DeepSpeed

27360990 - Add klxpu (KUNLUNXIN XPU) self-hosted CI for the v1 unit suite (#8923)

Add klxpu (KUNLUNXIN XPU) self-hosted CI for the v1 unit suite (#8923) Run the tests/unit/v1 suite on a self-hosted KUNLUNXIN XPU runner inside the vendor container, one test file per subprocess (torch_xmlir inits CUDA at import, so pytest --forked is unusable). A blocklist keeps new tests running by default: - ignore.txt whole files/dirs that hang or wholly fail - deselect.txt individual node ids that fail Environment handling, all per-machine via Actions variables so the same workflow works on hosts with different networks and storage: - proxy (KLXPU_HTTP_PROXY/KLXPU_HTTPS_PROXY): unset means direct access. - Hugging Face (KLXPU_HF_ENDPOINT + HF_HUB_DISABLE_XET): route model downloads through a reachable mirror and force plain-HTTPS LFS, since the Xet CDN is unreachable behind a mirror and otherwise stalls. Ignored files cover known KLX-stack limitations: BKCL/XCCL single-communicator hangs (autotp/moe/muon/zero/ulysses), Triton unsupported, torch.compile intercepted by torch_xmlir, and GDS unavailable. --------- Signed-off-by: guhangsong <guhangsong@outlook.com>