[Perf] Use the non-draining NPUStream getter in the kernel launch paths

[zihanlin-ai]

Fixes #219

What changed

Replace all six remaining
c10_npu::getCurrentNPUStream().stream() calls in kernel-launch paths with
stream(false).

stream() drains the torch_npu task queue before returning. These streams are
only consumed by OpCommand custom handlers enqueued on the current stream, so
queue and device-stream ordering already preserve dependencies. stream(false)
returns the same aclrtStream without the redundant drain.

csrc/pac_kernels.cpp already uses this form. csrc/framework_hal.cpp is
unchanged because device_index() does not drain the queue.

Validation

Environment: 910B2, CANN 8.5.1, torch_npu 2.9.0.post1.

Host-side multi_layer_kv_transfer latency over 200 iterations:

Operation	Pending ops	Mean (us)	P99 (us)	Max (us)
Scatter	0	127.8 → 18.0	163.4 → 34.1	236.6 → 34.7
Scatter	32	33.9 → 22.4	51.6 → 33.1	52.3 → 79.0
Gather	0	88.6 → 20.4	107.4 → 23.6	138.8 → 24.0
Gather	32	27.9 → 25.7	57.4 → 33.0	1098.1 → 33.1

• Three repeats measured 85–128 us before and 16–18 us after.
• All affected operations produced identical checksums; scatter/gather
  round trips were bit-exact.
• Device execution time was unchanged.
• With TASK_QUEUE_ENABLE=0, both builds measured 40–44 us.

csrc/mem_kernels.cpp and csrc/utils.cpp also overlap with #246; I can rebase
if needed.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-08-29T03:00:38.216017Z	32d4862	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[zihanlin-ai]

Also ran the repository's own kernel suites on 910B2 / CANN 8.5.1 / torch_npu 2.9.0.post1: tests/v1/test_mem_kernels.py plus tests/v1/test_pos_kernels.py give 341 passed on upstream main and 341 passed on this branch.

[matthewygf]

Thanks for the contribution, I was wondering whether you can add a simple integration test with gsm8k or the likes ? Wdyt ? @chloroethylene

[chloroethylene]

Could you please provide some information on the testing results regarding accuracy, such as on MMLU or GSM8K?

[zihanlin-ai]

Validated 169ee728 vs 32d4862f on Qwen3-8B/A3 with 5-shot MMLU (250 questions, LMCache cold+warm): accuracy was 66.4% vs 66.4%, with identical predictions/token IDs. Cold p95: 50.96→50.57 ms; warm p95: 49.15→49.58 ms; total time: +0.01%. No accuracy or end-to-end performance regression observed.

[zihanlin-ai]

GSM8K (250, Qwen3-8B/A3): base/head 85.6%; all outputs identical; wall +0.02%. No regression.

[zihanlin-ai]

@matthewygf @chloroethylene FYI—MMLU/GSM8K results above.

[chloroethylene]

lgtm! thanks for your contributions!