docker: add manifest to versioned docker releases

[ishandhanani]

Goal is to run docker pull lmsysorg/sglang:v0.5.3 and get the proper image corresponding to your machine. Also added in a latest as well @zhyncs

[gemini-code-assist]

Note

Gemini is unable to generate a summary for this pull request due to the file types involved not being currently supported.

[kyleliang-nv]

Can you also fix https://github.com/sgl-project/sglang/blob/main/docker/Dockerfile#L102 to add another OR condition for all_aarch64.
This line (https://github.com/sgl-project/sglang/blob/main/docker/Dockerfile#L115) should also be changed to
check BUILD_TYPE for either blackwell, blackwell_aarch64, or all_aarch64

[kyleliang-nv]

Another issue is this (https://github.com/sgl-project/sglang/blob/main/docker/Dockerfile#L143-L149) will fail if CUDA_VERISON=12.6.1.
You will want to set env FLASH_MLA_DISABLE_SM100=1 for CUDA_VERSION less than 12.9.1, otherwise will get error like this.

#22 4.375 Processing /sgl-workspace/flash-mla
#22 4.375   Preparing metadata (setup.py): started
#22 4.375   Running command python setup.py egg_info
#22 4.833   /usr/local/lib/python3.12/dist-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
#22 4.833     import pynvml  # type: ignore[import]
#22 5.573   W1006 05:49:58.407000 114 torch/utils/cpp_extension.py:118] No CUDA runtime is found, using CUDA_HOME='/usr/local/cuda'
#22 5.612   Traceback (most recent call last):
#22 5.612     File "<string>", line 2, in <module>
#22 5.612     File "<pip-setuptools-caller>", line 35, in <module>
#22 5.612     File "/sgl-workspace/flash-mla/setup.py", line 93, in <module>
#22 5.612       ] + get_features_args() + get_arch_flags() + get_nvcc_thread_args(),
#22 5.612                                 ^^^^^^^^^^^^^^^^
#22 5.612     File "/sgl-workspace/flash-mla/setup.py", line 39, in get_arch_flags
#22 5.613       assert DISABLE_SM100, "sm100 compilation for Flash MLA requires NVCC 12.9 or higher. Please set FLASH_MLA_DISABLE_SM100=1 to disable sm100 compilation, or update your environment."
#22 5.613              ^^^^^^^^^^^^^
#22 5.613   AssertionError: sm100 compilation for Flash MLA requires NVCC 12.9 or higher. Please set FLASH_MLA_DISABLE_SM100=1 to disable sm100 compilation, or update your environment.

[zhyncs]

we don't need -cu129 suffix

[merrymercy]

maybe still keep it somewhere?
Later we might have -cu129, -cu130 , which cannot be unified.

but for lmsys/sglang:latest, we can remove it and have a default cuda verison

[zhyncs]

e.g.
lmsysorg/sglang:v0.5.3 for cu129
lmsysorg/sglang:v0.5.3-cu130 for cu130

how about this

[ishandhanani]

This makes sense. I think default for almost all users will be cu129 therefore we don't need it. As we push on cu130 lets keep it tagged as such

[zhyncs]

I am actually worried that users may be troubled, for example, this image can also be used on cu126/cu128/cu129 hopper, not just cu129.