From 350635a04b13be6bc4883beddd7519c129002439 Mon Sep 17 00:00:00 2001
From: Chuang Zhu <111838961+chuangz0@users.noreply.github.com>
Date: Tue, 21 Jul 2026 05:28:17 -0700
Subject: [PATCH 1/4] [https://nvbugs/6426834][fix] Deflake test_kv_transfer:
 cap NIXL progress threads, pin UCX env

test_transfer_worker_v2[tp4_pp1_to_tp2_pp2] (and other multi-worker cases)
intermittently hit the 120s per-test timeout in CI. Root cause: each NIXL
transfer agent spawns TRTLLM_NIXL_NUM_THREADS (default 8) busy-polling UCX
progress threads, and a single case builds up to 8 TransferWorkers (64
spinning threads per pytest process). On CI nodes shared by up to 8
single-GPU Jenkins agents the resulting CPU oversubscription inflates agent
construction from ~3s to ~27s each, so 8 agents alone exceed the 120s
timeout depending on neighbor load.

Measured on an H100 node with 4 concurrent pytest processes: 8 agents take
~178s with the default 8 progress threads vs ~23s with 1 thread.

Fix, applied to test_kv_transfer.py and test_kv_transfer_mp.py:
- setdefault TRTLLM_NIXL_NUM_THREADS=1: these tests verify transfer logic,
  not transfer-engine threading (sender-side KV_TRANSFER_NUM_THREADS worker
  threads are unchanged).
- Force UCX_TLS=^ib,gdr_copy and drop UCX_NET_DEVICES instead of the
  previous setdefault, which was inert in CI (the CI agent bootstrap
  exports UCX_TLS before pytest starts) and broke local runs on clusters
  whose container runtime injects UCX_TLS/UCX_NET_DEVICES defaults.

Un-waive the affected case.

Validation: 8 concurrent whole-file runs x 5 iterations (40 runs) on one
H100 node under CI-parity env all passed; before the fix the same setup
timed out in the first iteration on every GPU.

Signed-off-by: Chuang Zhu <111838961+chuangz0@users.noreply.github.com>
---
 tests/integration/test_lists/waives.txt       |  4 +---
 .../disaggregated/test_kv_transfer.py         | 20 +++++++++++++++++--
 .../disaggregated/test_kv_transfer_mp.py      |  9 ++++++++-
 3 files changed, 27 insertions(+), 6 deletions(-)

diff --git a/tests/integration/test_lists/waives.txt b/tests/integration/test_lists/waives.txt
index 80e697faf3be..066878706ad8 100644
--- a/tests/integration/test_lists/waives.txt
+++ b/tests/integration/test_lists/waives.txt
@@ -363,9 +363,7 @@ unittest/_torch/visual_gen/multi_gpu/test_ulysses_async.py::test_capture_smoke S
 unittest/_torch/visual_gen/test_attention_integration.py::test_sage_attention_self_attention[fp8-2-1560] SKIP (https://nvbugs/6198760)
 unittest/_torch/visual_gen/test_attention_integration.py::test_sage_attention_self_attention[int8-2-1560] SKIP (https://nvbugs/6198760)
 unittest/auto_deploy/multigpu/custom_ops SKIP (https://nvbugs/6403920)
-unittest/disaggregated/test_kv_transfer.py SKIP (https://nvbugs/6403793)
-unittest/disaggregated/test_kv_transfer.py::test_transfer_worker_v2[tp1_pp4_to_tp2_pp2] SKIP (https://nvbugs/6445316)
-unittest/disaggregated/test_kv_transfer.py::test_transfer_worker_v2[tp4_pp1_to_tp2_pp2] SKIP (https://nvbugs/6426834)
+unittest/executor/test_proxy_fast_death.py::test_pool_session_shutdown_never_blocks_after_release SKIP (https://nvbugs/6480574)
 unittest/executor/test_rpc.py::TestRpcCorrectness::test_incremental_task_async SKIP (https://nvbugs/5741476)
 unittest/executor/test_rpc_proxy.py SKIP (https://nvbugs/5605741)
 unittest/executor/test_rpc_worker.py SKIP (https://nvbugs/5605741)
diff --git a/tests/unittest/disaggregated/test_kv_transfer.py b/tests/unittest/disaggregated/test_kv_transfer.py
index 76a0d8db25b3..8a8c13f34f48 100644
--- a/tests/unittest/disaggregated/test_kv_transfer.py
+++ b/tests/unittest/disaggregated/test_kv_transfer.py
@@ -5,8 +5,24 @@
 import time
 import uuid
 
