Adpat mori to RoCM 714 container

[QizhouZhang97]

Moves CI to rocm/pytorch:rocm7.14_ubuntu24.04_py3.12_pytorch_release_2.12.0, dropping 7.2.4.

The 7.14 image has no /opt/rocm — ROCm comes from the rocm-sdk-core wheel, which lacks CMake configs, and its hipconfig shim shadows the real one and ignores HIP_PATH.

• CMakeLists.txt: resolve ROCm before project(... HIP) without hipconfig; pin CMAKE_HIP_COMPILER; RPATH the ROCm libdir so no LD_LIBRARY_PATH is needed
• Dockerfile.dev: apt ROCm 7.14 dev stack, /usr/lib64/libc.so symlink
• Dockerfile.cco: pin flydsl<0.3
• nightly: py3.10 dropped (no 7.14 image)
• EPv2 gather hangs on 7.14 — a ROCm regression (passes on 7.2.4); gated off, scatter still covered.

Tested on 8×MI355X: CCO, SDMA, EPv1 (335 passed), mori check all green.

[jhchouuu]

现在应该已经支持flydsl0.3.0了，可以对应升级一下

[QizhouZhang97]

0.3.0貌似并没有bufferops?

[jhchouuu]

Two things in docker/Dockerfile.dev worth fixing before this lands. Both hit every base image, including the 7.2.4 py3.10 nightly lane that now builds from this same file.

1. ENV LD_LIBRARY_PATH=/opt/rocm/lib overwrites instead of appending (line 55). The Dockerfile.cco it replaces had :${LD_LIBRARY_PATH}, so whatever the base image set is now dropped.

ENV LD_LIBRARY_PATH=/opt/rocm/lib:${LD_LIBRARY_PATH}

2. ENV LD_PRELOAD=... is unconditional, but the install that provides the file is not (line 61). When the base already ships a ROCm dev stack the apt step is skipped, yet the preload still points at /opt/rocm/lib/libhsa-runtime64.so.1. If that path doesn't exist, every command in the image prints ld.so: object ... cannot be preloaded ...: ignored. Suggest gating it the same way as the install (ARG, or set it from the else branch).

Unrelated: the PR body still says "EPv2 gather hangs on 7.14 — ... gated off, scatter still covered", but e59ff2b removed the gating and later commits cleaned up the comments. Worth updating so the squashed commit message doesn't carry it, and it'd be useful to note whether gather turned out to be fixed on 7.14 or the original diagnosis was off.

[QizhouZhang97]

Two things in docker/Dockerfile.dev worth fixing before this lands. Both hit every base image, including the 7.2.4 py3.10 nightly lane that now builds from this same file.

1. ENV LD_LIBRARY_PATH=/opt/rocm/lib overwrites instead of appending (line 55). The Dockerfile.cco it replaces had :${LD_LIBRARY_PATH}, so whatever the base image set is now dropped.

ENV LD_LIBRARY_PATH=/opt/rocm/lib:${LD_LIBRARY_PATH}

2. ENV LD_PRELOAD=... is unconditional, but the install that provides the file is not (line 61). When the base already ships a ROCm dev stack the apt step is skipped, yet the preload still points at /opt/rocm/lib/libhsa-runtime64.so.1. If that path doesn't exist, every command in the image prints ld.so: object ... cannot be preloaded ...: ignored. Suggest gating it the same way as the install (ARG, or set it from the else branch).

Unrelated: the PR body still says "EPv2 gather hangs on 7.14 — ... gated off, scatter still covered", but e59ff2b removed the gating and later commits cleaned up the comments. Worth updating so the squashed commit message doesn't carry it, and it'd be useful to note whether gather turned out to be fixed on 7.14 or the original diagnosis was off.

Two things in docker/Dockerfile.dev worth fixing before this lands. Both hit every base image, including the 7.2.4 py3.10 nightly lane that now builds from this same file.

1. ENV LD_LIBRARY_PATH=/opt/rocm/lib overwrites instead of appending (line 55). The Dockerfile.cco it replaces had :${LD_LIBRARY_PATH}, so whatever the base image set is now dropped.

ENV LD_LIBRARY_PATH=/opt/rocm/lib:${LD_LIBRARY_PATH}

2. ENV LD_PRELOAD=... is unconditional, but the install that provides the file is not (line 61). When the base already ships a ROCm dev stack the apt step is skipped, yet the preload still points at /opt/rocm/lib/libhsa-runtime64.so.1. If that path doesn't exist, every command in the image prints ld.so: object ... cannot be preloaded ...: ignored. Suggest gating it the same way as the install (ARG, or set it from the else branch).

Unrelated: the PR body still says "EPv2 gather hangs on 7.14 — ... gated off, scatter still covered", but e59ff2b removed the gating and later commits cleaned up the comments. Worth updating so the squashed commit message doesn't carry it, and it'd be useful to note whether gather turned out to be fixed on 7.14 or the original diagnosis was off.

移除容器中对ld_library_path, ld_preload的依赖