Add CUDA 13 Support (#20)

Author: AryanAhadinia

src/swiss_ai_model_launch/assets/envs/sglang.toml

@@ -1,4 +1,4 @@
-image = "/iopsstor/scratch/cscs/edebened/ce-images/large_model_serving.sqsh"
+image = "/capstor/store/cscs/swissai/infra01/container-images/sglang_cuda13.sqsh"
 
 mounts = [
   "/iopsstor/store/cscs/swissai/a09/xyao/bin:/ocfbin",
@@ -7,16 +7,14 @@ mounts = [
   "/usr/lib64/libhwloc.so.15:/usr/lib/libhwloc.so.15",
   "/usr/lib64/libpciaccess.so.0:/usr/lib/libpciaccess.so.0",
   "/usr/lib64/libxml2.so.2:/usr/lib/libxml2.so.2",
-  "/opt/cray/libfabric/1.22.0/lib64:/opt/cray/libfabric/lib64",
-  "/opt/cscs/aws-ofi-ccl-plugin/cuda12:/opt/cscs/aws-ofi-ccl-plugin/cuda12",
+  "/usr/lib64/libnuma.so.1:/usr/lib/libnuma.so.1",
 ]
 
-workdir = "/opt" 
+workdir = "/opt"
 
 [env]
 # NCCL_DEBUG = "INFO"  # uncomment for debugging
 # NCCL_DEBUG_SUBSYS = "INIT,NET"  # uncomment for debugging
-NCCL_NET_PLUGIN = "/opt/cscs/aws-ofi-ccl-plugin/cuda12/libnccl-net.so"
 NCCL_NET = "AWS Libfabric"
 NCCL_CROSS_NIC = "1"
 NCCL_NET_GDR_LEVEL = "PHB"
@@ -34,4 +32,5 @@ VLLM_ALLREDUCE_USE_SYMM_MEM = "0"
 
 [annotations]
 com.hooks.aws_ofi_nccl.enabled = "true"
-com.hooks.aws_ofi_nccl.variant = "cuda12"
+com.hooks.aws_ofi_nccl.variant = "cuda13"
+com.hooks.cxi.enabled = "true"

src/swiss_ai_model_launch/assets/envs/sglang_glm.toml

@@ -1,4 +1,4 @@
-image = "vllm/vllm-openai:latest"
+image = "/capstor/store/cscs/swissai/infra01/container-images/vllm_cuda13.sqsh"
 
 mounts = [
   "/iopsstor/store/cscs/swissai/a09/xyao/bin:/ocfbin",
@@ -7,22 +7,19 @@ mounts = [
   "/usr/lib64/libhwloc.so.15:/usr/lib/libhwloc.so.15",
   "/usr/lib64/libpciaccess.so.0:/usr/lib/libpciaccess.so.0",
   "/usr/lib64/libxml2.so.2:/usr/lib/libxml2.so.2",
-  "/opt/cray/libfabric/1.22.0/lib64:/opt/cray/libfabric/lib64",
-  "/opt/cscs/aws-ofi-ccl-plugin/cuda12:/opt/cscs/aws-ofi-ccl-plugin/cuda12",
+  "/usr/lib64/libnuma.so.1:/usr/lib/libnuma.so.1",
 ]
 
-workdir = "/opt" 
+workdir = "/opt"
 
 [env]
-LD_LIBRARY_PATH = "/opt/cscs/aws-ofi-ccl-plugin/cuda12:/opt/cray/libfabric/lib64:/usr/lib:${LD_LIBRARY_PATH:-}"
-NCCL_NET_PLUGIN = "/opt/cscs/aws-ofi-ccl-plugin/cuda12/libnccl-net.so"
+# NCCL_DEBUG = "INFO"  # uncomment for debugging
+# NCCL_DEBUG_SUBSYS = "INIT,NET"  # uncomment for debugging
 NCCL_NET = "AWS Libfabric"
 NCCL_CROSS_NIC = "1"
 NCCL_NET_GDR_LEVEL = "PHB"
 NCCL_SOCKET_IFNAME = "hsn"
 NCCL_PROTO = "^LL128"
-NCCL_DEBUG = "INFO"
-NCCL_DEBUG_SUBSYS = "INIT,NET"
 FI_CXI_COMPAT = "0"
 FI_MR_CACHE_MONITOR = "userfaultfd"
 FI_CXI_RX_MATCH_MODE = "software"
@@ -34,4 +31,5 @@ VLLM_ALLREDUCE_USE_SYMM_MEM = "0"
 
 [annotations]
 com.hooks.aws_ofi_nccl.enabled = "true"
-com.hooks.aws_ofi_nccl.variant = "cuda12"
+com.hooks.aws_ofi_nccl.variant = "cuda13"
+com.hooks.cxi.enabled = "true"

src/swiss_ai_model_launch/assets/envs/sglang_kimi.toml

@@ -35,6 +35,8 @@ OCF_BOOTSTRAP_ADDR=$(curl -sf http://148.187.108.172:8092/v1/dnt/bootstraps \
     || echo "/ip4/148.187.108.172/tcp/43905/p2p/QmbUKJkCfotDzbFE5uoTsXD4GRyPHjzZC1f2yAGLoeBMn9")
 
 
+unset SLURM_CPU_BIND SLURM_CPU_BIND_TYPE SLURM_CPU_BIND_LIST SLURM_CPU_BIND_VERBOSE
+
 # Architecture detection
 ARCH=$(uname -m)
 if [[ "$ARCH" == "aarch64" ]]; then