feat(rocm): add RDNA3 and RDNA4 runtime foundation

[zihaomu]

Summary

• add an independent ROCm source-build foundation for RDNA3 (gfx1100-gfx1103) and RDNA4 (gfx1200, gfx1201);
• support both traditional /opt/rocm layouts and the modular ROCm SDK shipped by the official ROCm 7.14 PyTorch image;
• build and load the pinned-memory and CPU-MoE extensions on HIP;
• make the pinned-host fast-index-copy and Triton activation/norm paths work on ROCm;
• detect the visible gcnArchName, with explicit environment overrides for cross compilation.

Why this targets main directly

This PR is a standalone, current-main integration. It incorporates the useful work from #23 and the RDNA3 runtime follow-up at bouclem#1, then adds the RDNA4 and ROCm 7.14 changes. It therefore does not require either external feature branch to merge first.

This intentionally overlaps those open PRs so maintainers can review and merge a complete, hardware-tested ROCm foundation without being blocked by a cross-fork base chain.

Attribution

• @bouclem authored the initial RDNA3/ROCm foundation in feat: add simple ROCm GPU support for RDNA3 (gfx1100-1103) #23; that author information is preserved in commit 7caa62d (upstream source commit 27c0977).
• @nekomario28 authored the RDNA3 runtime completion in fix(rocm): complete RDNA3 runtime path bouclem/FreeToken#1; author information is preserved across commits af67560 through 62bb962 (upstream follow-up head 4e11ed0).
• @zihaomu added RDNA4 discovery, ROCm 7.14 modular-SDK support, versioned HIP-runtime linking, runtime DLPack handling, tests, and installation documentation.

Thank you to both contributors for establishing and validating the earlier ROCm paths.

Validation

Hardware and software:

• AMD Radeon AI PRO R9700 (gfx1201)
• PyTorch 2.11.0+rocm7.14.0
• HIP 7.14.60850
• AMD Triton 3.7.1

Results:

• clean python -m pip install -e . --no-build-isolation --no-deps: PASS; both native extensions compiled and loaded;
• cold TVM-FFI fast-index-copy JIT build: PASS;
• architecture, pinned/copy, Triton attention, Gemma norm, SwiGLU, and CPU-MoE graph replay tests: 48 passed;
• python -m pip check: PASS;
• git diff --check: PASS.

Compatibility

ROCm-specific code is guarded by HIP/ROCm detection and existing CUDA paths are retained. Physical NVIDIA regression testing is not claimed; upstream CUDA CI is requested before merge.

Follow-ups

Kept out of this foundation PR for separate review:

1. generic TVM-FFI index/store HIP portability;
2. CUDA-only optional-backend gating;
3. RCCL tensor-parallel routing;
4. native GGUF kernels on ROCm/RDNA4;
5. RDNA-specific performance tuning and full model-serving validation.

[zihaomu]

Draft follow-ups are now available:

• fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP #133 — TVM-FFI index/store JIT portability for HIP
• fix(rocm): gate CUDA-only optional backends on ROCm #134 — ROCm gating for CUDA-only optional backends
• fix(rocm): route tensor parallel communication through RCCL #135 — RCCL routing for tensor-parallel communication
• feat(rocm): enable native GGUF kernels on RDNA4 #136 — native GGUF kernels on RDNA4 (also follows fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP #133)

All four are opened as Draft PRs against main. Until this foundation PR merges, their GitHub “Files changed” views temporarily include the #132 foundation commits; each PR body lists its intended incremental review scope. They will remain drafts and be rebased/refreshed as their dependencies land.