[BUG] Crash in PD backend init due to NIXL change in LMCache v0.4.5

[marcobarlo]

Describe the bug

Ascend PD prefiller crashes on first peer connect with
AttributeError: 'AscendPDBackend' object has no attribute '_nixl_agent_lock'.
AscendPDSenderMixin._ensure_peer_connection called
super()._ensure_peer_connection(...), which resolved (via MRO) to upstream
PDBackend._ensure_peer_connection. That path serializes with
self._nixl_agent_lock, created only in PDBackend.__init__.
AscendPDBackend.__init__ reimplements init for HCCL/NPU and never calls
PDBackend.__init__, so the lock is never created. Note:
super()._init_sender() only opens the proxy ZMQ PUSH socket and does
not create _nixl_agent_lock.

To Reproduce

Steps to reproduce the behavior:

1. Launch Ascend PD disagg (prefiller kv_producer + decoder kv_consumer,
   transfer_channel: hccl), e.g. tests/run_pd_disagg_deepseek_v4.sh.
2. Send a chat/completions request through the LMCache disagg proxy
   (e.g. tests/query3.sh).
3. Prefiller hits batched_submit_put_task → _ensure_peer_connection.
4. See AttributeError: ... '_nixl_agent_lock'; proxy hangs waiting for
   KV-ready ZMQ notif.

Expected behavior

Peer connect should use the HCCL transfer channel
(lazy_init_peer_connection + alloc REQ socket) without touching NIXL
agent state or _nixl_agent_lock, then complete PD KV transfer.

Screenshots

N/A (log-only failure).

Platform & Software (please complete the following information):

• OS: Ubuntu 22.04.5 LTS (container, aarch64)
• CANN Version: 9.0.0
• Ascend HDK Version: [fill from local Ascend install]
• Hardware: Ascend 910B (TP8 prefiller + TP8 decoder)
• Framework: PyTorch / vLLM-Ascend
• vLLM Version: ~0.20.2 (Ascend build)
• LMCache / LMCache-Ascend: v0.4.5-aligned Ascend plugin
• Model: DeepSeek-V4-Flash (MLA)

Additional context

• Upstream commit that introduced _nixl_agent_lock:
  7f60057ce37102bf7e7a519901930d6b9a874136
  (feat(pd_backend): add bidirectional NIXL cache probe / #2972,
  Ziwen Ning, 2026-04-26). First release tag containing it: v0.4.5
  (also v0.4.5-cu129 / v0.4.6+). Absent from v0.4.4.
• Why upstream needed it: #2972 added a dedicated NIXL worker thread +
  queue so KV xfer does not block the vLLM worker (CUDA context /
  RPC timeouts on vLLM v0.19 multiprocess). _nixl_agent_lock serializes
  main-thread peer handshake vs worker-thread batched_write /
  batched_read on shared nixl_agent state.
• Do we need it on Ascend? No _nixl_agent_lock. Ascend PD has no
  NIXL worker thread/queue and does not share a NIXL agent. HCCL channel
  already serializes peer/state mutations via transfer_channel._state_lock.
  Copying the NIXL lock would paper over a wrong super() call without
  matching the async NIXL design.
• MRO: AscendPDBackend(AscendPDSenderMixin, AscendPDReceiverMixin, PDBackend).
  Mixin super() → PDBackend NIXL peer-connect.
• Root cause: Ascend stub (from Enable Direct rH2D / D2D KVtransfer in P2P and PD backend  #173, before #2972) delegated to upstream
  after LMCache ≥ v0.4.5; Ascend __init__ never constructs the lock.
• Fix: override _ensure_peer_connection with HCCL peer init; do not call
  super()._ensure_peer_connection.
• logs:

prefiller: AttributeError: 'AscendPDBackend' object has no attribute '_nixl_agent_lock'
  File ".../pd_backend.py", line ..., in _ensure_peer_connection
    with self._nixl_agent_lock:
prefiller: wait_for_save failed ... skipping save
proxy: waiting for KV ready (hang)