feat: merge python_native_v2 into main with review fixes

.bazelignore

@@ -1,2 +1,3 @@
 internal_source/rdma
-open_source
+open_source
+bazel-github-opensource

.bazelrc

@@ -2,8 +2,6 @@ common --experimental_downloader_config=bazel/bazel_downloader.cfg
 fetch:downloader --experimental_remote_downloader=grpc://com.taobao.search.bazel.server:9092 --remote_cache=grpc://com.taobao.search.bazel.server:9092
 
 
-build --python_top=//:python310 --incompatible_use_python_toolchains=false # force use /opt/conda310/bin/python3
-
 build --spawn_strategy=local # avoid nvcc conflicts
 build --action_env PYTHON_BIN_PATH="/opt/conda310/bin/python3"
 build --cxxopt="-std=c++17" --copt="-DGTEST_USE_OWN_TR1_TUPLE=0" --copt="-DEIGEN_MAX_CPP_VER=11"
@@ -30,8 +28,8 @@ build --define=xft_use_icx=true
 build --define=frontend=false
 build --linkopt -ldl # exception backtrace
 build --define=NVSHMEM_DIR=/
-build --action_env PATH="/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/tmp/cache_bust_20260321"
-build --host_action_env PATH="/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/tmp/cache_bust_20260321"
+build --action_env PATH="/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/tmp/cache_bust_20260423"
+build --host_action_env PATH="/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/tmp/cache_bust_20260423"
 
 build:cuda --copt="-DENABLE_BF16=1"
 build:cuda --copt="-DBUILD_CUTLASS_MIXED_GEMM=ON"
@@ -56,6 +54,9 @@ build:cuda --action_env TF_NCCL_VERSION="2"
 build:cuda --action_env CUDNN_INSTALL_PATH="/usr/local/cuda/"
 
 # 6.0 = P100, 7.0 = V100, 7.5 = T4, 8.6 = A10, 8.0 = A100 8.9 = L, 9.0 = H800
+# NOTE: cuda12 is a shared base config. Do not use it directly; it lacks the
+# platform-specific torch_deps/torch_libs selection. Use cuda12_6, cuda12_9, or
+# cuda12_9_arm instead.
 build:cuda12 --config=cuda
 build:cuda12 --action_env TF_CUDA_COMPUTE_CAPABILITIES="7.0,7.5,8.0,8.6,8.9,9.0"
 build:cuda12 --host_action_env TF_CUDA_COMPUTE_CAPABILITIES="7.0,7.5,8.0,8.6,8.9,9.0"
@@ -168,14 +169,14 @@ build:rocm --copt="-D_GLIBCXX_USE_CXX11_ABI=1"
 #build:rocm --action_env HIPBLASLT_LOG_MASK="32" # to log blasLt
 #build:rocm --action_env ROCBLAS_LAYER="4"
 #build:rocm --copt="-DENABLE_PROF=1"
-build:rocm --action_env LD_LIBRARY_PATH="/opt/rh/gcc-toolset-12/root/usr/lib64:/lib64:/opt/conda310/lib/:/opt/rocm/lib/:/opt/taobao/java/jre/lib/amd64/server/:/opt/amdgpu/lib64/"
-build:rocm --host_action_env LD_LIBRARY_PATH="/opt/rocm/lib/:/opt/taobao/java/jre/lib/amd64/server/:/opt/amdgpu/lib64/"
-build:rocm --action_env PATH="/opt/rh/gcc-toolset-12/root/usr/bin:/opt/conda310/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/tmp/cache_bust_20260321"
-build:rocm --host_action_env PATH="/opt/rh/gcc-toolset-12/root/usr/bin:/opt/conda310/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/tmp/cache_bust_20260321"
+build:rocm --action_env LD_LIBRARY_PATH="/opt/rh/gcc-toolset-12/root/usr/lib64:/lib64:/opt/rocm/lib/:/opt/amdgpu/lib64/"
+build:rocm --host_action_env  LD_LIBRARY_PATH="/opt/rh/gcc-toolset-12/root/usr/lib64:/lib64:/opt/rocm/lib/:/opt/amdgpu/lib64/"
+build:rocm --action_env PATH="/opt/rh/gcc-toolset-12/root/usr/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/tmp/cache_bust_20260423"
+build:rocm --host_action_env PATH="/opt/rh/gcc-toolset-12/root/usr/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/tmp/cache_bust_20260423"
 build:rocm --copt="-DUSE_ROCM=1"
 build:rocm --copt="-DFORCE_NVSHMEM_API=1"
 build:rocm --copt="-DFLASHINFER_CUB_SUBTRACTLEFT_DEFINED"  # ROCm 7.2 hipcub uses SubtractLeft instead of FlagHeads
-test:rocm --test_env="/opt/rocm/lib/:/opt/taobao/java/jre/lib/amd64/server/:/opt/amdgpu/lib64/"
+
 
