diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p1d-dep2-tep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p1d-dep2-tep8-1k1k.yaml
index e1c908228..e76827af3 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p1d-dep2-tep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p1d-dep2-tep8-1k1k.yaml
@@ -28,13 +28,11 @@ backend:
   connector: null
 
   prefill_environment:
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
     VLLM_FLOAT32_MATMUL_PRECISION: "high"
 
   decode_environment:
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
     VLLM_FLOAT32_MATMUL_PRECISION: "high"
 
   vllm_config:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p1d-dep2-tp8-marlin-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p1d-dep2-tp8-marlin-1k1k.yaml
index 452d748b4..c8362cb32 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p1d-dep2-tp8-marlin-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p1d-dep2-tp8-marlin-1k1k.yaml
@@ -28,13 +28,11 @@ backend:
   connector: null
 
   prefill_environment:
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
     VLLM_FLOAT32_MATMUL_PRECISION: "high"
 
   decode_environment:
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
     VLLM_FLOAT32_MATMUL_PRECISION: "high"
 
   vllm_config:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p2d-dep2-dep4-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p2d-dep2-dep4-1k1k.yaml
index 625927baf..349a125bb 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p2d-dep2-dep4-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p2d-dep2-dep4-1k1k.yaml
@@ -28,13 +28,11 @@ backend:
   connector: null
 
   prefill_environment:
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
     VLLM_FLOAT32_MATMUL_PRECISION: "high"
 
   decode_environment:
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
     VLLM_FLOAT32_MATMUL_PRECISION: "high"
 
   vllm_config:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p1d-dep2-dep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p1d-dep2-dep8-1k1k.yaml
index 951cdfc94..0f790c79b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p1d-dep2-dep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p1d-dep2-dep8-1k1k.yaml
@@ -28,13 +28,11 @@ backend:
   connector: null
 
   prefill_environment:
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
     VLLM_FLOAT32_MATMUL_PRECISION: "high"
 
   decode_environment:
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
     VLLM_FLOAT32_MATMUL_PRECISION: "high"
 
   vllm_config:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p1d-dep2-tep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p1d-dep2-tep8-1k1k.yaml
index 495fc9b78..1372ff29a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p1d-dep2-tep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p1d-dep2-tep8-1k1k.yaml
@@ -28,13 +28,11 @@ backend:
   connector: null
 
   prefill_environment:
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
     VLLM_FLOAT32_MATMUL_PRECISION: "high"
 
   decode_environment:
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
     VLLM_FLOAT32_MATMUL_PRECISION: "high"
 
   vllm_config:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p2d-dep2-tep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p2d-dep2-tep8-1k1k.yaml
index 58c533504..4447d971b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p2d-dep2-tep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p2d-dep2-tep8-1k1k.yaml
@@ -28,13 +28,11 @@ backend:
   connector: null
 
   prefill_environment:
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
     VLLM_FLOAT32_MATMUL_PRECISION: "high"
 
   decode_environment:
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
     VLLM_FLOAT32_MATMUL_PRECISION: "high"
 
   vllm_config:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/3p2d-dep2-tep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/3p2d-dep2-tep8-1k1k.yaml
index 165072cc5..b03d644d2 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/3p2d-dep2-tep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/3p2d-dep2-tep8-1k1k.yaml
@@ -28,13 +28,11 @@ backend:
   connector: null
 
   prefill_environment:
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
     VLLM_FLOAT32_MATMUL_PRECISION: "high"
 
   decode_environment:
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
     VLLM_FLOAT32_MATMUL_PRECISION: "high"
 
   vllm_config:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p2d-dep2-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p2d-dep2-dep8-8k1k.yaml
index cb9256e25..890014563 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p2d-dep2-dep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p2d-dep2-dep8-8k1k.yaml
@@ -29,13 +29,11 @@ backend:
 
   prefill_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
 
   decode_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
 
   vllm_config:
     prefill:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/2p2d-dep2-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/2p2d-dep2-dep8-8k1k.yaml
index 7a805df7f..6d9ecc425 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/2p2d-dep2-dep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/2p2d-dep2-dep8-8k1k.yaml
@@ -29,13 +29,11 @@ backend:
 
   prefill_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
 
   decode_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
 
   vllm_config:
     prefill:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/2p2d-dep2-tep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/2p2d-dep2-tep8-8k1k.yaml
index 8dc4c5c06..90d816592 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/2p2d-dep2-tep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/2p2d-dep2-tep8-8k1k.yaml
@@ -29,13 +29,11 @@ backend:
 
   prefill_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
 
   decode_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
 
   vllm_config:
     prefill:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/3p2d-dep2-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/3p2d-dep2-dep8-8k1k.yaml
index 8ff86e9d9..84d580452 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/3p2d-dep2-dep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/3p2d-dep2-dep8-8k1k.yaml
@@ -29,13 +29,11 @@ backend:
 
   prefill_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
 
   decode_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
 
   vllm_config:
     prefill:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/3p2d-dep2-tep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/3p2d-dep2-tep8-8k1k.yaml
index a90bfcf32..f272b21bc 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/3p2d-dep2-tep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/3p2d-dep2-tep8-8k1k.yaml
@@ -29,13 +29,11 @@ backend:
 
   prefill_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
 
   decode_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
 
   vllm_config:
     prefill:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/4p2d-dep2-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/4p2d-dep2-dep8-8k1k.yaml
index 8feb79b71..b087b0926 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/4p2d-dep2-dep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/4p2d-dep2-dep8-8k1k.yaml
@@ -29,13 +29,11 @@ backend:
 
   prefill_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
 
   decode_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
 
   vllm_config:
     prefill:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/4p3d-dep2-dep4-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/4p3d-dep2-dep4-8k1k.yaml
index 5639c2eda..94c36243e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/4p3d-dep2-dep4-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/4p3d-dep2-dep4-8k1k.yaml
@@ -29,13 +29,11 @@ backend:
 
   prefill_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
 
   decode_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
 
   vllm_config:
     prefill:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/5p2d-dep2-tep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/5p2d-dep2-tep8-8k1k.yaml
index 1f517d815..94f546ec2 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/5p2d-dep2-tep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/5p2d-dep2-tep8-8k1k.yaml
@@ -29,13 +29,11 @@ backend:
 
   prefill_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
 
   decode_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
 
   vllm_config:
     prefill:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/5p2d-dep2-tp8-marlin-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/5p2d-dep2-tp8-marlin-8k1k.yaml
index cf1840dfe..e77e77600 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/5p2d-dep2-tp8-marlin-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/5p2d-dep2-tp8-marlin-8k1k.yaml
@@ -29,13 +29,11 @@ backend:
 
   prefill_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
 
   decode_environment:
     VLLM_FLOAT32_MATMUL_PRECISION: high
-    UCX_NET_DEVICES: "all"
-    UCX_TLS: "rc,cuda_ipc,cuda_copy,sm,self,tcp"
+    UCX_TLS: "cuda_copy,rc"
 
   vllm_config:
     prefill:
diff --git a/runners/launch_b300-nv.sh b/runners/launch_b300-nv.sh
index c32516c9e..f2a83e4b3 100644
--- a/runners/launch_b300-nv.sh
+++ b/runners/launch_b300-nv.sh
@@ -85,7 +85,7 @@ elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm2.5" && $PRECIS
 elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" ]]; then
     git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR" || exit 1
-    git checkout main
+    git checkout sa-submission-q2-2026
     mkdir -p recipes/vllm/minimax-m3
     cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3" recipes/vllm/minimax-m3
 else