[JinyanYi]

#43711 superseded the tensor-bool fix by capturing transfer buffers during host-pool initialization without evaluating their truth value. The conflict has been resolved in favor of the new main implementation, and the redundant MHA/MLA changes have been removed from this PR.

Motivation

This PR addresses the following NPU CI issues:

1. NPU no longer supports --enable-torch-compile, but a dedicated compile test remained in registered CI, while DP-attention and load-balance tests also enabled the flag even though compile coverage was not their test target.

2. NPU CI left pyarrow unconstrained. After pyarrow==26.0.0 entered the package index, it was installed alongside the numpy==1.26.4 required by triton-ascend and failed at import time because PyArrow 26 requires NumPy 2 or newer. GPU CI currently uses the known-good pyarrow==23.0.1 with the same NumPy version.

The MLA HiCache test also requested 30 GB of pinned host memory per TP rank. With TP=4 that is approximately 120 GB, which can fail in aclrtMallocHostWithCfg on the CI host.

Modifications

2. Reduce test_npu_hicache_mla.py from 30 GB to 10 GB of pinned host memory per TP rank.

3. Remove test_npu_compile_graph_tp1_bf16.py, which existed specifically to cover the unsupported NPU --enable-torch-compile path.

4. Remove --enable-torch-compile and its compile-only max-batch-size option from the DP-attention and load-balance tests. Their test logic, targets, and assertions are otherwise unchanged.

5. Pin pyarrow==23.0.1 in the shared NPU CI dependency installer, matching the version used by GPU CI and preventing an incompatible unbounded upgrade to PyArrow 26.

Accuracy Tests

The targeted NPU HiCache PR smoke test passed on the updated branch:

GITHUB_EVENT_NAME=pull_request python3 test/registered/npu/basic_function/HiCache/test_npu_hicache_mla.py -v
Ran 1 test in 117.649s
OK

Test environment: CANN 9.1 A3 B082 image on four Ascend NPUs. All four TP ranks allocated a 10.00 GB Host KV pool, and the smoke generation requests completed successfully.

The NPU dependency-script change passes bash -n and git diff --check.

Speed Tests and Profiling

None.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

---

CI States

Latest PR Test (Base): ⏳ Run #38223489540
Latest PR Test (Extra): ❌ Run #38223489108
Latest PR Test (AMD ROCm 10): ⏳ Run #38223489405

[sglang-npu-bot]

failed CI test perf-2 and perf-16 can be fix by PR 43309

[sglang-npu-bot]

paas all GPU CI test: https://github.com/sgl-project/sglang/actions/runs/38191524617/job/114702181771?pr=43308

[JinyanYi]

#43711 Refactored the part where Boolean value of Tensor with more than one value is ambiguous error was possible, and is a better implementation. Therefore we will revert what we changed.