[https://nvbugs/6756848][fix] Exchange the handle as its raw CUDA_IPC_HANDLE_SIZE (64-byte) struct bytes…

[trtllm-agent]

Summary

• Root cause: open_ipc_memory read/wrote cudaIpcMemHandle_t.reserved, removed in cuda-python 13.4, which the unbounded cuda-python>=13 pin allows.
• Fix: Exchange the handle as its raw CUDA_IPC_HANDLE_SIZE (64-byte) struct bytes via getPtr(), which both 13.2.0 and 13.4.1 expose, so one path covers the supported range.
• Original test: pytest tests/integration/defs/accuracy/test_llm_api_pytorch.py::TestLlama4SpeculativeDecoding::test_llama4_eagle3[dynamic] -v
• Automated fix generated by repair-bot

Test plan

• Verify fix on the same GPU type as the original failure
• Check for regressions in related tests

Links

• Bug: https://nvbugs/6756848

Reproduction comparison

• Failed commit: no_failed_commit
• ToT: repro_on_tot at d26f734
  Signature: RuntimeError: Executor worker returned error
• Signature relation: not_comparable

Dev Engineer Review

• tensorrt_llm/_ipc_utils.py now serializes CUDA IPC handles as raw 64-byte buffers through getPtr().
• The change removes direct access to cudaIpcMemHandle_t.reserved, restoring compatibility with cuda-python 13.2.0–13.4.1.
• The implementation avoids version-specific branching and preserves byte-identical handle round-tripping.
• Per-file QA perspective: Verify multi-GPU allocation and IPC handle exchange across supported cuda-python versions. Confirm that ctypes.memmove copies exactly CUDA_IPC_HANDLE_SIZE bytes.

QA Engineer Review

No test changes.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 782f066f-e702-475b-a011-b90404bc771e

📥 Commits

Reviewing files that changed from the base of the PR and between a85c26c and 868f983.

📒 Files selected for processing (1)

• tensorrt_llm/_ipc_utils.py

Included review availability: Your plan provides up to 12 included reviews per hour; 10 remain after this review.

---

Walkthrough

The CUDA IPC memory path now serializes handles as raw bytes. It uses ctypes, CUDA_IPC_HANDLE_SIZE, getPtr(), and ctypes.memmove to reconstruct handles.

Changes

CUDA IPC handle exchange

Layer / File(s)	Summary
Raw handle serialization and reconstruction tensorrt_llm/_ipc_utils.py	open_ipc_memory replaces reserved-field access with raw CUDA IPC handle buffers and ctypes.memmove reconstruction.

Priority: ⬇️ Low

Estimated code review effort: 2 (Simple) | ~10 minutes

Change: Bug fix

Merge Risk: ⚪ Minimal · up to 868f9

The IPC handle compatibility update is ready to merge.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 1 functions across 1 files.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title identifies the NVBugs fix and clearly states the main change: exchanging CUDA IPC handles as raw 64-byte struct bytes.
Description check	✅ Passed	The description explains the root cause, fix, affected cuda-python versions, test coverage, bug link, and reproduction results. It uses different headings from the template and does not include the PR…

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands.}