Release 012026

[diptorupd]

A running checklist of TODOs to prepare for the next release during 01/2026 time frame.

• Complete port of the CUDA versions of prefill.cuh to ROCm 7.1
  • Complete BatchPrefillWithPagedKVCache port
• Refactor the HIP port
  • Rename libflashinfer/include/flashinfer/attention/generic to libflashinfer/include/flashinfer/attention/rocm
  • Remove all CUDA code blocks from HIP implementation
  • Rename source files to *.hip
  • Create a rocm directory under csrc
  • Create a rocm directory under tests
• Rename Wheels package to amd-flashinfer #53
• Complete MMA layout abstraction to clarify indexing logic
• Upgrade rocm-flashinfer to latest upstream trunk. #55
  • Revert directory layout changes
  • Sync JIT and AOT code generation to upstream
  • Remove all sources removed from upstream: e.g. C++ tests, benchmarks, tvm etc.
• Create a feature matrix listing the supported features from flashinfer-ai/flashinfer that have been enabled for rocm-flashinfer #56
• Update Changelog
• Update Readme
• docker/Dockerfile.rocm_ci uses micromamba redundantly #26
• Improve Docker image of rocm flashinfer published at https://hub.docker.com/r/rocm/flashinfer #80
• Support manylinux wheel builds for amd-flashinfer #81

[diptorupd]

Release was completed with the most relevant infra work completed.