RDNA4 fat wheels — gfx1200 + gfx1201 (vLLM 0.22.0, aiter 0.1.14, flash_attn 2.8.4)

Fat gfx1200;gfx1201 wheels: one set runs on BOTH RDNA4 dies — gfx1200 (RX 9060 XT / 9060, Navi44) and gfx1201 (RX 9070 XT / 9070, Navi48). py3.12 + torch 2.10+rocm7.14. Verified to contain both arches' code objects (flash_attn 2662+2662; vllm _C/_moe_C/_rocm_C all equal-count; aiter JITs per-arch at runtime). gfx1200 is built-correct but NOT yet hardware-validated — see issue #2. The gfx1201-only release v0.22.0-gfx1201 remains as a fallback.

gfx1201 wheels (vLLM 0.22.0 + aiter 0.1.14 + flash_attn 2.8.4)

Prebuilt RDNA4 (gfx1201) wheels for py3.12 + torch 2.10+rocm7.14, consumed by the default Dockerfile in this repo. Built on bare metal from the patched aiter/flash-attention/vllm trees (see Dockerfile.fromsource for commits). ABI-pinned: will not load on a different arch/python/torch.