b1c5dbf..1fecb4b 100644
+--- a/src/host/init/init.cu
+++ b/src/host/init/init.cu
+@@ -43,6 +43,8 @@
+ #include "internal/host/nvshmemi_types.h"
+ #include "internal/host/shared_memory.h"
+ #include "internal/host/nvshmemi_symmetric_heap.hpp"
+// eep-dev
+#include "internal/host/nvshmemi_mem_transport.hpp"
+ 
+ extern __constant__ nvshmemi_device_host_state_t nvshmemi_device_state_d;
+ static std::map<void *, int> registered_device_states;
+@@ -1293,6 +1295,14 @@ void nvshmemid_hostlib_finalize(void *device_ctx, void *transport_device_ctx) {
+         /* Multi-init Multi-fini*/
+         nvshmemi_state = NULL;
+         nvshmemi_device_state.nvshmemi_is_nvshmem_initialized = 0;
+        
+        // eep-dev
+        nvshmemi_mem_p2p_transport::destroy_instance();
+        nvshmemi_mem_remote_transport::destroy_instance();
+        free(nvshmemi_default_session);
+        nvshmemi_default_session = nullptr;
+        nvshmemi_device_state.nvshmemi_is_nvshmem_bootstrapped = false;
+        
+         nvshmemi_is_device_state_ready = false;
+     } else
+         nvshmemi_boot_handle.barrier(&nvshmemi_boot_handle);
+diff --git a/src/include/internal/host/nvshmemi_mem_transport.hpp b/src/include/internal/host/nvshmemi_mem_transport.hpp
+index 2495844..e4f408a 100644
+--- a/src/include/internal/host/nvshmemi_mem_transport.hpp
+++ b/src/include/internal/host/nvshmemi_mem_transport.hpp
+@@ -36,6 +36,13 @@ class nvshmemi_mem_p2p_transport final {
+             return p2p_objref_;
+         }
+     }
+    // eep-dev
+    static void destroy_instance(void) {
+        if (p2p_objref_ != nullptr) {
+            delete p2p_objref_;
+            p2p_objref_ = nullptr;
+        }
+    }
+ 
+     void print_mem_handle(int pe_id, int transport_idx, nvshmemi_symmetric_heap &obj);
+ 
+@@ -87,6 +94,14 @@ class nvshmemi_mem_remote_transport final {
+         }
+     }
+ 
+    // eep-dev
+    static void destroy_instance(void) {
+        if (remote_objref_ != nullptr) {
+            delete remote_objref_;
+            remote_objref_ = nullptr;
+        }
+    }
+
+     int gather_mem_handles(nvshmemi_symmetric_heap &obj, uint64_t heap_offset, size_t size);
+     /* On-demand registration and release of memory */
+     int register_mem_handle(nvshmem_mem_handle_t *local_handles, int transport_idx,
+diff --git a/src/modules/bootstrap/uid/bootstrap_uid.cpp b/src/modules/bootstrap/uid/bootstrap_uid.cpp
+index a1fa748..788fa96 100644
+--- a/src/modules/bootstrap/uid/bootstrap_uid.cpp
+++ b/src/modules/bootstrap/uid/bootstrap_uid.cpp
+@@ -630,6 +630,11 @@ int nvshmemi_bootstrap_plugin_pre_init(bootstrap_handle_t* handle, const int abi
+     // Discover the network for bootstrap, if not done previously.
