fix: propagate TP KV mapping failures

[shipiyouniao]

Summary

Fixes #372.

This PR tightens TP IPC error handling around KV tensor mapping:

• propagate worker-side map_to_kv_tensors / unmap_from_kv_tensors failures instead of replying success
• preserve pp_rank and group_id from the KVCacheManager when checking KV tensor creation and when invoking map/unmap callbacks
• include the last observed KV tensor creation exception in timeout errors
• synchronize CUDA after TP map and before TP unmap so mapped pages are visible before schedulers touch them, and unmap follows the single-process allocator safety ordering

Why

Without this, the coordinator can lose the real map/unmap failure and later report only a generic KV tensor wait timeout. In TP/PP deployments the default callback context can also make the wait path observe the wrong group. The additional synchronization mirrors the allocator ordering used outside the TP IPC path.

Validation

python -m pytest tests/test_kvcache_manager_post_init.py tests/test_tp_ipc_unmap_sync.py -q
python -m py_compile kvcached/kv_cache_manager.py kvcached/tp_ipc_util.py tests/test_kvcache_manager_post_init.py tests/test_tp_ipc_unmap_sync.py

[Copilot]

Pull request overview

This PR addresses issue #372 by tightening TP IPC error propagation for KV tensor map/unmap operations and improving KV tensor creation diagnostics so the coordinator doesn’t lose the root-cause error behind a generic timeout.

Changes:

• Propagate worker-side map_to_kv_tensors / unmap_from_kv_tensors failures through the IPC response instead of always returning success.
• Ensure broadcast map callbacks use the KVCacheManager’s recorded pp_rank / group_id context.
• Augment KV tensor creation wait timeouts with the last observed exception to improve debuggability.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
kvcached/tp_ipc_util.py	Makes worker IPC replies reflect map/unmap failures so callers receive the real error.
kvcached/kv_cache_manager.py	Fixes callback context usage and improves timeout error reporting when waiting for KV tensors.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.