New features

• Cache Drama 3 prompt prefixes with LRU eviction, cache controls, and hit/miss diagnostics (#2133, #2144) by @jordimas
• Add native CPU reference-audio encoding for Drama 3 voice cloning (#2141) by @jordimas

Fixes and improvements

• Preserve FP32 attention scores through softmax to fix CUDA alignment precision with FP16 and BF16 models (#2143) by @jordimas
• Improve Drama 3 prefill performance by increasing the default chunk size from 16 to 128, and fuse gated-delta projections during GPU decoding (#2133) by @jordimas
• Fix caching allocator shutdown crashes when the allocating thread exits (#2136) by @jordimas
• Fix FP16 overflow in CUDA attention score accumulation (#2131) by @jordimas
• Fix the eole converter scale for relative position keys (#2138) by @maksimcialehin
• Add a Windows ROCm/HIP build guide and document AMD GPU support (#2041) by T0nd3 and @jordimas
• Add CUDA build instructions for NVIDIA Jetson devices (#2019) by @vimalk78

New features

• Build Python 3.15 wheels (#2129) by @shaanmajid

• Add an experimental implementation of Drama 3 offline text-to-speech, including model conversion and waveform decoding on CPU with optional CUDA and ROCm acceleration by @jordimas

• Add CPU and GPU implementations of CausalDepthwiseConv1D, ConvTranspose1D, DequantizeFP8, GatedDeltaNet, MoeTopK, and Snake operations for audio models (#2121) by @jordimas

See: https://opennmt.net/CTranslate2/guides/drama3.html

Fixes and improvements

• Guard integer overflow in tensor size computation, reshape, and growth (#2097) by @Kyozzz and @jordimas
• Fix CPU Conv1D dilation and INT8 quantization scale overflow for tiny weight rows (#2117) by @jordimas
• Use the caching allocator by default on HIP (#2118) by @jordimas
• Fix Wav2Vec2 lookups of packed linear weights (#2119) by @arcusbuilds
• Fix Wav2Vec2-BERT float16 inference (#2126) by @jordimas
• Rebuild rotary caches when the device or data type changes by @jordimas
• Restore the Transformer include in the public attention header by @jordimas
• Improve CI reliability with updated GitHub Actions and retries for ROCm package index updates and macOS OneAPI downloads (#2125) by @jordimas

New features

• Support converting UMT5 models from Hugging Face Transformers (#2103) by @the-cross-art

Fixes and improvements

• Support Transformer models where each layer has its own relative attention bias by @the-cross-art
• Fix Windows ROCm Translator hangs and DLL loading (#2106) by @0xDELUXA
• Fix crashes when loading model files from Python file-like objects (#2110) by @Purfview
• Upgrade pybind11 to 2.13.6 for CPython 3.14 allocator safety (#2108) by @cepeter
• Guard exhausted candidates in CUDA TopK stage 2 (#2104) by @mzggr0914
• Fix Windows shutdown deadlocks and memory leaks with the Ruy backend (#2076) by @timo9378
• Add Clang ThreadSanitizer testing (#1986) by @3manifold
• Improve ROCm wheel testing and release checks; upgrade build actions to Node.js 24 (#2112) by @jordimas
• Add local CUDA and ROCm release validation with CI wheels and commit status reporting by @jordimas

Fixes and improvements

• Import Python converters and specs lazily to avoid loading torch for inference-only use (#2080) by @Anai-Guo
• Fix out-of-bounds reads in StorageView index validation (#2073) by @jordimas, reported by Nathan Keys (Halo Forge Labs)
• Update cxxopts to v3.3.1 for GCC 15 compatibility (#2084) by @bekard
• Validate model variable sizes before allocation (#2091) by @jordimas, reported by Hyeonbell
• Check StorageView reserve byte size before allocation (#2094) by @infinityscroll
• Remove the runtime dependency on setuptools (#2096) by @Ravencentric

New features

• Support for Gemma4 12B dense model (#2060) by @jordimas

Fixes and improvements

• Harden legacy converter checkpoint loading (#2036) by @igorsantosgrueiro
• Fix model load heap overflow (#2068) by @jordimas, reported by Chegne Eu Joe (Project Umbra)
• Fix process-killing integer division by zero when Whisper align() gets a window with no frames (#2065) by @MAkcanca
• CI/CD compilation fixes for Windows (#2069) by @jordimas
• Optimize attention softmax buffer reuse (#2066) by @jordimas

v4.8.0 (2026-06-06)

New features

• Add support for Google T5Gemma2 (#2058) by @jordimas
• Enables PACKED_GEMM by default for Intel MKL to improve Intel CPU performance (#2055) by @jordimas
• Upgrade Thrust submodule from 1.12.0 to CCCL 2.7.0 (#2062) by @jordimas

Fixes and improvements

• Adopt latest BS_thread_pool library version v5.1.0 (#2053) by @3manifold
• Upgrade CI system dependencies and reduce build time (#2056) by @jordimas
• Fix PyPI publish: download wheels to dist/ to avoid directory hash error (#2050) by @jordimas
• Fix flaky transformers_wav2vec2 test (#1996) by @3manifold

v4.7.2 (2026-05-18)

New features

• Gemma4 support for dense model (#2048) by @jordimas

Fixes and improvements

• Gemma 3 model conversion fixes (#2037) by @jordimas
• Update source ROCM version from 7.2 to 7.2.1 (#2030) by @racedale
• Free curand states before the thread is destroyed (#1912) by @no1d

Fixes and improvements

• Fix Windows build (#2007) @sssshhhhhh

New features

• Introduce AMD GPU support with ROCm HIP (#1989) @sssshhhhhh
• Compatibility with Transformers v5 (#1999) by @jordimas

Fixes and improvements

• Assume less about whisper vocab (#2000) by @sssshhhhhh
• Use LLVM ThreadSanitizer instead of Google (#1993) by @3manifold
• Optimize all builds with parallel execution (#1992) by @3manifold
• Remove unecessary zero init from conv1d (#1990) by @sssshhhhhh
• Integrate Clang AddressSanitizer in tests (#1903) by @3manifold
• Enable multiple of 16 padding for INT8 Tensor Cores (#1982) by @Purfview
• Add activation and dilation to conv1d (#1979) by @sssshhhhhh
• Minor refactor to CMakeLists.txt (#1980) by @sssshhhhhh
• Remove unnecessary check from wav2vec2 (#1977) by @plan9better
• Add optional residual add to gemm op (#1975) by @sssshhhhhh
• Implement cuda layernorm axis (#1971) by @sssshhhhhh
• Fix Eole conversion (#1998) by @vince62s
• Gemma 3 conversion improvements (#1991) by @sssshhhhhh
• Add causal flag to fa2 (#1976) by @sssshhhhhh
• Fixes cross attention tests and refactors code (#1974) by @jordimas
• Fix CUDA bf16 median filter (#1972) by @sssshhhhhh
• Fix various compiler warnings (#1970) by @sssshhhhhh

v4.6.3 (2026-01-06)

New features

• T5Gemma model conversion and inference (#1962) by @jordimas
• Support for CUDA 12.8 (#1937, #1940) by @Purfview
• Conv1d pure CUDA implementation (#1949), makes cuDNN an optional dependency by @jordimas
• Add CUDA implementation for median filter (#1917) by @ja2d8a4v

Fixes and improvements

• Fix issue- tokens batch_type may exceed max_batch_size (#1948) by @anterart
• Fixes AWQ support for Qwen 2.5/3 and Phi3 (#1951, #1952, #1953) by @BBC-Esq