[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>

Author: chuangz0

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

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"
 
 

tests/unittest/disaggregated/test_cache_transceiver_single_process.py

@@ -16,7 +16,6 @@
 # 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

tests/unittest/disaggregated/test_deepseek_v4_kv_transfer.py

@@ -42,7 +42,6 @@
 # default 8 NIXL progress threads per agent cause heavy contention when
 # creating multiple agents on a single GPU in the same process.
 os.environ["UCX_TLS"] = "^ib,gdr_copy"
-os.environ.pop("UCX_NET_DEVICES", None)
 os.environ["TRTLLM_NIXL_NUM_THREADS"] = "1"
 
 

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

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).

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

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

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