fix(ascend): keep unsafe operators out of device graphs

[baominghelly]

Summary

• Mark Ascend HCCL AllReduce as unsafe for device-graph capture in csrc/infinicore/src/ops/distributed/allreduce.cc.
• Mark Ascend MultiheadAttentionVarlen as unsafe for device-graph capture because its provider synchronizes cumulative sequence lengths on the host.
• Add static runtime-contract tests covering both graph segmentation decisions.

Motivation

Ascend graph replay produced corrupted model output when host-dependent attention and HCCL collective operations were captured into device-graph segments. The graph runtime already supports splitting operator sequences by GraphOperator::is_device_graph_capture_safe(); this change uses that existing mechanism to execute the two currently unsafe Ascend operator classes eagerly between captured segments.

This is a correctness fallback rather than native graph support for those operators. Once the relevant backends expose validated graph-capture capability, InfiniCore should query that capability instead of checking the device type here.

No issue is currently linked to this change.

Type of Change

• feat — new feature / new model
• fix — bug fix
• perf — performance improvement (no behavioral change)
• refactor — code restructuring without behavior change
• test — adding or fixing tests only
• docs — documentation only
• build / ci — build system or CI configuration
• chore — tooling, formatting, or other non-code changes
• Breaking change

Test Results of Involved Models on Supported Platforms (Please attach screenshots)

Validated on Ascend 910C with CANN 9.0.0, the modern InfiniRT/InfiniOps stack, InfiniOps paged flash-attention PR #984, and InfiniLM refactor/adopt-modern-infini-stack.

Single-request graph tests used examples/test_infer.py with --device ascend --enable-paged-attn --attn=flash-attn --enable-graph.

ChatGLM3-6B TP1: exit 0; coherent self-introduction; segmented graph runtime enabled
InternLM3-8B-Instruct TP1: exit 0; coherent response; segmented graph runtime enabled
MiniCPM4-8B TP1: exit 0; coherent model introduction; segmented graph runtime enabled
ChatGLM3-6B TP2: exit 0 in 53 s; communicator established; segmented graph runtime enabled; coherent self-introduction

The final-head TP2 log is /workspace/model-matrix-route-a-20260910/infinilm-graph-pr-final/chatglm3_tp2_graph.log on the validation host.

Benchmark / Performance Impact

N/A. This is a correctness change. No performance claim is made.

Notes for Reviewers

• AllReduce remains capture-safe on all existing platforms except Ascend; other platform behavior is unchanged.
• MultiheadAttentionVarlen was already excluded on NVIDIA and is now also excluded on Ascend.
• The long-term replacement for these device checks is an explicit backend capability query after HCCL graph capture/replay and host-free attention metadata handling are validated.
• Diagnostic changes to graph.cc were intentionally excluded.

CI / ChatOps

CI does not run automatically on pull requests. It should be triggered manually from the Actions tab or by a maintainer with /retest or /test.

---

Checklist

Title, Branch, and Commits

• PR title follows Conventional Commits.
• Branch name follows <type>/xxx-yyyy-zzzz.
• Each commit message follows Conventional Commits.
• Small PR is a single squashable commit.
• Branch is rebased cleanly on the current target branch.
• No fixup! / squash! / wip commits remain.
• Existing PR/branch/commit follows the current contribution format.

Scope and Design

• Changes are minimal and limited to Ascend graph segmentation safety.
• No dead code, commented-out blocks, debug prints, or unrelated formatting churn.
• No unrelated formatting churn.
• No public API changes.

General Code Hygiene

• Code is self-explanatory; comments explain non-obvious backend contracts.
• Every modified file ends with a single trailing newline.
• No trailing whitespace, tab/space mixing, or stray BOMs.
• Identifiers in added comments and messages use Markdown backticks where appropriate.
• Added comments and messages are in English.
• Added comments and messages follow language conventions.

C++ Specific

• Code follows the existing C++ style.
• Error and warning wording follows existing conventions.
• Constructor initializer-list order matches member declaration order.
• No new raw new/delete was introduced.
• Changed files pass scripts/format.py.
• No changes to csrc/models/llama_legacy/.

Python Specific

• Test code follows PEP 8.
• Added Python code is formatted by scripts/format.py.
• No docstring changes were needed.
• No changes to python/infinilm/auto_config.py.

Testing

• Affected Ascend platform tested; other platforms are unchanged by this diff.
• Passed single-request tests with graph mode enabled.
• Offline performance test not run; this correctness change makes no performance claim.
• Sanity suite not run; targeted graph and full static contract validation covered the changed runtime contracts.
• Service test not run; no server behavior was changed.
• Build and install passed on Ascend.

Build, CI, and Tooling

• _infinilm rebuilt and installed successfully on Ascend.
• CI still needs to be triggered manually.

Documentation

• No user-facing build flag, API, or workflow documentation change is needed.

Security and Safety

• No secrets, access tokens, internal URLs, customer data, or personal hardware identifiers were committed.
• No third-party code was introduced.
• No unsafe pointer arithmetic, uninitialized reads, or missing bounds checks were introduced.

Validation Evidence

python scripts/format.py --check --path <four changed files>
# passed

python -m unittest discover -s test/static -p 'test_*.py' -q
Ran 106 tests in 0.169s
OK

xmake build -r -j16 _infinilm
xmake install -y _infinilm
# build ok, install ok

python test/static/test_infinicore_runtime_contracts.py \
  InfiniCoreRuntimeContractsTest.test_ascend_hccl_allreduce_stays_out_of_device_graph \
  InfiniCoreRuntimeContractsTest.test_ascend_attention_with_host_lengths_stays_out_of_device_graph
# Ran 2 tests, OK