Here's a summary of all changes applied to installer/generated/post_install_v1.8.7b0.py:

        #                                                           Change                                                           Status

   1    verify_existing_torch_stack(): creates GPU constraints for pre-existing torch

   2    install_pytorch_and_torchaudio(): use_constraints=False

   3    install_pytorch(): creates GPU constraints after torch verification

   4    install_llama_cpp_python(): removed [server] extras calls, added installed_backend tracking, added
        _pin_llama_cpp_version() calls, passes expected_backend to validation

   5    Added _pin_llama_cpp_version() function

   6    Rewrote validate_llama_cpp_installation() with expected_backend param, trust-the-installer approach

   7    Removed _check_llama_cpp_has_gpu() entirely

Author: meizhong986

docs/release_notes_v1.8.7b0.md

@@ -0,0 +1,38 @@
+# v1.8.7-beta — Improvements & Bug Fixes
+
+Pre-release for testing. Feedback welcome via [Issues](https://github.com/meizhong986/WhisperJAV/issues).
+
+---
+
+## Enhancements
+
+- **Local LLM translation reliability** — Fixed token limit handling for small-context models (8K). Streaming is now correctly enabled for local LLM servers, and max_tokens is properly capped to prevent truncated translations. (#196)
+
+- **Apple Silicon (MPS) support** — Metal GPU acceleration now works across more components:
+  - Transformers pipeline detects and uses MPS instead of falling back to CPU (#198)
+  - Speech enhancement backends (ZipEnhancer, BS-RoFormer) try MPS with automatic CPU fallback
+  - Installer correctly installs PyTorch with MPS support on Apple Silicon (no longer pulls CPU-only wheels)
+  - Prebuilt Metal wheel support for llama-cpp-python on Mac
+
+- **VAD controls** — New `--vad-threshold` and `--speech-pad-ms` CLI flags for direct control over voice activity detection sensitivity, with per-pass overrides for ensemble mode. (#159)
+
+- **Enhanced VAD for ChronosJAV pipelines** — Decoupled and Qwen pipelines now support speech segmentation with Silero v6.2 and TEN backends, giving these pipelines the same VAD quality as the classic balanced pipeline.
+
+- **ZipEnhancer dependency fix** — The `enhance` extra now correctly declares all required ModelScope dependencies (einops, oss2, addict, attrs, datasets, etc.), eliminating manual pip installs after `pip install whisperjav[enhance]`.
+
+## Bug Fixes
+
+- **#196** — Local LLM translation failing with "No matches found" on 8K context models. Three root causes fixed: max_tokens cap, forced streaming, and streaming support flag in CustomClient.
+- **#198** — Apple Silicon Macs running Transformers pipeline on CPU instead of MPS. MPS device detection added with proper float16 dtype handling.
+- **#195** — UnicodeDecodeError when processing files with Japanese metadata (M4A/M4B). Added `errors='replace'` to FFmpeg subprocess calls. *(fix backported to v1.8.6)*
+
+## Installation
+
+```bash
+pip install "whisperjav @ git+https://github.com/meizhong986/whisperjav.git@v1.8.7-beta"
+```
+
+Or for Mac users:
+```bash
+curl -fsSL https://raw.githubusercontent.com/meizhong986/whisperjav/main/install_mac.sh | bash
+```