From 98d5574de5182b9736767cda9166fede9e6770d5 Mon Sep 17 00:00:00 2001 From: Misha Chornyi Date: Wed, 15 Oct 2025 08:02:01 -0700 Subject: [PATCH 01/17] Add libnvshmem installation --- build.py | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-) diff --git a/build.py b/build.py index d02cf14200..690558b61a 100755 --- a/build.py +++ b/build.py @@ -1318,8 +1318,17 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach backend_dependencies = "libgomp1" # libgfortran5 is needed by pytorch backend on ARM - if ("pytorch" in backends) and (target_machine == "aarch64"): - backend_dependencies += " libgfortran5" + if "pytorch" in backends: + if target_machine == "aarch64": + backend_dependencies += " libgfortran5" + + if "CUDA_VERSION" in os.environ: + cuda_version_major = int(os.environ["CUDA_VERSION"].split(".")[0]) + if cuda_version_major >= 13: + backend_dependencies += " libnvshmem3-cuda-{}".format( + cuda_version_major + ) + # openssh-server is needed for fastertransformer if "fastertransformer" in backends: backend_dependencies += " openssh-server" From 7c2586ee06158c86a8ff4ede5a7782e9c93ee932 Mon Sep 17 00:00:00 2001 From: Misha Chornyi Date: Wed, 15 Oct 2025 08:26:32 -0700 Subject: [PATCH 02/17] Set explicit dependencies --- build.py | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/build.py b/build.py index 690558b61a..7c6542cca7 100755 --- a/build.py +++ b/build.py @@ -1322,12 +1322,12 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach if target_machine == "aarch64": backend_dependencies += " libgfortran5" - if "CUDA_VERSION" in os.environ: - cuda_version_major = int(os.environ["CUDA_VERSION"].split(".")[0]) - if cuda_version_major >= 13: - backend_dependencies += " libnvshmem3-cuda-{}".format( - cuda_version_major - ) + if ( + target_platform != "igpu" + or target_platform != "windows" + or target_platform != "rhel" + ): + backend_dependencies += " libnvshmem3-cuda-13" # openssh-server is needed for fastertransformer if "fastertransformer" in backends: From dec55119abfe3baad0f26dde58fbbb0c7f845d52 Mon Sep 17 00:00:00 2001 From: Misha Chornyi Date: Wed, 15 Oct 2025 09:24:59 -0700 Subject: [PATCH 03/17] Fix condition --- build.py | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-) diff --git a/build.py b/build.py index 7c6542cca7..83b44ff7e1 100755 --- a/build.py +++ b/build.py @@ -1322,11 +1322,7 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach if target_machine == "aarch64": backend_dependencies += " libgfortran5" - if ( - target_platform != "igpu" - or target_platform != "windows" - or target_platform != "rhel" - ): + if target_platform() not in ["igpu", "windows", "rhel"]: backend_dependencies += " libnvshmem3-cuda-13" # openssh-server is needed for fastertransformer From 6bfa68f00b691c676ae6317e226690229e91f2c3 Mon Sep 17 00:00:00 2001 From: Misha Chornyi Date: Wed, 15 Oct 2025 10:29:49 -0700 Subject: [PATCH 04/17] Updating build.py --- build.py | 11 ++++++----- 1 file changed, 6 insertions(+), 5 deletions(-) diff --git a/build.py b/build.py index 83b44ff7e1..28be0e10ea 100755 --- a/build.py +++ b/build.py @@ -1318,12 +1318,13 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach backend_dependencies = "libgomp1" # libgfortran5 is needed by pytorch backend on ARM - if "pytorch" in backends: - if target_machine == "aarch64": - backend_dependencies += " libgfortran5" + if ("pytorch" in backends) and (target_machine == "aarch64"): + backend_dependencies += " libgfortran5" - if target_platform() not in ["igpu", "windows", "rhel"]: - backend_dependencies += " libnvshmem3-cuda-13" + # libnvshmem3-cuda-13 is needed by pytorch backend but will be required by vLLM and TensorRT-LLM + # on platforms other than IGPU we still using CUDA 12 it means we need to add this conditionally + if target_platform() not in ["igpu", "windows", "rhel"]: + backend_dependencies += " libnvshmem3-cuda-13" # openssh-server is needed for fastertransformer if "fastertransformer" in backends: From d725b73d350d37ba2d032b9f9e389205db14ab72 Mon Sep 17 00:00:00 2001 From: Misha Chornyi Date: Wed, 15 Oct 2025 08:58:00 -0700 Subject: [PATCH 05/17] Set vLLM version 0.10.2 --- build.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/build.py b/build.py index 28be0e10ea..4fd7347d79 100755 --- a/build.py +++ b/build.py @@ -78,7 +78,7 @@ "ort_openvino_version": "2025.3.0", "standalone_openvino_version": "2025.3.0", "dcgm_version": "4.4.0-1", - "vllm_version": "0.10.1.1", + "vllm_version": "0.10.2", "rhel_py_version": "3.12.3", } From 37ff94e22c13c1e2bee0da59e1a59edcf7bed20d Mon Sep 17 00:00:00 2001 From: Misha Chornyi Date: Wed, 15 Oct 2025 11:46:34 -0700 Subject: [PATCH 06/17] Restructure instructions for Dockerfile --- build.py | 36 +++++++++++++++++++----------------- 1 file changed, 19 insertions(+), 17 deletions(-) diff --git a/build.py b/build.py index 4fd7347d79..618ad9a664 100755 --- a/build.py +++ b/build.py @@ -78,7 +78,7 @@ "ort_openvino_version": "2025.3.0", "standalone_openvino_version": "2025.3.0", "dcgm_version": "4.4.0-1", - "vllm_version": "0.10.2", + "vllm_version": "0.10.1.1", "rhel_py_version": "3.12.3", } @@ -879,6 +879,7 @@ def install_dcgm_libraries(dcgm_version, target_machine): else: if target_machine == "aarch64": return """ +ENV DEBIAN_FRONTEND=noninteractive ENV DCGM_VERSION {} # Install DCGM. Steps from https://developer.nvidia.com/dcgm#Downloads RUN curl -o /tmp/cuda-keyring.deb \\ @@ -894,6 +895,7 @@ def install_dcgm_libraries(dcgm_version, target_machine): ) else: return """ +ENV DEBIAN_FRONTEND=noninteractive ENV DCGM_VERSION {} # Install DCGM. Steps from https://developer.nvidia.com/dcgm#Downloads RUN curl -o /tmp/cuda-keyring.deb \\ @@ -1323,7 +1325,7 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach # libnvshmem3-cuda-13 is needed by pytorch backend but will be required by vLLM and TensorRT-LLM # on platforms other than IGPU we still using CUDA 12 it means we need to add this conditionally - if target_platform() not in ["igpu", "windows", "rhel"]: + if target_platform() not in ["igpu", "windows", "rhel"] and enable_gpu: backend_dependencies += " libnvshmem3-cuda-13" # openssh-server is needed for fastertransformer @@ -1352,6 +1354,21 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach gpu_enabled=gpu_enabled ) + if enable_gpu: + df += install_dcgm_libraries(argmap["DCGM_VERSION"], target_machine) + # This segment will break the RHEL SBSA build. Need to determine whether + # this is necessary to incorporate. + if target_platform() != "rhel": + df += """ +# Extra defensive wiring for CUDA Compat lib +RUN ln -sf ${_CUDA_COMPAT_PATH}/lib.real ${_CUDA_COMPAT_PATH}/lib \\ + && echo ${_CUDA_COMPAT_PATH}/lib > /etc/ld.so.conf.d/00-cuda-compat.conf \\ + && ldconfig \\ + && rm -f ${_CUDA_COMPAT_PATH}/lib +""" + else: + df += add_cpu_libs_to_linux_dockerfile(backends, target_machine) + if target_platform() == "rhel": df += """ # Common dependencies. @@ -1416,21 +1433,6 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach exec(response.content, fastertransformer_buildscript.