v0.2.5+amd.2

Added

• Make single prefill example script standalone (#104) @demandal25
• Run tests/test_non_contiguous_prefill.py on HIP (#96) @demandal25
• Run tests/test_activation.py on HIP (#98) @demandal25
• Add more PyTests from upstream to ROCm CI (#93) @demandal25
• Unify pytests for batch prefill for HIP and add to pyproject.toml to run on CI (#90) @demandal25
• Port over BatchPrefillWithPagedKVCacheDevice kernel to HIP (#63) @rtmadduri
• Batch prefill example script (#58) @demandal25
• Port over BatchPrefillWithRaggedKVCache to HIP (#50) @rtmadduri
• Add test_batch_prefill.cpp to HIP (#43) @rtmadduri
• PyTest script for single prefill (#48) @demandal25
• Add tests for LSE in single prefill example script (#46) @demandal25
• Enable FP8 support for Flashinfer ROCm decode kernels on CDNA3 (#40) @rtmadduri
• Single prefill example script (#36) @demandal25
• Support only gfx942 arch for ROCm (#45) @demandal25
• Enable JIT Installation for Prefill (#22) @rtmadduri
• Enable AOT Installation for Prefill (#21) @rtmadduri
• Adds a HIPified version of the SinglePrefillWithKVCacheDevice kernel (#31) @diptorupd
• Copy CUDA versions of prefill into attention/generic (#30) @diptorupd
• Port changes to vec_dtypes from prefill branch to amd-integration (#9) @diptorupd
• Add bit-width specific masks (#7) @diptorupd
• Feature/layout transformation function to transform from A matrix layout to B matrix layout for MFMA (#5) @diptorupd

Changed

• Updated changelog @diptorupd
• Update REAMDE Wheel installation section (#112) @diptorupd
• Change pyproject.toml to generate revised whl name (#113) @rtmadduri
• Update README with clearer usage guide (#72) @diptorupd
• Change Swizzle mode from kLinear to k128B for prefill (#103) @demandal25
• Update logging for using JIT in absence of AOT (#94) @demandal25
• Improvements to Python packaging infrastructure (#76) @diptorupd
• pytest Configuration Improvements for ROCm/HIP Testing (#65) @diptorupd
• Refactor single prefill tests (#51) @demandal25
• Update CHANGELOG for v0.2.5+rocm.0.2 release (#34) @rtmadduri
• Decode feature chunking logic and shared mem optimization (#25) @rtmadduri
• Update hyperlinks in the TOC of README (#33) @demandal25
• Update build instructions for C++ tests in README.md (#27) @demandal25
• Incoporate gpu_iface changes needed for prefill (#14) @diptorupd
• Update CMake infra to run HIP CXX tests using top-level cmake (#10) @diptorupd
• Updates to permuted_smem.cuh (#8) @diptorupd

Removed

• Removes leftover src and all tvm bindings (#99) @diptorupd
• Remove verbose CMake installation messages for editable JIT (#97) @demandal25
• Chore: Refactors the codebase to remove libflashinfer (#88) @diptorupd
• Remove xfail markers about HIP support from pytests (#92) @demandal25
• Chore: Reduce tech debt by removing CUDA sections from generic/prefill.cuh (#87) @diptorupd
• Removes the test_transpose_4x4_half_registers (#11) @diptorupd

Fixed

• Add custom ROCm version scheme to fix wheels version naming (#110) @diptorupd
• Fix datatypes for HIP when using customized attention kernels (#111) @demandal25
• Fix partition-kv=True case and memory allocation issues in batch prefill (#89) @demandal25
• Fixes the single prefill kernel dispatch for HEAD_DIM_QK values gt. 64 (#86) @diptorupd
• Fix/threadblock sync mdo (#62) @diptorupd
• Fix batch prefill example script for ragged kv cache (#73) @demandal25
• Fixes to the single prefill dispatch for HIP devices (#64) @diptorupd
• Skip failing C++ tests and fix mma_debug_utils (#59) @diptorupd
• Fix Log-sum-exp (LSE) write back for single prefill kernels for CDNA3 (#42) @diptorupd
• Implemented fix for the write_o_reg_gmem kernel (#39) @diptorupd
• Fix few more leftover SPDX headers (#38) @diptorupd
• Fix SPDX headers for AMD authored files (#37) @diptorupd
• Improvements to the S-matrix (s_frag) materialization to LDS for debugging (#20) @diptorupd
• fix the pipe at the end of a table (#19) @demandal25
• Fix some compiler warnings in Cxx unit tests (#13) @diptorupd
• Adds debug utility functions for CDNA3 MMA ops. (#3) @diptorupd
• Fixes fragment loading to properly pack 16b values into a 32b register (#2) @diptorupd

Maintenance

• Update rocm+torch versions in Dev Dockerfile (#118) @demandal25
• Update Dockerfile (#116) @rtmadduri
• Update Dockerfile.rocm_ci (#78) @diptorupd
• Fix Fockerfile.rocm_ci (#77) @diptorupd
• Removes hardcoded mamba env name from Docker image (#71) @diptorupd
• Fix/dockerfile rocm.ci (#70) @diptorupd
• Change base image to rocm/dev-ubuntu (#69) @diptorupd
• Revert tar changes from Dockerfile (#68) @rtmadduri
• Update ROCm CI Dockerfile (#67) @rtmadduri
• Update devcontainer dockerfile to rocm7.0.2 (#24) @demandal25
• Enhance devcontainer Dockerfile for ROCm (#28) @demandal25
• Update/devcontainer dockerfile (#23) @diptorupd
• Update Dockerfile.rocm_ci (#6) @clint

Miscellaneous

• Mark copied git repo as a safe directory (#114) @diptorupd
• Disable test_custom_allreduce (#108) @rtmadduri
• Ignore nfs related files in git tracking (#75) @demandal25
• Update .clangd to support HIP specific flags (#74) @demandal25
• Add clangd configuration (#32) @demandal25
• Lint README.md with suggestions from markdown linter (#17) @demandal25

---

Contributors: @diptorupd , @rtmadduri, @demandal25, @clint

Summary: This release brings full prefill kernel support to ROCm, including single and batch prefill with paged and ragged KV cache. Major performance improvements include k128B swizzle mode and FP8 support. Significant infrastructure improvements include complete pytest coverage, improved build system, and updated to ROCm 7.1.1 / PyTorch 2.8.0.