-# Exclude IB (no fabric) and gdr_copy (UCX rcache SIGABRT at teardown).
-os.environ.setdefault("UCX_TLS", "^ib,gdr_copy")
+# Force a deterministic UCX config regardless of what the cluster/CI injects
+# (enroot may inject UCX_TLS=rc_v,sm + UCX_NET_DEVICES=mlx5_*; the CI agent
+# bootstrap exports UCX_TLS=tcp,cuda_copy,cuda_ipc):
+# - exclude IB (no fabric assumed) and gdr_copy (UCX rcache SIGABRT at
+#   teardown);
+# - drop UCX_NET_DEVICES, otherwise with ^ib the remaining transports cannot
+#   use the injected mlx5_* device list and UCX fails with "no active
+#   messages transport".
+os.environ["UCX_TLS"] = "^ib,gdr_copy"
+os.environ.pop("UCX_NET_DEVICES", None)
+# Each NIXL agent spawns TRTLLM_NIXL_NUM_THREADS (default 8) busy-polling
+# progress threads, and a single case builds up to 8 TransferWorkers (one per
+# rank). On CI nodes shared with other single-GPU jobs the resulting CPU
+# oversubscription inflates agent construction from ~3s to ~30s each, blowing
+# the 120s per-test timeout intermittently (https://nvbugs/6426834). One
+# progress thread is enough here: these tests verify transfer logic, not
+# transfer-engine threading.
+os.environ.setdefault("TRTLLM_NIXL_NUM_THREADS", "1")
 from dataclasses import dataclass
 from typing import List, Optional
 
diff --git a/tests/unittest/disaggregated/test_kv_transfer_mp.py b/tests/unittest/disaggregated/test_kv_transfer_mp.py
index 76e5f4aa7573..797a582526bb 100644
--- a/tests/unittest/disaggregated/test_kv_transfer_mp.py
+++ b/tests/unittest/disaggregated/test_kv_transfer_mp.py
@@ -8,7 +8,14 @@
 import torch.multiprocessing as mp
 
 # Exclude IB (no fabric) and gdr_copy (UCX rcache SIGABRT at teardown).
-os.environ.setdefault("UCX_TLS", "^ib,gdr_copy")
+# Force a deterministic UCX config regardless of what the cluster/CI injects;
+# see test_kv_transfer.py for the full rationale.
+os.environ["UCX_TLS"] = "^ib,gdr_copy"
+os.environ.pop("UCX_NET_DEVICES", None)
+# Limit NIXL busy-polling progress threads; see test_kv_transfer.py for the
+# full rationale (intermittent 120s timeouts on shared CI nodes,
+# https://nvbugs/6426834).
+os.environ.setdefault("TRTLLM_NIXL_NUM_THREADS", "1")
 
 import tensorrt_llm
 import tensorrt_llm.bindings

From fe0f929e028c31c1f42dd5c8efafcd6f57aec24b Mon Sep 17 00:00:00 2001
From: Chuang Zhu <111838961+chuangz0@users.noreply.github.com>
Date: Wed, 22 Jul 2026 08:25:25 +0000
Subject: [PATCH 2/4] [https://nvbugs/6426834][fix] Pin UCX/NIXL env in all
 transceiver unit tests

Address review: assign TRTLLM_NIXL_NUM_THREADS explicitly instead of
setdefault so an externally supplied value cannot defeat the
single-progress-thread cap, and apply the same deterministic env
(UCX_TLS=^ib,gdr_copy, drop UCX_NET_DEVICES, TRTLLM_NIXL_NUM_THREADS=1)
to every unit test that creates real UCX/NIXL transfer agents, not just
test_kv_transfer*.py.

Signed-off-by: Chuang Zhu <111838961+chuangz0@users.noreply.github.com>
---
 tests/unittest/disaggregated/test_agent.py               | 7 +++++--
 .../unittest/disaggregated/test_agent_multi_backends.py  | 8 ++++++--
 .../test_cache_transceiver_single_process.py             | 9 +++++----
 tests/unittest/disaggregated/test_kv_transfer.py         | 2 +-
 tests/unittest/disaggregated/test_kv_transfer_mp.py      | 2 +-
 tests/unittest/disaggregated/test_mamba_transfer.py      | 7 +++++++
 .../disaggregated/test_py_cache_transceiver_mp.py        | 7 +++++--
 tests/unittest/others/test_kv_cache_transceiver.py       | 7 +++++++
 8 files changed, 37 insertions(+), 12 deletions(-)

diff --git a/tests/unittest/disaggregated/test_agent.py b/tests/unittest/disaggregated/test_agent.py
index f2212a6c5f82..ad4688ef2f0d 100644
--- a/tests/unittest/disaggregated/test_agent.py
+++ b/tests/unittest/disaggregated/test_agent.py
@@ -6,8 +6,11 @@
 import pytest
 import torch
 
-# Exclude IB (no fabric) and gdr_copy (UCX rcache SIGABRT at teardown).
-os.environ.setdefault("UCX_TLS", "^ib,gdr_copy")
+# Force a deterministic UCX/NIXL config regardless of what the cluster/CI
+# injects; see test_kv_transfer.py for the full rationale.
+os.environ["UCX_TLS"] = "^ib,gdr_copy"
+os.environ.pop("UCX_NET_DEVICES", None)
+os.environ["TRTLLM_NIXL_NUM_THREADS"] = "1"
 
 from tensorrt_llm import logger
 from tensorrt_llm._torch.disaggregation.base.agent import (
diff --git a/tests/unittest/disaggregated/test_agent_multi_backends.py b/tests/unittest/disaggregated/test_agent_multi_backends.py
index d983af152542..c8fa0d0b85d6 100644
--- a/tests/unittest/disaggregated/test_agent_multi_backends.py
+++ b/tests/unittest/disaggregated/test_agent_multi_backends.py
@@ -3,8 +3,12 @@
 
 import pytest
 
-# Exclude IB (no fabric) and gdr_copy (UCX rcache SIGABRT at teardown).
-os.environ.setdefault("UCX_TLS", "^ib,gdr_copy")
+# Force a deterministic UCX/NIXL config regardless of what the cluster/CI
+# injects; see test_kv_transfer.py for the full rationale. The subprocesses
+# spawned below inherit these via os.environ.copy().
+os.environ["UCX_TLS"] = "^ib,gdr_copy"
+os.environ.pop("UCX_NET_DEVICES", None)
+os.environ["TRTLLM_NIXL_NUM_THREADS"] = "1"
 
 
 def test_load_agent_missing_module():
diff --git a/tests/unittest/disaggregated/test_cache_transceiver_single_process.py b/tests/unittest/disaggregated/test_cache_transceiver_single_process.py
index 651a347b2f08..4d3ba3779c0d 100644
--- a/tests/unittest/disaggregated/test_cache_transceiver_single_process.py
+++ b/tests/unittest/disaggregated/test_cache_transceiver_single_process.py
@@ -28,7 +28,11 @@
 # Exclude UCX IB transport (avoid NIXL setup hangs without IB) and gdr_copy
 # (avoid SIGSEGV at process exit from UCX rcache cleanup; gdr_copy disabled
 # falls back to cuda_ipc / cuda_copy without affecting correctness).
-os.environ.setdefault("UCX_TLS", "^ib,gdr_copy")
+# Force a deterministic UCX/NIXL config regardless of what the cluster/CI
+# injects; see test_kv_transfer.py for the full rationale.
+os.environ["UCX_TLS"] = "^ib,gdr_copy"
+os.environ.pop("UCX_NET_DEVICES", None)
+os.environ["TRTLLM_NIXL_NUM_THREADS"] = "1"
 from dataclasses import dataclass
 from typing import Dict, List, Optional
 
@@ -55,9 +59,6 @@
 
 AttentionTypeCpp = tensorrt_llm.bindings.internal.batch_manager.AttentionType
 
-# Reduce NIXL threads for unit test: default 8 threads per agent causes heavy
-# contention when creating multiple agents on a single GPU in the same process.
-os.environ.setdefault("TRTLLM_NIXL_NUM_THREADS", "0")
 # ---------------------------------------------------------------------------
 # Constants
 # ---------------------------------------------------------------------------
diff --git a/tests/unittest/disaggregated/test_kv_transfer.py b/tests/unittest/disaggregated/test_kv_transfer.py
index 8a8c13f34f48..821fcc431c03 100644
--- a/tests/unittest/disaggregated/test_kv_transfer.py
+++ b/tests/unittest/disaggregated/test_kv_transfer.py
@@ -22,7 +22,7 @@
 # the 120s per-test timeout intermittently (https://nvbugs/6426834). One
 # progress thread is enough here: these tests verify transfer logic, not
 # transfer-engine threading.
-os.environ.setdefault("TRTLLM_NIXL_NUM_THREADS", "1")
+os.environ["TRTLLM_NIXL_NUM_THREADS"] = "1"
 from dataclasses import dataclass
 from typing import List, Optional
 
diff --git a/tests/unittest/disaggregated/test_kv_transfer_mp.py b/tests/unittest/disaggregated/test_kv_transfer_mp.py
index 797a582526bb..d58315e1a93a 100644
--- a/tests/unittest/disaggregated/test_kv_transfer_mp.py
+++ b/tests/unittest/disaggregated/test_kv_transfer_mp.py
@@ -15,7 +15,7 @@
 # Limit NIXL busy-polling progress threads; see test_kv_transfer.py for the
 # full rationale (intermittent 120s timeouts on shared CI nodes,
 # https://nvbugs/6426834).
-os.environ.setdefault("TRTLLM_NIXL_NUM_THREADS", "1")
+os.environ["TRTLLM_NIXL_NUM_THREADS"] = "1"
 
 import tensorrt_llm
 import tensorrt_llm.bindings
diff --git a/tests/unittest/disaggregated/test_mamba_transfer.py b/tests/unittest/disaggregated/test_mamba_transfer.py
index 0cf0b3209899..f844a31597b9 100644
--- a/tests/unittest/disaggregated/test_mamba_transfer.py
+++ b/tests/unittest/disaggregated/test_mamba_transfer.py
@@ -12,6 +12,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import os
 import threading
 import uuid
 from typing import Dict, List
@@ -19,6 +20,12 @@
 import pytest
 import torch
 
+# Force a deterministic UCX/NIXL config regardless of what the cluster/CI
+# injects; see test_kv_transfer.py for the full rationale.
+os.environ["UCX_TLS"] = "^ib,gdr_copy"
+os.environ.pop("UCX_NET_DEVICES", None)
+os.environ["TRTLLM_NIXL_NUM_THREADS"] = "1"
+
 import tensorrt_llm
 import tensorrt_llm.bindings
 import tensorrt_llm.tensorrt_llm_transfer_agent_binding  # noqa: F401
diff --git a/tests/unittest/disaggregated/test_py_cache_transceiver_mp.py b/tests/unittest/disaggregated/test_py_cache_transceiver_mp.py
index a81e05026cbb..50d7a06d0012 100644
--- a/tests/unittest/disaggregated/test_py_cache_transceiver_mp.py
+++ b/tests/unittest/disaggregated/test_py_cache_transceiver_mp.py
@@ -15,8 +15,11 @@
 import torch.distributed as dist
 import torch.multiprocessing as mp
 
