laya.cpp r0004

Raw Linux and Windows x64 plus macOS arm64 binaries; no installer.

Commit: e632da0428236be56978e53ddc92cdcd565fb493

Choose CUDA 12, CUDA 13, Vulkan or Core ML. See RUNTIME-REQUIREMENTS.md for external libraries.
Automated build/host tests passed; these rolling prereleases are not GPU correctness certifications.

Changes

• Add independent CUDA 12 and CUDA 13 binary release profiles (95f200e)

laya.cpp r0003

Raw Linux and Windows x64 plus macOS arm64 binaries; no installer.

Commit: 58c68ae72849e9f03a6629bbd9446edf0e82fce8

Choose CUDA, Vulkan or Core ML. See RUNTIME-REQUIREMENTS.md for external libraries.
Automated build/host tests passed; these rolling prereleases are not GPU correctness certifications.

Changes

• Clarify HTTP model discovery contract (8f26e10)
• Reject truncated requests by default (0ebd5d4)
• Skip scheduled releases for non-release changes (e1c6e78)
• Run legacy build checks only on manual dispatch (fc41d44)

laya.cpp r0002

Raw Linux and Windows x64 plus macOS arm64 binaries; no installer.

Commit: c785d0b6d293de6b6903072131b5f87e52ca06eb

Choose CUDA, Vulkan or Core ML. See RUNTIME-REQUIREMENTS.md for external libraries.
Automated build/host tests passed; these rolling prereleases are not GPU correctness certifications.

Changes

• Document executable permission for macOS downloads (c785d0b)

laya.cpp r0001

Raw Linux and Windows x64 plus macOS arm64 binaries; no installer.

Commit: 38b2055494a24105cc3fa89d1075046765bd2f10

Choose CUDA, Vulkan or Core ML. See RUNTIME-REQUIREMENTS.md for external libraries.
Automated build/host tests passed; these rolling prereleases are not GPU correctness certifications.

Changes

• Skip CUDA toolkit setup when reusing tested executable (38b2055)
• Document macOS binary releases in README (03090bd)
• Add standalone macOS Core ML binary release target (39250db)
• Split binary builds into reusable platform workflows (1ec2d77)
• Audit Windows CUDA runtime-loaded driver dependencies correctly (d4455d8)
• Write release metadata and notices as UTF-8 on Windows (2346d2a)
• Include string explicitly in Vulkan precision test (9c7d145)
• Declare the project language level for all native test targets (eb31534)
• Ship license text referenced by the static C++ runtime notices (604e19d)
• Set C++17 explicitly for custom CUDA BF16 compilation (654344a)
• Include Vulkan and SPIR-V header notices with binary releases (db1492b)
• Keep diagnostic builds separate from serialized release publication (3d1ba1c)
• Install complete Windows Vulkan SDK instead of extracting tool files (baedc93)
• Cache completed static dependencies before GPU compilation (2c69d0a)
• Require C++20 for the Vulkan precision test on release compilers (b3b8677)
• Check release notices and cover unchanged-commit gating (3ab82c5)
• Use packaged MSVC setup action for Windows release builds (ecdb7cd)
• Automate twice-daily raw Windows and Linux GPU binary releases (5b2d301)
• Add portable static application builds and Windows environment support (550b873)
• Add contributor guidance for agents (8d01c3b)
• Add reproducible Core ML performance sweep (5dc5166)
• Build Core ML on Apple Silicon in CI (cfd8b86)
• Automate Core ML builds on macOS (a1ba17f)
• Add Apple Core ML inference for macOS (5763d96)
• Simplify README and consolidate latest backend measurements (3ca6a32)
• Document completed optimized AMD BF16 performance and Vulkan regression (b1d4a52)
• Record Vulkan regression after BF16 scan optimization (519c476)
• Record optimized AMD BF16 versus Python performance for typed-decisions-bf16 (231b416)
• Record optimized AMD BF16 versus Python performance for multilingual-bf16 (f007e2f)
• Record optimized AMD BF16 versus Python performance for english-bf16 (90d2215)
• Replace stale Vulkan measurement status with completed result links (b33429f)
• Extend BF16 residual tests and record rejected cache optimization (d9083fe)
• Document complete AMD BF16 subgroup scan acceptance (3c0fc9a)
• Record AMD BF16 parallel scan correctness for typed-decisions (6e4a971)
• Record AMD BF16 parallel scan correctness for multilingual (8d4827b)
• Record AMD BF16 parallel scan correctness for english (505ff11)
• Parallelize AMD BF16 scaling scans across GPU subgroups (8efb869)
• Link refreshed native CUDA versus Vulkan BF16 comparison (fc522f4)
• Record integrated Vulkan operator regression on both GPUs (7d310f3)
• Record final native CUDA versus Vulkan BF16 performance for typed-decisions (a7f2ad3)
• Record final native CUDA versus Vulkan BF16 performance for multilingual (b53ada0)
• Link complete optimized RTX 16-bit Python measurements (b0c955a)
• Record final native CUDA versus Vulkan BF16 performance for english (dd63f30)
• Record NVIDIA Vulkan versus Python performance for typed-decisions-bf16 (8c7ab3c)
• Record NVIDIA Vulkan versus Python performance for multilingual-bf16 (28d663a)
• Record NVIDIA Vulkan versus Python performance for english-bf16 (e6e4e16)
• Record NVIDIA Vulkan versus Python performance for typed-decisions-fp16 (5dbcbbe)
• Record NVIDIA Vulkan versus Python performance for multilingual-fp16 (610eecf)
• Record NVIDIA Vulkan versus Python performance for english-fp16 (5f46842)
• Summarize complete AMD 16-bit Python baseline measurements (9116e0a)