__dict__) df += fastertransformer_buildscript.create_postbuild(is_multistage_build=False) - if enable_gpu: - df += install_dcgm_libraries(argmap["DCGM_VERSION"], target_machine) - # This segment will break the RHEL SBSA build. Need to determine whether - # this is necessary to incorporate. - if target_platform() != "rhel": - df += """ -# Extra defensive wiring for CUDA Compat lib -RUN ln -sf ${_CUDA_COMPAT_PATH}/lib.real ${_CUDA_COMPAT_PATH}/lib \\ - && echo ${_CUDA_COMPAT_PATH}/lib > /etc/ld.so.conf.d/00-cuda-compat.conf \\ - && ldconfig \\ - && rm -f ${_CUDA_COMPAT_PATH}/lib -""" - else: - df += add_cpu_libs_to_linux_dockerfile(backends, target_machine) - # Add dependencies needed for python backend if "python" in backends: if target_platform() == "rhel": From 21922995931c7b88449dd385eb17fac0310a660e Mon Sep 17 00:00:00 2001 From: Misha Chornyi Date: Wed, 15 Oct 2025 11:47:38 -0700 Subject: [PATCH 07/17] Set proper version for vLLM --- build.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/build.py b/build.py index 618ad9a664..e8813ba079 100755 --- a/build.py +++ b/build.py @@ -78,7 +78,7 @@ "ort_openvino_version": "2025.3.0", "standalone_openvino_version": "2025.3.0", "dcgm_version": "4.4.0-1", - "vllm_version": "0.10.1.1", + "vllm_version": "0.10.2", "rhel_py_version": "3.12.3", } From cd3c948f85728e312d29614060512e52a66b237d Mon Sep 17 00:00:00 2001 From: Misha Chornyi Date: Wed, 15 Oct 2025 12:25:43 -0700 Subject: [PATCH 08/17] Adding package installation to pass CPU build --- build.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/build.py b/build.py index e8813ba079..22f7273d3b 100755 --- a/build.py +++ b/build.py @@ -1586,7 +1586,7 @@ def add_cpu_libs_to_linux_dockerfile(backends, target_machine): # patchelf is needed to add deps of libcublasLt.so.12 to libtorch_cuda.so RUN apt-get update \\ - && apt-get install -y --no-install-recommends openmpi-bin + && apt-get install -y --no-install-recommends openmpi-bin python3-pip RUN pip3 install patchelf==0.17.2 ENV LD_LIBRARY_PATH /usr/local/cuda/targets/{cuda_arch}-linux/lib:/usr/local/cuda/lib64/stubs:${{LD_LIBRARY_PATH}} From ff76327b0c99c97ae7c40d7f73d0829be201c8ba Mon Sep 17 00:00:00 2001 From: Misha Chornyi Date: Wed, 15 Oct 2025 14:17:16 -0700 Subject: [PATCH 09/17] Adding CMake argument to iGPU. --- build.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/build.py b/build.py index 22f7273d3b..2d34d7801a 100755 --- a/build.py +++ b/build.py @@ -660,6 +660,8 @@ def pytorch_cmake_args(images): cargs.append( cmake_backend_enable("pytorch", "TRITON_ENABLE_NVTX", FLAGS.enable_nvtx) ) + if target_platform() == "igpu": + cargs.append(cmake_backend_enable("pytorch", "TRITON_PYTORCH_NVSHMEM", False)) return cargs From 5b87832a8454db31b62597becf48c135349fd41e Mon Sep 17 00:00:00 2001 From: Misha Chornyi Date: Wed, 15 Oct 2025 17:32:08 -0700 Subject: [PATCH 10/17] Fix the base image container to unblock the build --- build.py | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/build.py b/build.py index 2d34d7801a..db1f47a466 100755 --- a/build.py +++ b/build.py @@ -1312,8 +1312,13 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach # Necessary for libtorch.so to find correct HPCX libraries if "pytorch" in backends: - df += """ + if not enable_gpu: + df += """ ENV LD_LIBRARY_PATH /opt/hpcx/ucc/lib/:/opt/hpcx/ucx/lib/:${LD_LIBRARY_PATH} +""" + else: + df += f""" +ENV LD_LIBRARY_PATH /usr/lib/{target_machine}-linux-gnu/nvshmem/13/:/opt/hpcx/ucc/lib/:/opt/hpcx/ucx/lib/:${{LD_LIBRARY_PATH}} """ backend_dependencies = "" From 95241f4ecb6f55a659033a8b015faf9cd055bb40 Mon Sep 17 00:00:00 2001 From: Misha Chornyi Date: Wed, 15 Oct 2025 19:57:17 -0700 Subject: [PATCH 11/17] Revert changes --- build.py | 51 ++++++++++++++++++--------------------------------- 1 file changed, 18 insertions(+), 33 deletions(-) diff --git a/build.py b/build.py index db1f47a466..d02cf14200 100755 --- a/build.py +++ b/build.py @@ -78,7 +78,7 @@ "ort_openvino_version": "2025.3.0", "standalone_openvino_version": "2025.3.0", "dcgm_version": "4.4.0-1", - "vllm_version": "0.10.2", + "vllm_version": "0.10.1.1", "rhel_py_version": "3.12.3", } @@ -660,8 +660,6 @@ def pytorch_cmake_args(images): cargs.append( cmake_backend_enable("pytorch", "TRITON_ENABLE_NVTX", FLAGS.enable_nvtx) ) - if target_platform() == "igpu": - cargs.append(cmake_backend_enable("pytorch", "TRITON_PYTORCH_NVSHMEM", False)) return cargs @@ -881,7 +879,6 @@ def install_dcgm_libraries(dcgm_version, target_machine): else: if target_machine == "aarch64": return """ -ENV DEBIAN_FRONTEND=noninteractive ENV DCGM_VERSION {} # Install DCGM. Steps from https://developer.nvidia.com/dcgm#Downloads RUN curl -o /tmp/cuda-keyring.deb \\ @@ -897,7 +894,6 @@ def install_dcgm_libraries(dcgm_version, target_machine): ) else: return """ -ENV DEBIAN_FRONTEND=noninteractive ENV DCGM_VERSION {} # Install DCGM. Steps from https://developer.nvidia.com/dcgm#Downloads RUN curl -o /tmp/cuda-keyring.deb \\ @@ -1312,13 +1308,8 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach # Necessary for libtorch.so to find correct HPCX libraries if "pytorch" in backends: - if not enable_gpu: - df += """ + df += """ ENV LD_LIBRARY_PATH /opt/hpcx/ucc/lib/:/opt/hpcx/ucx/lib/:${LD_LIBRARY_PATH} -""" - else: - df += f""" -ENV LD_LIBRARY_PATH /usr/lib/{target_machine}-linux-gnu/nvshmem/13/:/opt/hpcx/ucc/lib/:/opt/hpcx/ucx/lib/:${{LD_LIBRARY_PATH}} """ backend_dependencies = "" @@ -1329,12 +1320,6 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach # libgfortran5 is needed by pytorch backend on ARM if ("pytorch" in backends) and (target_machine == "aarch64"): backend_dependencies += " libgfortran5" - - # libnvshmem3-cuda-13 is needed by pytorch backend but will be required by vLLM and TensorRT-LLM - # on platforms other than IGPU we still using CUDA 12 it means we need to add this conditionally - if target_platform() not in ["igpu", "windows", "rhel"] and enable_gpu: - backend_dependencies += " libnvshmem3-cuda-13" - # openssh-server is needed for fastertransformer if "fastertransformer" in backends: backend_dependencies += " openssh-server" @@ -1361,21 +1346,6 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach gpu_enabled=gpu_enabled ) - if enable_gpu: - df += install_dcgm_libraries(argmap["DCGM_VERSION"], target_machine) - # This segment will break the RHEL SBSA build. Need to determine whether - # this is necessary to incorporate. - if target_platform() != "rhel": - df += """ -# Extra defensive wiring for CUDA Compat lib -RUN ln -sf ${_CUDA_COMPAT_PATH}/lib.real ${_CUDA_COMPAT_PATH}/lib \\ - && echo ${_CUDA_COMPAT_PATH}/lib > /etc/ld.so.conf.d/00-cuda-compat.conf \\ - && ldconfig \\ - && rm -f ${_CUDA_COMPAT_PATH}/lib -""" - else: - df += add_cpu_libs_to_linux_dockerfile(backends, target_machine) - if target_platform() == "rhel": df += """ # Common dependencies. @@ -1440,6 +1410,21 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach exec(response.content, fastertransformer_buildscript.__dict__) df += fastertransformer_buildscript.create_postbuild(is_multistage_build=False) + if enable_gpu: + df += install_dcgm_libraries(argmap["DCGM_VERSION"], target_machine) + # This segment will break the RHEL SBSA build. Need to determine whether + # this is necessary to incorporate. + if target_platform() != "rhel": + df += """ +# Extra defensive wiring for CUDA Compat lib +RUN ln -sf ${_CUDA_COMPAT_PATH}/lib.real ${_CUDA_COMPAT_PATH}/lib \\ + && echo ${_CUDA_COMPAT_PATH}/lib > /etc/ld.so.conf.d/00-cuda-compat.conf \\ + && ldconfig \\ + && rm -f ${_CUDA_COMPAT_PATH}/lib +""" + else: + df += add_cpu_libs_to_linux_dockerfile(backends, target_machine) + # Add dependencies needed for python backend if "python" in backends: if target_platform() == "rhel": @@ -1593,7 +1578,7 @@ def add_cpu_libs_to_linux_dockerfile(backends, target_machine): # patchelf is needed to add deps of libcublasLt.so.12 to libtorch_cuda.so RUN apt-get update \\ - && apt-get install -y --no-install-recommends openmpi-bin python3-pip + && apt-get install -y --no-install-recommends openmpi-bin RUN pip3 install patchelf==0.17.2 ENV LD_LIBRARY_PATH /usr/local/cuda/targets/{cuda_arch}-linux/lib:/usr/local/cuda/lib64/stubs:${{LD_LIBRARY_PATH}} From 7a62b03d117f0e954b224ae898f3b1b52d106db4 Mon Sep 17 00:00:00 2001 From: Misha Chornyi Date: Wed, 15 Oct 2025 19:57:44 -0700 Subject: [PATCH 12/17] Update vLLM version --- build.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/build.py b/build.py index d02cf14200..7873f49d26 100755 --- a/build.py +++ b/build.py @@ -78,7 +78,7 @@ "ort_openvino_version": "2025.3.0", "standalone_openvino_version": "2025.3.0", "dcgm_version": "4.4.0-1", - "vllm_version": "0.10.1.1", + "vllm_version": "0.10.2", "rhel_py_version": "3.12.3", } From d9fbf21c0fd5777a9464a55df21b3992a5d58889 Mon Sep 17 00:00:00 2001 From: Misha Chornyi Date: Wed, 15 Oct 2025 19:58:51 -0700 Subject: [PATCH 13/17] Straight forward approach to solve missed dependency for Pytorch --- build.py | 16 ++++++++++++++++ 1 file changed, 16 insertions(+) diff --git a/build.py b/build.py index 7873f49d26..7bc53a9613 100755 --- a/build.py +++ b/build.py @@ -1514,6 +1514,22 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach ENV PYTHONPATH=/opt/tritonserver/backends/dali/wheel/dali:$PYTHONPATH """ + if target_platform() not in ["igpu", "windows", "rhel"]: + libs_arch = "aarch64" if target_machine == "aarch64" else "x86_64" + df += f""" +RUN curl -o /tmp/cuda-keyring.deb \\ + https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/{libs_arch}/cuda-keyring_1.1-1_all.deb \\ + && apt install /tmp/cuda-keyring.deb \\ + && rm /tmp/cuda-keyring.deb + && apt update -qq \\ + && apt install --yes --no-install-recommends libnvshmem3-cuda-13 \\ + && rm -rf /var/lib/apt/lists/* \\ + && dpkg -L libnvshmem3-cuda-13 | grep libnvshmem_host.so | sed -e 's/libnvshmem_host.