llama : return 0 on invalid ON_DEVICE seq restore

[xiehuanyi]

Overview

Fixes #27439
ON_DEVICE seq restore was reading magic/seq_id before the existing try/catch, so a bad buffer could throw out of the C API or hit GGML_ASSERT. Invalid input now returns 0. Added a CPU test
on the generated llama-dense model: valid round-trip plus the four malformed cases.

Additional information

CPU-only. Did not add a second catch on llama_state_seq_set_data_ext(); the inner try covers the reporter cases.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - used Cursor to locate the try-block gap, write the patch, and add the regression test. I reviewed the diff and can explain it.

[ggml-gh-bot]

Hi @xiehuanyi, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• PR Template not respected: Please respect the template when creating a new pull request. Make sure to fill out all required sections.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[Copilot]

Pull request overview

This PR hardens ON_DEVICE sequence-state restoration against malformed input and adds CPU regression coverage.

Changes:

• Handles invalid headers and sequence IDs through the existing error path.
• Adds valid and malformed restore tests.
• Registers the new test with generated-model fixtures.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Summary
tests/test-state-seq-on-device.cpp	Adds valid round-trip and malformed-buffer regression tests.
tests/CMakeLists.txt	Builds and registers the regression test.
src/llama-context.cpp	Adds ON_DEVICE validation, but malformed payload cleanup can still assert; the failure path must be made non-fatal and covered by a truncation test.

Suppressed comments (1)

tests/CMakeLists.txt:229

• This test is registered inside the NOT WIN32 OR NOT BUILD_SHARED_LIBS guard, which skips it for Windows shared-library builds. Unlike the tests covered by the guard's comment, test-state-seq-on-device.cpp includes only llama.h and exercises exported C APIs, so the new C API regression remains untested on that configuration; register it in an applicable cross-platform test block and provide the model fixture there.

    llama_test(
        test-state-seq-on-device
        LABEL main
        ARGS "${MODEL_DIR}/llama-dense.gguf"
    )
    set_tests_properties(test-state-seq-on-device PROPERTIES
        FIXTURES_REQUIRED generate-models

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.