Skip to content

Strix Halo Low-bit Kernel Pack v0.1.0

Latest

Choose a tag to compare

@hstolte11-collab hstolte11-collab released this 03 Sep 12:41

Strix Halo Low-bit Kernel Pack v0.1.0

Initial source-bound release for AMD Radeon 8060S (gfx1151, RDNA 3.5) on ROCm 10.

Release assets

Asset SHA-256
strix_halo_lowbit_kernel_pack-0.1.0-py3-none-any.whl e06d773292cb05c45fc61fffb251fcace49ef387c284c00208fb1827248ed8ad
vllm-0.27.0+haloloom.full.gfx1151.rocm10.rocm100-cp314-cp314-linux_x86_64.whl e5f65076a2a5144e36cd78f89e5e61133fa13660574b014424a7dffdd890b831
sglang-0.5.15-py3-none-any.whl ea578533a14f8912643e25357e7b225026a6159d4ea379955e55ebd30daa8763
sglang_kernel-0.4.4-cp39-abi3-linux_x86_64.whl b11966dbebaa08840fc5f9b3cd0bdc7708d2d6c29c3ab8506a4e10c92fa1fe91

release/SHA256SUMS is authoritative for downloaded assets.

Source reconstruction

  • Six-route HIP source: kernel/rdna35_lowbit.hip
  • Qualified HSACO: kernel/rdna35_lowbit.hsaco
  • vLLM full composition: apply adapters/vllm-v16-full.patch directly to public upstream commit 4bdc8a788d2e2ce9165d552b3d4d8b72604626bf
  • SGLang adapter: apply adapters/sglang.patch to f63458b5beaceabbd9d749b9fc956370e1b649e6
  • Recursive source identities: provenance/

The vLLM public-base patch reconstructs the exact 6,383-file public source root fd7d2374b2f51c65cfe2bfc8a6ce02a0f04f82f1a69f55db2b6e804c0a8f097f. The only post-build source change is test-fixture sanitization in tests/quantization/test_gfx1151_archive_port_contract.py; tests are absent from the wheel.

Physical qualification

  • Direct six-route native gfx1151: qualification/direct-gfx1151-final-v0.1.0.json
  • Final vLLM W4A8 request route: qualification/vllm-full-w4a8-final-v0.1.0.json
  • Final SGLang W4A8 request route: qualification/sglang-full-w4a8-final-v0.1.0.json

Both framework receipts return exact RDNA35_OK, cover 16 eligible layers, use the qualified module hash, complete asynchronously through hipEventRecord+hipEventSynchronize, and report zero fallback or foreign request rows.

Limits

  • Reference correctness/dispatch routes, not claimed tuned GEMMs or performance wins.
  • No broad model-quality claim.
  • Archived W4A4 subpack remains disabled by default. Its retained counts are a historical summary; the underlying matrix/evidence index is not bundled, so this release makes no current archive functional, quality or performance qualification claim.
  • No production activation or promotion authority.
  • Requires the matching HaloLoom ROCm 10/Torch/framework dependency closure; do not install framework wheels into an empty environment with --no-deps.