# Validated environment

## Hardware

- AMD Radeon 8060S / Strix Halo
- GPU target: `gfx1151`, Wave32
- 128 GB unified memory
- Linux TTM/GTT mapping limit: 124 GiB

## Software

- NixOS host
- Python 3.12.12
- Torch `2.13.0+rocm7.15.0a20260728`
- HIP `7.15.0`
- Triton `3.8.0+git4cff872c.rocm7.15.0a20260728`
- AMD clang 23
- torchvision `0.28.0+rocm7.15.0a20260728`
- torchaudio `2.11.0.2+rocm7.15.0a20260728`
- FlashAttention `2.8.3`, Dao-AILab commit `0e60e39473e8df549a20fb5353760f7a65b30e2d`, Python-only AMD Triton path
- vLLM upstream base `d35eb6c44071ea806018841c490f0d2f3219c485`
- CIRU runtime head `838616875c5dd4913d9f753221bf05f64cb7a7ed`

## Build facts

- Fresh Python 3.12 environment.
- Exact AMD multi-architecture wheel transaction installed in one resolver operation.
- `rocm-sdk init` and all 27 SDK tests passed.
- vLLM compiled for `gfx1151` only with 16 jobs.
- The clean package derived the NixOS GCC/glibc bridge without fixed store
  hashes and supplied ROCr host dependencies through an ephemeral `nix-shell`.
- Clean native build completed in 92 seconds on the validation host.
- Imports passed for `vllm._C`, `vllm._C_stable_libtorch`, `vllm._moe_C_stable_libtorch`, `vllm._rocm_C`, and `vllm.cumem_allocator`.

The Ubuntu/Debian, Fedora, Arch, and Windows/WSL build paths in this package are portability candidates and have not yet been clean-host validated by CIRU.

The published native-256K launcher defaults to `gpu_memory_utilization=0.72`, `max_model_len=262144`, `max_num_seqs=6`, and `max_num_batched_tokens=8192`. The validated CIRU profile exposed a 1,457,313-token KV pool and reported 5.56x maximum concurrency at the full native context. Six active short/aggregate-context requests are supported, but six requests cannot all hold the complete 256K limit simultaneously. Users who do not need that KV-cache capacity can lower the memory fraction to leave more unified memory free.

The release head includes a ROCm MLA adapter fix that normalizes upstream FlashAttention LSE output to vLLM's head-first accumulator contract. Focused attention coverage passed 31 tests, with one unrelated CUDA-only FA4 test deliberately deselected. The exact heterogeneous five-request reproducer passed 5/5 after the fix. A later three-repetition exact-2K/256-output C1-C6 sweep passed all 18 measured runs with no errors, preemptions, post-warmup JIT, engine restart, or speed regression; the production profile also passed a separate C6 smoke 6/6.

The retained 60K+512 benchmark was measured earlier at `gpu_memory_utilization=0.75`; that historical measurement completed without faults but left only about 2 GiB of available system memory at peak. Its result is not relabeled as a 0.72 measurement.