v0.5.3+amd.1

Updated Upstream

• Updated to upstream v0.5.3 tag of FlashInfer (#173) @diptorupd

Added

• Port new upstream RoPE kernels to ROCm: fused RoPE + FP8 quantization and paged KV cache append (#196) @diptorupd
• Add test_hip_utils (#197) @rtmadduri
• Add AITER version check to FlashInfer and configure AITER during Docker build (#191) @rtmadduri
• Add AITER_ROCM_VERSION (#193) @rtmadduri
• Enable gfx950 support for FlashInfer+ROCm (#188) @rtmadduri
• Add prefill and decode aliases to sys.modules (#185) @diptorupd
• Support environments built by TheRock build system (#172) @eppaneamd
• Run tests in parallel with pytest-xdist and multi-GPU scheduling (#176) @diptorupd
• JIT core and env: IS_HIP block with cpp_ext_hip, JitSpecRegistry, and amd-jit-cache integration (#173) @diptorupd

Changed

• Update decode/prefill_rocm to v0.5.3 API; expand HIP exports (#173) @diptorupd
• Re-HIPify sampling kernels (#173) @diptorupd
• Update and reorganize HIP tests (#173) @diptorupd
• amd-flashinfer-jit-cache: align version with v0.5.3 (#173) @diptorupd
• Fix README post v0.5.3 upgrade (#179) @diptorupd

Fixed

• Fix AITER prefill graph-capture path + update page-size support (#170) @rtmadduri
• Add support for both CUDA and HIP generator headers in sampling.cu (#180) @eppaneamd
• Move HIP_VISIBLE_DEVICES setting to top-level conftests.py (#183) @diptorupd
• Fix HIP import errors blocking canary test on ROCm (pynvml lazy import, fp4 guard, modules_hip generate_additional_params, mypy) (#173) @diptorupd
• Fix flaky Sampling Tests (#189) @rtmadduri
• Reduce sampling threshold from 0.99 to 0.98 (#190) @diptorupd
• Improve test_logits_processor_hip tolerances (#192) @rtmadduri
• Fix arch check (#187) @diptorupd

Maintenance

• Fixes devcontainer Dockerfile (#195) @diptorupd
• Update dev Dockerfile to install AITER (#194) @diptorupd
• Infra: improve dockerfile.rocm ci (#186) @diptorupd
• Fix linter issues identified by pre-commit (#184) @diptorupd
• chore: Update coverage include list (#182) @diptorupd
• Fix the coverage include list workflow (#178) @diptorupd

---

Contributors: @diptorupd, @rtmadduri, @eppaneamd

Summary: This release rebases amd-flashinfer onto the upstream v0.5.3 tag. It adds a full port of the fused RoPE + FP8 quantization and paged KV cache append kernels to ROCm, gfx950 (MI350) support, and AITER version checking with Docker build integration. JIT and env layers now use a dedicated IS_HIP path with cpp_ext_hip and amd-jit-cache. Fixes include AITER prefill graph-capture and page-size handling, HIP/CUDA generator header compatibility in sampling, and ROCm import/canary test fixes. Testing improvements include parallel pytest runs with multi-GPU scheduling, test_hip_utils, and TheRock build system support.

---