+     // This code needs to be stateful to be able to be called multiple times by the caller
+     BOOTSTRAP_CHECK(bootstrap_net_init());
+    // eep-dev
+    if (handle->pre_init_ops != nullptr) {
+        BOOTSTRAP_PTR_FREE(handle->pre_init_ops);
+        handle->pre_init_ops = nullptr;
+    }
+     if (handle->pre_init_ops == nullptr) {
+         BOOTSTRAP_CALLOC(&handle->pre_init_ops, 1);
+         handle->pre_init_ops->get_unique_id = bootstrap_get_unique_id;
+-- 
+2.43.0
+
--- a/tools/ep_kernels/elastic_ep/install_eep_libraries.sh
+++ b/tools/ep_kernels/elastic_ep/install_eep_libraries.sh
@@ -0,0 +1,86 @@
+#!/bin/bash
+
+set -ex
+
+# Default workspace directory
+WORKSPACE=$(pwd)/eep_kernels_workspace
+INSTALL_NVSHMEM=true
+
+# Parse command line arguments
+while getopts "w:n" opt; do
+  case $opt in
+    w)
+      WORKSPACE="$OPTARG"
+      ;;
+    n)
+      INSTALL_NVSHMEM=false
+      ;;
+    \?)
+      echo "Invalid option: -$OPTARG" >&2
+      exit 1
+      ;;
+  esac
+done
+
+if [ ! -d "$WORKSPACE" ]; then
+    mkdir -p $WORKSPACE
+fi
+
+
+# install dependencies if not installed
+pip3 install cmake torch ninja
+
+# build nvshmem
+pushd $WORKSPACE
+# Reset NVSHMEM build if requested
+if [ "$INSTALL_NVSHMEM" = true ]; then
+    mkdir -p nvshmem_src
+    wget https://developer.download.nvidia.com/compute/redist/nvshmem/3.2.5/source/nvshmem_src_3.2.5-1.txz
+    tar -xvf nvshmem_src_3.2.5-1.txz -C nvshmem_src --strip-components=1
+    pushd nvshmem_src
+    wget https://github.com/deepseek-ai/DeepEP/raw/main/third-party/nvshmem.patch
+    git init
+    git apply -vvv nvshmem.patch
+    git apply --reject --whitespace=fix ../../eep_nvshmem.patch 
+else
+    pushd nvshmem_src
+fi
+
+# assume CUDA_HOME is set correctly
+if [ -z "$CUDA_HOME" ]; then
+    echo "CUDA_HOME is not set, please set it to your CUDA installation directory."
+    exit 1
+fi
+
+# disable all features except IBGDA
+export NVSHMEM_IBGDA_SUPPORT=1
+
+export NVSHMEM_SHMEM_SUPPORT=0
+export NVSHMEM_UCX_SUPPORT=0
+export NVSHMEM_USE_NCCL=0
+export NVSHMEM_PMIX_SUPPORT=0
+export NVSHMEM_TIMEOUT_DEVICE_POLLING=0
+export NVSHMEM_USE_GDRCOPY=0
+export NVSHMEM_IBRC_SUPPORT=0
+export NVSHMEM_BUILD_TESTS=0
+export NVSHMEM_BUILD_EXAMPLES=0
+export NVSHMEM_MPI_SUPPORT=0
+export NVSHMEM_BUILD_HYDRA_LAUNCHER=0
+export NVSHMEM_BUILD_TXZ_PACKAGE=0
+export NVSHMEM_TIMEOUT_DEVICE_POLLING=0
+
+cmake -G Ninja -S . -B $WORKSPACE/nvshmem_build/ -DCMAKE_INSTALL_PREFIX=$WORKSPACE/nvshmem_install
+cmake --build $WORKSPACE/nvshmem_build/ --target install
+
+popd
+
+export CMAKE_PREFIX_PATH=$WORKSPACE/nvshmem_install:$CMAKE_PREFIX_PATH
+
+# build and install pplx, require pytorch installed
+pushd $WORKSPACE
+git clone https://github.com/ppl-ai/pplx-kernels
+cd pplx-kernels
+# see https://github.com/pypa/pip/issues/9955#issuecomment-838065925
+# PIP_NO_BUILD_ISOLATION=0 disables build isolation
+PIP_NO_BUILD_ISOLATION=0 TORCH_CUDA_ARCH_LIST=9.0a+PTX pip install . --no-deps -v
+