feat: Metal GPU backend for Apple Silicon

[discordwell]

Summary

Adds a native Metal compute shader backend for GPU-accelerated histogram construction on Apple Silicon Macs. This fills the gap left by the deprecated OpenCL backend which crashes on Apple Silicon (#6189).

• New device_type="metal" option for training on Apple Silicon GPUs
• Three Metal compute kernels ported from the OpenCL originals (histogram16/64/256)
• MetalTreeLearner using Objective-C++ with unified memory (no pinned buffers needed)
• Runtime .metal source compilation when build-time metallib is unavailable
• 14-test suite covering binary/regression/multiclass, all 3 bin variants, scalability, bagging

Design

Follows the existing OpenCL GPU backend architecture:

• MetalTreeLearner extends SerialTreeLearner — only histogram construction on GPU
• Split finding, data partitioning, tree construction remain on CPU
• Apple Silicon unified memory eliminates PCIe transfer overhead (simpler than OpenCL)

Test results (Apple M4 Max)

14 passed in 1.60s
- Binary classification, regression, multiclass ✓
- histogram16 (max_bin≤16), histogram64 (max_bin≤64), histogram256 (max_bin≤255) ✓
- Dataset sizes 100 to 10,000 ✓
- Bagging, constant hessian, gpu_use_dp override ✓
- Max prediction diff vs CPU: <5e-6 (for n≥1000)

Known limitations

• Single workgroup per feature (POWER=0) — limits GPU occupancy for large datasets. Multi-workgroup reduction kernels are included but have a layout mismatch that needs investigation. This will be addressed in a follow-up.
• Metal Toolchain must be installed for build-time kernel compilation; falls back to runtime compilation from .metal source files.
• macOS only, Apple Silicon only, FP32 only (no gpu_use_dp).

Build

cmake -DUSE_METAL=ON ..
make -j

Usage

params = {'device': 'metal', 'objective': 'binary', ...}
model = lgb.train(params, train_data)

Test plan

• Binary classification correctness vs CPU
• Regression correctness vs CPU
• Multiclass correctness vs CPU
• All 3 histogram kernel variants (max_bin 15/63/255)
• Dataset sizes 100 to 10,000
• Bagging mode
• Constant hessian path
• gpu_use_dp=true gracefully degrades to FP32
• Benchmark on multiple Apple Silicon generations (M1/M2/M3/M4)
• Multi-workgroup reduction for large datasets

🤖 Generated with Claude Code

[discordwell]

Closing — opened prematurely, needs more work on multi-workgroup reduction before it's ready for review.

[jameslamb]

Replaced by #7215, locking this.