[luyuzhe111]

During MigrationBench agent training with Qwen3-Coder-30B-A3B-Instruct on verl, an AgentLoopWorkerTQ stopped making progress in async_kv_batch_put. Its asyncio thread was blocked in zmq.Context.term() inside TransferQueue's per-request socket cleanup, preventing pending rollouts from completing. The trainer remained waiting for the batch, stalling the entire training run.

Pin TransferQueue to 0.1.10 in the verl extra and verl-ci group, and update the lockfile. This picks up Ascend/TransferQueue#145, which reuses a long-lived ZMQ context and closes per-request sockets with linger=0, removing context termination from the request path.

Upstream verl also adopted 0.1.10 in its requirements and CI in verl-project/verl#7951, with a minimum version of >=0.1.10 in its pyproject.toml.

Validation:

• MigrationBench agent training with Qwen3-Coder-30B-A3B-Instruct completed 15 training steps after the upgrade, including step-15 validation and checkpoint saving, without reproducing the hang.
• uv lock --check --offline and git diff --check passed; all other locked packages are unchanged.
• pytest tests/backends/verl -q: 95 passed with verl 0.9.0 and TransferQueue 0.1.10.
• TransferQueue 0.1.10 smoke test: 8 Ray workers completed 512 concurrent KV lifecycles, including writes, reads, status updates, and cleanup; no entries remained.