 build:arm --copt="-DENABLE_BF16=1"
 build:arm --define=using_cuda=false --define=using_cuda_nvcc=false
@@ -213,20 +214,41 @@ build:asan --copt -fPIC # for "fix relocation truncated to fit: R_X86_64_32 agai
 build:asan --copt -fdebug-types-section # for "fix relocation truncated to fit: R_X86_64_32 against `.debug_info'" collect2 error
 build:asan --linkopt -fsanitize=address
 
-test:rocm --test_env PATH="/opt/rocm/bin:/opt/rh/gcc-toolset-12/root/usr/bin:/opt/conda310/bin:/opt/conda310/condabin:/usr/share/Modules/bin:/sbin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/usr/X11R6/bin:/opt/cmake/cmake-3.26.4/bin"
-test:rocm --test_env LD_LIBRARY_PATH="/opt/rh/gcc-toolset-12/root/usr/lib64:/opt/rocm/lib:/opt/conda310/lib/:/usr/lib64:/opt/amdgpu/lib64"
-test --test_env LD_LIBRARY_PATH="/opt/rocm/lib:/opt/conda310/lib/:/usr/local/nvidia/lib64:/usr/lib64:/usr/local/cuda/lib64:/opt/amdgpu/lib64:/usr/local/cuda/extras/CUPTI/lib64"
 test --test_env OMP_NUM_THREADS=8
 test --test_env FT_SERVER_TEST="1"
 test --test_env LOG_LEVEL="INFO"
 test --test_env NCCL_DEBUG="INFO"
 # for dp scatter add stable result
 test --test_env TZ=Asia/Shanghai
+# Global minimal test LD_LIBRARY_PATH so platforms without a dedicated stanza
+# (cpu / arm / ppu) can still find the conda310 python/torch runtime libs.
+# The cuda/rocm/compat stanzas below set fuller paths and override this for
+# their --config (single-valued --test_env: last applied wins).
+test --test_env LD_LIBRARY_PATH="/opt/conda310/lib/:/usr/lib64:/lib64"
 # for flashinfer find cuda; scope to CUDA-specific test runs
 test:cuda --test_env CUDA_HOME=/usr/local/cuda
+
+# NOTE: /opt/conda310/bin MUST stay on PATH — many cpp tests use
+# `#!/usr/bin/env python3` shebangs / subprocess invocations and conda310
+# is the only python3 in the dev images.  Dropping it caused
+# `/usr/bin/env: 'python3': No such file or directory` to fail all 54 cpp-ut
+# tests on REAPI workers (PR 965 run 39271211 cpp-ut job).
+test:cuda --test_env PATH="/opt/conda310/bin:/usr/local/nvidia/bin:/usr/local/cuda/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin"
+test:cuda --test_env LD_LIBRARY_PATH="/opt/conda310/lib/:/usr/local/nvidia/lib64:/usr/lib64:/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64"
+
+# ROCm: gcc-toolset-12 lib64 MUST come first so dlopen of aiter JIT-built
+# .so finds GCC 12 libstdc++ (has __cxx11::ostringstream::str() const).
+# conda310 lib AFTER gcc-toolset-12 to avoid clobbering libstdc++ resolution
+# but BEFORE rocm/amdgpu so python finds its own runtime.
+test:rocm --test_env LD_LIBRARY_PATH="/opt/rh/gcc-toolset-12/root/usr/lib64:/opt/conda310/lib/:/lib64:/opt/rocm/lib/:/opt/amdgpu/lib64/"
+test:rocm --test_env PATH="/opt/rh/gcc-toolset-12/root/usr/bin:/opt/conda310/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/tmp/cache_bust_20260423"
+
 # compat for low driver < 535
 test:compat --test_env LD_LIBRARY_PATH="/opt/rocm/lib:/opt/conda310/lib/:/usr/local/cuda/compat/:/usr/local/nvidia/lib64:/usr/lib64:/usr/local/cuda/lib64:/opt/amdgpu/lib64:/usr/local/cuda/extras/CUPTI/lib64:/usr/local/cuda/lib64/stubs"
 