-# Exclude IB (no fabric) and gdr_copy (UCX rcache SIGABRT at teardown).
-os.environ.setdefault("UCX_TLS", "^ib,gdr_copy")
+# Force a deterministic UCX/NIXL config regardless of what the cluster/CI
+# injects; see test_kv_transfer.py for the full rationale.
+os.environ["UCX_TLS"] = "^ib,gdr_copy"
+os.environ.pop("UCX_NET_DEVICES", None)
+os.environ["TRTLLM_NIXL_NUM_THREADS"] = "1"
 
 import tensorrt_llm
 import tensorrt_llm.bindings
diff --git a/tests/unittest/others/test_kv_cache_transceiver.py b/tests/unittest/others/test_kv_cache_transceiver.py
index 0d6b31e367d1..86dc2ad6dd7c 100644
--- a/tests/unittest/others/test_kv_cache_transceiver.py
+++ b/tests/unittest/others/test_kv_cache_transceiver.py
@@ -3,6 +3,7 @@
 
 import gc
 import multiprocessing
+import os
 import sys
 import time
 import uuid
@@ -11,6 +12,12 @@
 import pytest
 import torch
 
+# Force a deterministic UCX/NIXL config regardless of what the cluster/CI
+# injects; see disaggregated/test_kv_transfer.py for the full rationale.
+os.environ["UCX_TLS"] = "^ib,gdr_copy"
+os.environ.pop("UCX_NET_DEVICES", None)
+os.environ["TRTLLM_NIXL_NUM_THREADS"] = "1"
+
 import tensorrt_llm
 import tensorrt_llm.bindings
 import tensorrt_llm.bindings.executor as trtllm

From fd6764de7292d1c5e3e513a5f6d4a2ad3f9c7d7f Mon Sep 17 00:00:00 2001
From: Chuang Zhu <111838961+chuangz0@users.noreply.github.com>
Date: Wed, 22 Jul 2026 08:28:02 +0000
Subject: [PATCH 3/4] [https://nvbugs/6426834][fix] Drop UCX_NET_DEVICES
 unsetting from transceiver unit tests

Keep only UCX_TLS pinning and the NIXL progress-thread cap; removing
UCX_NET_DEVICES from the environment is not needed.

Signed-off-by: Chuang Zhu <111838961+chuangz0@users.noreply.github.com>
---
 tests/unittest/disaggregated/test_agent.py            |  1 -
 .../disaggregated/test_agent_multi_backends.py        |  1 -
 .../test_cache_transceiver_single_process.py          |  1 -
 tests/unittest/disaggregated/test_kv_transfer.py      | 11 +++--------
 tests/unittest/disaggregated/test_kv_transfer_mp.py   |  1 -
 tests/unittest/disaggregated/test_mamba_transfer.py   |  1 -
 .../disaggregated/test_py_cache_transceiver_mp.py     |  1 -
 tests/unittest/others/test_kv_cache_transceiver.py    |  1 -
 8 files changed, 3 insertions(+), 15 deletions(-)

diff --git a/tests/unittest/disaggregated/test_agent.py b/tests/unittest/disaggregated/test_agent.py
index ad4688ef2f0d..c8425ab49e18 100644
--- a/tests/unittest/disaggregated/test_agent.py
+++ b/tests/unittest/disaggregated/test_agent.py
@@ -9,7 +9,6 @@
 # Force a deterministic UCX/NIXL config regardless of what the cluster/CI
 # injects; see test_kv_transfer.py for the full rationale.
 os.environ["UCX_TLS"] = "^ib,gdr_copy"
-os.environ.pop("UCX_NET_DEVICES", None)
 os.environ["TRTLLM_NIXL_NUM_THREADS"] = "1"
 
 from tensorrt_llm import logger
diff --git a/tests/unittest/disaggregated/test_agent_multi_backends.py b/tests/unittest/disaggregated/test_agent_multi_backends.py
index c8fa0d0b85d6..d82815c99a79 100644
--- a/tests/unittest/disaggregated/test_agent_multi_backends.py
+++ b/tests/unittest/disaggregated/test_agent_multi_backends.py
@@ -7,7 +7,6 @@
 # injects; see test_kv_transfer.py for the full rationale. The subprocesses
 # spawned below inherit these via os.environ.copy().
 os.environ["UCX_TLS"] = "^ib,gdr_copy"
-os.environ.pop("UCX_NET_DEVICES", None)
 os.environ["TRTLLM_NIXL_NUM_THREADS"] = "1"
 
