[Model][Hardware][AMD][Kernel] Part 2/2 -> Enable e2e QK Norm + RoPE + KV Cache runtime fusion for Qwen3-30B-A3B on ROCM_ATTN (#42755)

Signed-off-by: Jack Hu <Jack.Hu@amd.com>
Signed-off-by: Jack Hu <jingxujaster@hotmail.com>
Signed-off-by: Jack Hu <jack.hu@amd.com>
Co-authored-by: Douglas Lehr <91553416+dllehr-amd@users.noreply.github.com>
This commit is contained in:
Jack Hu
2026-10-15 17:10:14 -07:00
committed by GitHub
co-authored by Douglas Lehr
parent af28c171f5
commit ef38e8d006
23 changed files with 701 additions and 205 deletions
+4
View File
@@ -122,6 +122,7 @@ def paged_attention_rocm(
v_scale: torch.Tensor,
fp8_out_scale: torch.Tensor | None = None,
mfma_type: str = "fp8" if envs.VLLM_ROCM_FP8_MFMA_PAGE_ATTN else "f16",
use_interleaved_v_cache: bool = False,
) -> None:
torch.ops._rocm_C.paged_attention(
out,
@@ -144,6 +145,7 @@ def paged_attention_rocm(
v_scale,
fp8_out_scale,
mfma_type,
use_interleaved_v_cache,
)
@@ -2554,6 +2556,7 @@ def reshape_and_cache(
kv_cache_dtype: str,
k_scale: torch.Tensor,
v_scale: torch.Tensor,
interleaved_v: bool = False,
) -> None:
torch.ops._C_cache_ops.reshape_and_cache(
key,
@@ -2564,6 +2567,7 @@ def reshape_and_cache(
kv_cache_dtype,
k_scale,
v_scale,
interleaved_v,
)