*//g' | sort -u > /etc/ld.so.conf.d/libnvshmem3-cuda-13.conf \\ + && ldconfig +""".format( + libs_arch=libs_arch + ) + df += """ WORKDIR /opt/tritonserver RUN rm -fr /opt/tritonserver/* From d9dd6f7ed8c9912d257262eafe32a48f5aab6176 Mon Sep 17 00:00:00 2001 From: Misha Chornyi Date: Wed, 15 Oct 2025 20:02:20 -0700 Subject: [PATCH 14/17] Add flag for iGPU build --- build.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/build.py b/build.py index 7bc53a9613..e0ec56becd 100755 --- a/build.py +++ b/build.py @@ -660,6 +660,10 @@ def pytorch_cmake_args(images): cargs.append( cmake_backend_enable("pytorch", "TRITON_ENABLE_NVTX", FLAGS.enable_nvtx) ) + if target_platform() == "igpu": + cargs.append( + cmake_backend_enable("pytorch", "TRITON_PYTORCH_NVSHMEM", True) + ) return cargs From bc3264aa25bd43fc44250fa928cf14a64aaadb5b Mon Sep 17 00:00:00 2001 From: Misha Chornyi Date: Wed, 15 Oct 2025 20:17:12 -0700 Subject: [PATCH 15/17] Add missed symbol --- build.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/build.py b/build.py index e0ec56becd..30749d5aa8 100755 --- a/build.py +++ b/build.py @@ -1524,7 +1524,7 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach RUN curl -o /tmp/cuda-keyring.deb \\ https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/{libs_arch}/cuda-keyring_1.1-1_all.deb \\ && apt install /tmp/cuda-keyring.deb \\ - && rm /tmp/cuda-keyring.deb + && rm /tmp/cuda-keyring.deb \\ && apt update -qq \\ && apt install --yes --no-install-recommends libnvshmem3-cuda-13 \\ && rm -rf /var/lib/apt/lists/* \\ From 7d24a8a3a1b1563e55886f94c20d39deef27695c Mon Sep 17 00:00:00 2001 From: Misha Chornyi Date: Wed, 15 Oct 2025 20:38:29 -0700 Subject: [PATCH 16/17] Update source url --- build.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/build.py b/build.py index 30749d5aa8..ac44a23879 100755 --- a/build.py +++ b/build.py @@ -1519,10 +1519,10 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach """ if target_platform() not in ["igpu", "windows", "rhel"]: - libs_arch = "aarch64" if target_machine == "aarch64" else "x86_64" + repo_arch = "sbsa" if target_machine == "aarch64" else "x86_64" df += f""" RUN curl -o /tmp/cuda-keyring.deb \\ - https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/{libs_arch}/cuda-keyring_1.1-1_all.deb \\ + https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/{repo_arch}/cuda-keyring_1.1-1_all.deb \\ && apt install /tmp/cuda-keyring.deb \\ && rm /tmp/cuda-keyring.deb \\ && apt update -qq \\ @@ -1531,7 +1531,7 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach && dpkg -L libnvshmem3-cuda-13 | grep libnvshmem_host.so | sed -e 's/libnvshmem_host.*//g' | sort -u > /etc/ld.so.conf.d/libnvshmem3-cuda-13.conf \\ && ldconfig """.format( - libs_arch=libs_arch + repo_arch=repo_arch ) df += """ From ab55273e75b69b780b96cbb0fd7d8fe63afc70f2 Mon Sep 17 00:00:00 2001 From: Misha Chornyi Date: Wed, 15 Oct 2025 20:42:29 -0700 Subject: [PATCH 17/17] Update property for iGPU --- build.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/build.py b/build.py index ac44a23879..c5dc3b3920 100755 --- a/build.py +++ b/build.py @@ -662,7 +662,7 @@ def pytorch_cmake_args(images): ) if target_platform() == "igpu": cargs.append( - cmake_backend_enable("pytorch", "TRITON_PYTORCH_NVSHMEM", True) + cmake_backend_enable("pytorch", "TRITON_PYTORCH_NVSHMEM", False) ) return cargs