 
diff --git a/tests/unittest/disaggregated/test_cache_transceiver_single_process.py b/tests/unittest/disaggregated/test_cache_transceiver_single_process.py
index 4d3ba3779c0d..26b9e2e9199d 100644
--- a/tests/unittest/disaggregated/test_cache_transceiver_single_process.py
+++ b/tests/unittest/disaggregated/test_cache_transceiver_single_process.py
@@ -31,7 +31,6 @@
 # Force a deterministic UCX/NIXL config regardless of what the cluster/CI
 # injects; see test_kv_transfer.py for the full rationale.
 os.environ["UCX_TLS"] = "^ib,gdr_copy"
-os.environ.pop("UCX_NET_DEVICES", None)
 os.environ["TRTLLM_NIXL_NUM_THREADS"] = "1"
 from dataclasses import dataclass
 from typing import Dict, List, Optional
diff --git a/tests/unittest/disaggregated/test_kv_transfer.py b/tests/unittest/disaggregated/test_kv_transfer.py
index 821fcc431c03..c2cfc1382e37 100644
--- a/tests/unittest/disaggregated/test_kv_transfer.py
+++ b/tests/unittest/disaggregated/test_kv_transfer.py
@@ -6,15 +6,10 @@
 import uuid
 
 # Force a deterministic UCX config regardless of what the cluster/CI injects
-# (enroot may inject UCX_TLS=rc_v,sm + UCX_NET_DEVICES=mlx5_*; the CI agent
-# bootstrap exports UCX_TLS=tcp,cuda_copy,cuda_ipc):
-# - exclude IB (no fabric assumed) and gdr_copy (UCX rcache SIGABRT at
-#   teardown);
-# - drop UCX_NET_DEVICES, otherwise with ^ib the remaining transports cannot
-#   use the injected mlx5_* device list and UCX fails with "no active
-#   messages transport".
+# (the CI agent bootstrap exports UCX_TLS=tcp,cuda_copy,cuda_ipc before pytest
+# starts, which a setdefault would leave in place): exclude IB (no fabric
+# assumed) and gdr_copy (UCX rcache SIGABRT at teardown).
 os.environ["UCX_TLS"] = "^ib,gdr_copy"
-os.environ.pop("UCX_NET_DEVICES", None)
 # Each NIXL agent spawns TRTLLM_NIXL_NUM_THREADS (default 8) busy-polling
 # progress threads, and a single case builds up to 8 TransferWorkers (one per
 # rank). On CI nodes shared with other single-GPU jobs the resulting CPU
diff --git a/tests/unittest/disaggregated/test_kv_transfer_mp.py b/tests/unittest/disaggregated/test_kv_transfer_mp.py
index d58315e1a93a..a52d95bde7a4 100644
--- a/tests/unittest/disaggregated/test_kv_transfer_mp.py
+++ b/tests/unittest/disaggregated/test_kv_transfer_mp.py
@@ -11,7 +11,6 @@
 # Force a deterministic UCX config regardless of what the cluster/CI injects;
 # see test_kv_transfer.py for the full rationale.
 os.environ["UCX_TLS"] = "^ib,gdr_copy"
-os.environ.pop("UCX_NET_DEVICES", None)
 # Limit NIXL busy-polling progress threads; see test_kv_transfer.py for the
 # full rationale (intermittent 120s timeouts on shared CI nodes,
 # https://nvbugs/6426834).
diff --git a/tests/unittest/disaggregated/test_mamba_transfer.py b/tests/unittest/disaggregated/test_mamba_transfer.py
index f844a31597b9..8688d75f9a53 100644
--- a/tests/unittest/disaggregated/test_mamba_transfer.py
+++ b/tests/unittest/disaggregated/test_mamba_transfer.py
@@ -23,7 +23,6 @@
 # Force a deterministic UCX/NIXL config regardless of what the cluster/CI
 # injects; see test_kv_transfer.py for the full rationale.
 os.environ["UCX_TLS"] = "^ib,gdr_copy"
-os.environ.pop("UCX_NET_DEVICES", None)
 os.environ["TRTLLM_NIXL_NUM_THREADS"] = "1"
 
