ai-dynamo · ovidiusm · Mar 1, 2026 · Jan 21, 2026 · Jan 21, 2026 · Jan 22, 2026
@@ -43,7 +43,7 @@ env:
   TEST_TIMEOUT: 30
   UCX_TLS: "^shm"
   STORAGE_DRIVER: 'overlay'
-  CI_IMAGE_TAG: "20260219-1"
+  CI_IMAGE_TAG: "20260226-1"
 
 
 runs_on_dockers:

@@ -49,7 +49,7 @@ env:
   SLURM_JOB_TIMEOUT: '02:20:00'
   SLURM_IMMEDIATE_TIMEOUT: "3600"
   STORAGE_DRIVER: overlay
-  CI_IMAGE_TAG: "20260219-1"
+  CI_IMAGE_TAG: "20260226-1"
 
 empty_volumes:
   - {mountPath: /var/lib/containers/storage, memory: false}

@@ -35,7 +35,7 @@ LIBFABRIC_VERSION=${LIBFABRIC_VERSION:-v1.21.0}
 # LIBFABRIC_INSTALL_DIR can be set via environment variable, defaults to INSTALL_DIR
 LIBFABRIC_INSTALL_DIR=${LIBFABRIC_INSTALL_DIR:-$INSTALL_DIR}
 # UCCL_COMMIT_SHA is the commit SHA of UCCL.
-UCCL_COMMIT_SHA="a962f611021afc2e3c9358f6da4ae96539cbca0f"
+UCCL_COMMIT_SHA="2de728f1a27ea3f3b66059baf838f940e243ebc6"
 AZURITE_VER="3.35.0"
 TMPDIR=$(mktemp -d)
 

@@ -105,7 +105,7 @@ if $HAS_GPU ; then
     for op_type in READ WRITE; do
         for initiator in $seg_types; do
             for target in $seg_types; do
-                UCCL_RCMODE=1 run_nixlbench_two_workers --backend UCCL --op_type $op_type --initiator_seg_type $initiator --target_seg_type $target --check_consistency
+                run_nixlbench_two_workers --backend UCCL --op_type $op_type --initiator_seg_type $initiator --target_seg_type $target --check_consistency
             done
         done
     done

@@ -311,6 +311,9 @@ xferBenchNixlWorker::xferBenchNixlWorker(int *argc, char ***argv, std::vector<st
 }
 
 xferBenchNixlWorker::~xferBenchNixlWorker() {
+    delete rt;
+    rt = nullptr;
+
     if (agent) {
         delete agent;
         agent = nullptr;

@@ -34,20 +34,18 @@ UCCL engine would auto discover the right NIC to be used for the GPU based on th
 
 Refer to [README](https://github.com/uccl-project/uccl/tree/main/collective/rdma#environment-variables-in-uccl) for the complete list of environment variables that can be set to customize UCCL.
 
-**Important**: For `NIXL_READ` operations set `UCCL_RCMODE=1`. By default, UCCL uses RDMA UC (Unreliable Connection). However, `READ` operations need to operate on RDMA RC (Reliable Connection). `WRITE` operations can work on both RDMA UC and RC.
-
 ### Usage References
 
 1) [NIXL Benchmark](https://github.com/uccl-project/uccl/blob/main/p2p/benchmarks/benchmark_nixl.py) in UCCL P2P: Refer to  this [README](https://github.com/uccl-project/uccl/tree/main/p2p) on how to run the script.
 
-2) [NIXL connector](https://github.com/vllm-project/vllm/commit/e731733d30d0aed3252dc60427927768bfc0ca73) in vLLM. vLLM's NIXL connector uses `NIXL_READ` operations, hence set env `UCCL_RCMODE` to 1.
+2) [NIXL connector](https://github.com/vllm-project/vllm/commit/e731733d30d0aed3252dc60427927768bfc0ca73) in vLLM.
 
 ### Road Map
 
-- [ ] Add Intra-node communication support
+- ✅ Add asynchronous posting of reads over multiple workers to mitigate latency increase upon fragmentation
 
-- [ ] Add Progress Thread support
+- 🚧 Add Intra-node communication support
 
-- [ ] Add asynchronous posting of reads over multiple workers to mitigate latency increase upon fragmentation
+- 🚧 Add Progress Thread support
 
-- [ ] Add support for other transport (TCP, TCP-X, etc.)
+- 🚧 Add support for other transport (TCP, TCP-X, etc.)
-2) [NIXL connector](https://github.com/vllm-project/vllm/commit/e731733d30d0aed3252dc60427927768bfc0ca73) in vLLM.
-
-### Road Map
-
- [ ] Add Intra-node communication support
- ✅ Add asynchronous posting of reads over multiple workers to mitigate latency increase upon fragmentation
-
- [ ] Add Progress Thread support
- 🚧 Add Intra-node communication support
-
- [ ] Add asynchronous posting of reads over multiple workers to mitigate latency increase upon fragmentation
- 🚧 Add Progress Thread support
-
- [ ] Add support for other transport (TCP, TCP-X, etc.)
-\ No newline at end of file
- 🚧 Add support for other transport (TCP, TCP-X, etc.)
+2) [NIXL connector](https://github.com/vllm-project/vllm/commit/e731733d30d0aed3252dc60427927768bfc0ca73) in vLLM.
+
+### Road Map
+
+- ✅ Add asynchronous posting of reads over multiple workers to mitigate latency increase upon fragmentation
+
+- 🚧 Add Intra-node communication support
+
+- 🚧 Add Progress Thread support
+
+- 🚧 Add support for other transport (TCP, TCP-X, etc.)
+
-2) [NIXL connector](https://github.com/vllm-project/vllm/commit/e731733d30d0aed3252dc60427927768bfc0ca73) in vLLM.
-
-### Road Map
-
- [ ] Add Intra-node communication support
- ✅ Add asynchronous posting of reads over multiple workers to mitigate latency increase upon fragmentation
-
- [ ] Add Progress Thread support
- 🚧 Add Intra-node communication support
-
- [ ] Add asynchronous posting of reads over multiple workers to mitigate latency increase upon fragmentation
- 🚧 Add Progress Thread support
-
- [ ] Add support for other transport (TCP, TCP-X, etc.)
-\ No newline at end of file
- 🚧 Add support for other transport (TCP, TCP-X, etc.)
+2) [NIXL connector](https://github.com/vllm-project/vllm/commit/e731733d30d0aed3252dc60427927768bfc0ca73) in vLLM.
+
+### Road Map
+
+- ✅ Add asynchronous posting of reads over multiple workers to mitigate latency increase upon fragmentation
+
+- 🚧 Add Intra-node communication support
+
+- 🚧 Add Progress Thread support
+
+- 🚧 Add support for other transport (TCP, TCP-X, etc.)
+