master

Xxuxingleifeat: support Qwen3.5 and Qwen3-VL-MoE training

427916e6创建于 7月10日历史提交

文件	最后提交记录	最后更新时间
pre-commit	Support pre-commint and update code-style rules.	6 个月前
build_custom_ops.sh	refactor: simplify native build env var values and add strict build mode	3 个月前
build_multicore.sh	refactor: simplify native build env var values and add strict build mode	3 个月前
build_multicore_local.sh	Refactor the multicore code	4 个月前
build_symmetric_memory.sh	refactor: simplify native build env var values and add strict build mode	3 个月前
check_gcc_version.sh	fix(packaging): tag wheel per python/arch and unify cxx11 abi per backend The wheel was published as py3-none-any while its package_data ships prebuilt .so files linked against a specific CPython ABI and CPU arch. Installing the wheel under a different Python minor version triggered 'Python version mismatch' at import. - setup.py: add BinaryDistribution + BdistWheel so the wheel is tagged cp3xy-cp3xy-linux_{aarch64|x86_64}; narrow python_requires to [3.10, 3.13); enforce GCC in [7.3.0, 11.3.0] mirroring mindspore; mirror native build outputs from build/lib into the platform-tagged build dir so the symmetric_memory .so files survive into the wheel; drop the broken setup_requires=['ninja>=1.10.0'] (ninja is only used by the opt-in torch extension path). - Per-backend CXX11 ABI: mindspore-link CMakeLists and CustomOpBuilder cflags pin =0 (matches mindspore); torch-link CMakeLists and the NpuExtension setup.py pin =1 (matches PyTorch/torch_npu defaults since 2.7). The symmetric_memory ops public layer keeps =0 with a POD interface so both wrappers link it safely (mirrors CANN's intf_pub policy). - scripts/check_gcc_version.sh: new shared host-GCC policy sourced by the three build scripts. The dead USE_CXX11_ABI shell switch and matching cmake args are removed. - README zh/en: rewrite the install/build section into general, build- only, and runtime subsections; declare the host glibc lower-bound expectation. Pre-existing pylint/markdownlint debt in setup.py and the READMEs is left untouched and the commit uses --no-verify per repo guidance; a separate cleanup PR can address it.	4 个月前
pylint_hyperparallel.py	chore: make agent tooling assistant-agnostic for any AI coding assistant Make the in-repo agent tooling work with any AI coding assistant instead of being tied to a specific vendor, so contributors can use whichever assistant they prefer.	4 个月前
train_lm.py	feat: support Qwen3.5 and Qwen3-VL-MoE training Add dense, MoE, and vision-language model integration with checkpoint conversion, data loading, and unified trainer execution. Correct TP, CP, EP, PP, FSDP, and HSDP synchronization semantics so supported BF16 single-card and hybrid-parallel configurations follow the same training behavior.	3 个月前
train_vl.py	feat: support Qwen3.5 and Qwen3-VL-MoE training Add dense, MoE, and vision-language model integration with checkpoint conversion, data loading, and unified trainer execution. Correct TP, CP, EP, PP, FSDP, and HSDP synchronization semantics so supported BF16 single-card and hybrid-parallel configurations follow the same training behavior.	3 个月前