Use libfabric fork with CUDA GDR hotfix

Author: Madeeks

examples/container/comm-fwk/Containerfile

@@ -83,7 +83,8 @@ RUN git clone --branch ${libcxi_version} --depth 1 https://github.com/HewlettPac
 
 # Install libfabric
 ARG libfabric_version=2.1.0
-RUN git clone --branch v${libfabric_version} --depth 1 https://github.com/ofiwg/libfabric.git \
+#RUN git clone --branch v${libfabric_version} --depth 1 https://github.com/ofiwg/libfabric.git \
+RUN git clone --branch cuda_gdrcopy_unregister_fix --depth 1 https://github.com/Madeeks/libfabric.git \
     && cd libfabric \
     && ./autogen.sh \
     && ./configure --prefix=/usr --with-cuda=/usr/local/cuda --enable-cuda-dlopen --enable-gdrcopy-dlopen --enable-xpmem=/usr --enable-cxi --enable-lnx --enable-efa \

examples/container/comm-fwk/README.md

@@ -13,7 +13,7 @@ Builds of the image are currently hosted on the Quay.io registry at https://quay
 - CUDA 12.8.1 (including other dependencies like NCCL from the NVIDIA Docker Hub image)
 - GDRCopy 2.5
 - XPMEM (commit 0d0bad4e1d07b38d53ecc8f20786bb1328c446da - corresponds to version 2.6.5-36 in Spack)
-- Libfabric 2.1.0 with the following providers explicitly enabled:
+- A patched Libfabric 2.1.0-dev (see notes) with the following providers explicitly enabled:
     - CXI
     - AWS EFA
     - LINKx
@@ -25,6 +25,7 @@ Builds of the image are currently hosted on the Quay.io registry at https://quay
 ## Notes
 
 - This image and its derivatives are self-sufficient with respect to Slingshot connectivity, and do not require hooks to inject a custom CXI stack from the host.
+- The libfabric in this image contains an experimental fix for CUDA GDR support (more details [here](https://github.com/ofiwg/libfabric/issues/10865#issuecomment-2735866065)); the fix is applied on top of libfabric's `main` commit at the moment of writing (commit faf13301a4a9628b6c9a28a06d936258c6d368af), and the code is available from [this forked branch](https://github.com/Madeeks/libfabric/tree/cuda_gdrcopy_unregister_fix).
 - The libfabric EFA provider is included to leave open the possibility to experiment with derived images on AWS infrastructure as well.
 - The libfabric LINKx provider is included to allow for experimentation.
 - Although only the libfabric framework and its CXI provider are required to support the Slingshot network, this image also packages the UCX communication framework to allow building a broader set of software (e.g. some OpenSHMEM implementations) and supporting optimized Infiniband communication as well.