feat(rocm): f32-query decode GQA kernel with bf16 and fp8 KV cache support - #2168

Merged

mudler merged 10 commits into

mudler:mainfrom

ghazni101:row/fp8-kv-decode-attn

Sep 2, 2026

Commits on Sep 1, 2026

spec(rocm): fp8 KV cache decode attention arm (GFX1100-TG200, fork issue #7)
ghazni101
committed
7dfb08b
feat(rocm): f32-query decode GQA kernel with bf16 and fp8 KV cache support
ghazni101
committed
5b00016
feat(rocm): advertise fp8 and fp8_e4m3 on RocmAttentionBackend
ghazni101
committed
37cd7ba
docs(ENVIRONMENT): document VT_ATTN_DECODE_GQA4, the f32-query GQA4 gate
ghazni101
committed
4a91a14
fix(rocm): LoadRowEplFp8 EPL=16 loaded two uint4 into a 16-element buffer
ghazni101
committed
a28ef2a
feat(abi): expose kv_cache_dtype through the C ABI and vllm-cli
ghazni101
committed
724b6bd
test(rocm): gate the fp8 KV decode attention surface end to end
ghazni101
committed
d195b4c
docs: update public C-ABI/FP8 capability documentation for ABI v24
ghazni101
committed
e35f5ae
spec(rocm): reconcile owning row and record review-driven updates
ghazni101
committed
4968243
record(GFX1100-TG200): advance KERNEL-ATTN-MLA-SPARSE citation to backend.h:587
ghazni101
committed
e93c0a9