Skip to content

Add macOS 15 CI target and fix ARM i8mm feature conflict#4

Merged
bernardladenthin merged 1 commit into
masterfrom
claude/fix-macos-15-arm-features-HPebZ
Mar 22, 2026
Merged

Add macOS 15 CI target and fix ARM i8mm feature conflict#4
bernardladenthin merged 1 commit into
masterfrom
claude/fix-macos-15-arm-features-HPebZ

Conversation

@bernardladenthin

Copy link
Copy Markdown
Owner

Summary

This PR adds comprehensive unit tests for the Pair generic utility class and introduces developer documentation to guide contributors working with the codebase.

Key Changes

  • Added PairTest.java: Comprehensive test suite with 18 test cases covering:

    • Basic getter functionality (getKey(), getValue())
    • Null handling for both key and value
    • Equality semantics (same pairs, different keys/values, null comparisons, type mismatches, self-equality)
    • Hash code consistency and null safety
    • String representation with various input types
    • Generic type support with different type combinations (primitives, arrays, complex types)
  • Added CLAUDE.md: Developer guidance document including:

    • Project overview and architecture explanation
    • Build commands for Java (Maven) and native C++ (CMake) components
    • Detailed architecture description of the two-layer design (Java/Native)
    • Parameter flow and native library resolution mechanisms
    • Testing requirements and setup
    • Key constraints and best practices
  • Updated CI workflow: Added macOS 15 runner support with appropriate CMake flags (-DGGML_NATIVE=OFF) and improved artifact naming to be runner-specific for better debugging.

Implementation Details

The PairTest class uses JUnit 4 and covers edge cases including null values, type safety with generics, and proper equals() and hashCode() contract compliance. The test suite ensures the Pair utility class is robust for use throughout the codebase.

The CLAUDE.md file serves as a reference guide for understanding the project structure, build process, and architectural decisions, particularly useful for AI-assisted development workflows.

https://claude.ai/code/session_01PjfeKeUZE8spcY4atNbymE

On macOS 15 (Apple Clang 16), stricter always_inline enforcement causes a
build failure when llama.cpp generates -mcpu=native+dotprod+noi8mm flags
while code includes i8mm-requiring always_inline functions. Adding
-DGGML_NATIVE=OFF for the macos-15 runner disables native CPU feature
detection, avoiding the contradictory i8mm enable/disable state.

Also use runner-specific artifact names to avoid upload conflicts when
multiple macOS matrix jobs fail simultaneously.

https://claude.ai/code/session_01PjfeKeUZE8spcY4atNbymE
@bernardladenthin
bernardladenthin merged commit 737ff83 into master Mar 22, 2026
3 of 5 checks passed
@bernardladenthin
bernardladenthin deleted the claude/fix-macos-15-arm-features-HPebZ branch March 23, 2026 10:05
vaiju1981 pushed a commit to vaiju1981/java-llama.cpp that referenced this pull request Jun 25, 2026
Re-checked all six findings raised after PR bernardladenthin#268 against the pinned upstream
tts.cpp @ b9739. Two were false alarms; four were genuine.

Verified upstream-faithful (NOT divergences) — added provenance comments only:
- bernardladenthin#1 llama_model_n_embd_out: upstream tts.cpp:1042 uses the exact same call.
  Comment now notes it reads the vocoder OUTPUT embedding width, matching upstream.
- bernardladenthin#2 0.25 s silence lead-in: upstream tts.cpp:1077-1080 zeroes the first 0.25 s
  identically. Comment notes it mirrors upstream and that our `i < audio.size()`
  bound is an added safety guard over upstream's fixed 24000/4.

Genuine findings fixed:
- bernardladenthin#3 heavy include in shared header: tts_upstream.h now includes
  <nlohmann/json_fwd.hpp> (forward-declares ordered_json) instead of the full
  <nlohmann/json.hpp>, and drops the json default argument. The single caller,
  tts_engine.cpp, includes the full json and passes an explicit empty object.
  Future includers of the shared interface no longer pull in ~25k lines of json.
- bernardladenthin#4 unasserted duplicate enum: generate-tts-upstream.cmake now captures the
  upstream `outetts_version` enum body and pins its enumerators + order against
  the hand-written copy in tts_upstream.h, so a reorder/rename fails the configure
  instead of silently assigning different integer values across the two TUs.
- bernardladenthin#5 prompt_add overload coverage: the bare `void prompt_add(` prefix de-statics
  all three upstream overloads but only proved >=1 existed. The generator now
  pins (whitespace-tolerant) both overloads the header declares, turning a future
  cryptic link error into a clear configure-time failure.
- bernardladenthin#6 weak WAV assertion: TtsIntegrationTest now parses the RIFF/WAVE header
  (PCM format, mono, 24 kHz, 16-bit), checks chunk-size self-consistency, and
  scans the PCM payload for non-zero samples — so a near-empty or all-silent
  result no longer passes the way `length > 44` did.

Generator regexes validated against the real b9739 source via `cmake -P`
(including a negative drift control); Java test compiles; clang-format 22.1.5
and Spotless both clean. Native build not run here (sandbox proxy blocks the
dependency FetchContent clones) — exercised in CI.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QTQ8mBM9tyKkHbXVBpGwET
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants