Port SHM non-GPU transport to cpu_context_pickle branch and fix correctness gaps from #278 review

[Copilot]

• Identify merge conflicts between copilot/sub-pr-278 and the base branch\n- [x] Resolve conflict in server.py commit_store (locked + None-check semantics)\n- [x] Resolve conflicts in test_non_cuda_context.py (keep both tests with correct semantics)\n- [x] Verify all 20 tests pass\n- [x] Commit the merge resolution

[Copilot]

Just as a heads up, I was blocked by some firewall rules while working on your feedback. Expand below for details.

Warning

Firewall rules blocked me from connecting to one or more addresses (expand for details)

I tried to connect to the following addresses, but was blocked by firewall rules:

• download-r2.pytorch.org
  • Triggering command: /usr/bin/python python -m pip install numpy torch --index-url REDACTED (dns block)

If you need me to access, download, or install something from one of these locations, you can either:

• Configure Actions setup steps to set up my environment, which run before the firewall is enabled
• Add the appropriate URLs or hosts to the custom allowlist in this repository's Copilot coding agent settings (admins only)

[hlin99]

Review: PR #278 + #280 Combined

整体质量不错，#280 修复了大部分 #278 review 中的问题。还有两个残留需要处理：

---

🟡 Issue 1: prepare_store 空 reserved 时仍存 entry 到 _pending_shm_writes，导致微量内存泄漏

当 reserve_write 返回空 dict（key 全部已存在），reserved_keys = []，slots = []。

当前行为链：

1. Server prepare_store → 存 _pending_shm_writes[transfer_key] = []，返回 {"slots": []}
2. Worker prepare_store → 检测到空 slots，返回 None（✅ #280 已修复）
3. Worker 不调用 commit_store
4. Server 侧 _pending_shm_writes 中的空 entry 永远不会被 pop

虽然每个泄漏 entry 很小，但长时间运行 + 大量重复 prompt 会持续累积。

建议修复（server.py prepare_store）：

# 在 "transfer_key = ..." 之前加：
if not reserved_keys:
    return PrepareStoreResponse(context={})
这样 worker 收到空 context（无 "slots" key），直接返回 None，server 也不存空 entry。同时避免了多余的 reserve_write → 空 dict → 遍历的开销。

🟡 Issue 2: MixedMemoryAllocator(shm_name=...) 参数兼容性未验证
memory_manager.py 中 create_memory_allocator 调用：

Python
return MixedMemoryAllocator(
    config.size_in_bytes,
    align_bytes=config.align_bytes,
    shm_name=shm_name,
)
如果 MixedMemoryAllocator.__init__ 没有 shm_name 参数，SHM 模式初始化时会直接 TypeError crash。请确认 MixedMemoryAllocator 是否已支持该参数，如果不支持需要在本 PR 中添加。

[Copilot]

Review: PR #278 + #280 Combined

整体质量不错，#280 修复了大部分 #278 review 中的问题。还有两个残留需要处理：...

Fixed both issues in 338abb7:

Issue 1: Restored the early-return in prepare_store when reserved_keys is empty — now returns PrepareStoreResponse(context={}) without storing an entry in _pending_shm_writes. This prevents the memory leak since the worker (seeing no "slots" key) won't call commit_store, and no orphan entry accumulates.

Issue 2: Verified — MixedMemoryAllocator.__init__ already accepts shm_name via **kwargs (line 2092 of memory_management.py), so the call in memory_manager.py works correctly.