 import tensorrt_llm
diff --git a/tests/unittest/disaggregated/test_py_cache_transceiver_mp.py b/tests/unittest/disaggregated/test_py_cache_transceiver_mp.py
index 50d7a06d0012..d414d18fb48b 100644
--- a/tests/unittest/disaggregated/test_py_cache_transceiver_mp.py
+++ b/tests/unittest/disaggregated/test_py_cache_transceiver_mp.py
@@ -18,7 +18,6 @@
 # Force a deterministic UCX/NIXL config regardless of what the cluster/CI
 # injects; see test_kv_transfer.py for the full rationale.
 os.environ["UCX_TLS"] = "^ib,gdr_copy"
-os.environ.pop("UCX_NET_DEVICES", None)
 os.environ["TRTLLM_NIXL_NUM_THREADS"] = "1"
 
 import tensorrt_llm
diff --git a/tests/unittest/others/test_kv_cache_transceiver.py b/tests/unittest/others/test_kv_cache_transceiver.py
index 86dc2ad6dd7c..241a4aee0d09 100644
--- a/tests/unittest/others/test_kv_cache_transceiver.py
+++ b/tests/unittest/others/test_kv_cache_transceiver.py
@@ -15,7 +15,6 @@
 # Force a deterministic UCX/NIXL config regardless of what the cluster/CI
 # injects; see disaggregated/test_kv_transfer.py for the full rationale.
 os.environ["UCX_TLS"] = "^ib,gdr_copy"
-os.environ.pop("UCX_NET_DEVICES", None)
 os.environ["TRTLLM_NIXL_NUM_THREADS"] = "1"
 
 import tensorrt_llm

From 127e7f4b0b3e9e79bdb0a112832eb175198487f5 Mon Sep 17 00:00:00 2001
From: Chuang Zhu <111838961+chuangz0@users.noreply.github.com>
Date: Fri, 24 Jul 2026 02:43:35 +0000
Subject: [PATCH 4/4] [https://nvbugs/6426834][chore] Drop stale
 test_proxy_fast_death waive re-added during rebase (fixed by #16630)

Signed-off-by: Chuang Zhu <111838961+chuangz0@users.noreply.github.com>
---
 tests/integration/test_lists/waives.txt | 1 -
 1 file changed, 1 deletion(-)

diff --git a/tests/integration/test_lists/waives.txt b/tests/integration/test_lists/waives.txt
index 066878706ad8..c02464b7ecfb 100644
--- a/tests/integration/test_lists/waives.txt
+++ b/tests/integration/test_lists/waives.txt
@@ -363,7 +363,6 @@ unittest/_torch/visual_gen/multi_gpu/test_ulysses_async.py::test_capture_smoke S
 unittest/_torch/visual_gen/test_attention_integration.py::test_sage_attention_self_attention[fp8-2-1560] SKIP (https://nvbugs/6198760)
 unittest/_torch/visual_gen/test_attention_integration.py::test_sage_attention_self_attention[int8-2-1560] SKIP (https://nvbugs/6198760)
 unittest/auto_deploy/multigpu/custom_ops SKIP (https://nvbugs/6403920)
-unittest/executor/test_proxy_fast_death.py::test_pool_session_shutdown_never_blocks_after_release SKIP (https://nvbugs/6480574)
 unittest/executor/test_rpc.py::TestRpcCorrectness::test_incremental_task_async SKIP (https://nvbugs/5741476)
 unittest/executor/test_rpc_proxy.py SKIP (https://nvbugs/5605741)
 unittest/executor/test_rpc_worker.py SKIP (https://nvbugs/5605741)