-# internal source overrides (only loaded when internal_source symlink exists)
-try-import %workspace%/internal_source/.internal_bazelrc
-try-import %workspace%/internal_source/.cicd_bazelrc
+# internal source overrides (only loaded in the internal monorepo checkout)
+try-import %workspace%/../internal_source/.internal_bazelrc
+# torch path from host environment (generated by setup.py)
+try-import %workspace%/.torch_bazelrc
+# always use for aone ci
+try-import %workspace%/../internal_source/.cicd_bazelrc

.gitignore

@@ -1,12 +1,26 @@
 *~
 *.o
 __pycache__/
+
+# Local venv: use $HOME/venvs/<name> (see scripts/create_venv.sh). .venv/ is uv/IDE default when
+# running `uv run`/`uv sync` in-tree without UV_PROJECT_ENVIRONMENT — ignore so it never gets committed.
+.venv/
+venv/
 .vscode
 .idea
 ./translation
 .cache
 *.npy
 *.pth
+.torch_bazelrc
+*.egg-info/
+
+# Build artifacts - entire libs/ dir is generated by setup.py build_ext
+rtp_llm/libs/
+# Generated .pyi stubs (created by setup.py build_ext via pybind11_stubgen)
+rtp_llm/ops/**/*.pyi
+smoke_actual/
+main_logs/
 !tests/data/**/*.npy
 /notebooks
 **/.ipynb_checkpoints/
@@ -42,8 +56,11 @@ package/rtp_llm*.whl
 package/maga-ci.Dockerfile
 rtp_llm/test-data/*
 
-model_rpc_service_pb2.py
-model_rpc_service_pb2_grpc.py
+# Protobuf / gRPC generated Python (grpc_tools.protoc / setup.py build_ext)
+rtp_llm/cpp/model_rpc/proto/*_pb2.py
+rtp_llm/cpp/model_rpc/proto/*_pb2_grpc.py
+rtp_llm/test/remote_tests/*_pb2.py
+rtp_llm/test/remote_tests/*_pb2_grpc.py
 rtp_llm/cpp/deep_gemm/cutlass_hdr
 
 cuda_profiler*.json
@@ -69,6 +86,7 @@ git.diff
 
 *.mo
 *build*/
+!_build/
 .doctrees/
 node_modules/
 docs/index/dist/

3rdparty/BUILD

@@ -1,5 +1,4 @@
-load("//:def.bzl", "rpm_library", "copts", "cuda_copts")
-load("@arch_config//:arch_select.bzl", "torch_deps")
+load("//:def.bzl", "copts", "cuda_copts")
 package(default_visibility = ["//visibility:public"])
 
 cc_library(
@@ -13,7 +12,7 @@ cc_library(
         "@local_config_cuda//cuda:cuda_driver",
     ],
     copts = copts(),
-    visibility = ["//visibility:public"],  
+    visibility = ["//visibility:public"],
 )
 
 cc_library(

3rdparty/contextFusedMultiHeadAttention/BUILD

@@ -1,6 +1,6 @@
 load("//:def.bzl", "rpm_library", "copts", "cuda_copts",)
 load("@local_config_cuda//cuda:build_defs.bzl", "cuda_default_copts_without_arch", "if_cuda")
-load("@arch_config//:arch_select.bzl", "torch_deps")
+load("@rtp_llm//bazel:defs.bzl", "torch_deps")
 load("@rules_cc//examples:experimental_cc_shared_library.bzl", "cc_shared_library")
 package(default_visibility = ["//visibility:public"])
 

3rdparty/contextFusedMultiHeadAttentionSm70/BUILD

@@ -1,5 +1,5 @@
 load("//:def.bzl", "rpm_library", "copts", "cuda_copts",)
-load("@arch_config//:arch_select.bzl", "torch_deps")
+load("@rtp_llm//bazel:defs.bzl", "torch_deps")
 package(default_visibility = ["//visibility:public"])
 
 cc_library(