fix naive mha out-of-bound read

[nihui]

No description provided.

[tencent-adm]

Thank you for your submission, we really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 94.70%. Comparing base (bd58421) to head (81db1d5).

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #6812      +/-   ##
==========================================
- Coverage   95.13%   94.70%   -0.43%     
==========================================
  Files         947      947              
  Lines      416258   416884     +626     
==========================================
- Hits       396004   394808    -1196     
- Misses      20254    22076    +1822     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Pull request overview

Fixes an out-of-bounds read in the MultiHeadAttention KV-cache path when decoding with cur_seqlen > 1, and extends the corresponding tests to cover multi-token decode.

Changes:

• Copy only past_seqlen elements from cached K/V into the new k_affine/v_affine buffers (instead of dst_seqlen).
• Update KV-cache decode test helpers to accept a configurable cur_seqlen.
• Add decode test cases with cur_seqlen = 32 for both fp32 and int8 paths.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
tests/test_multiheadattention_kvcache.cpp	Extends decode tests to cover multi-token decode (cur_seqlen > 1) for fp32 and int8.
src/layer/multiheadattention.cpp	Fixes KV-cache memcpy length to prevent reading past cached K/V buffers during decode.

---

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