Fp8 paged attention update (#22222)

Signed-off-by: Xiao Yu <xiao.yu@amd.com>
Signed-off-by: xiao-llm <xiao.yu.dc@outlook.com>
Co-authored-by: Xiao Yu <xiao.yu@metamaterial.com>
Co-authored-by: Xiao Yu <xiao.yu@amd.com>
Co-authored-by: Bowen Bao <bowenbao@amd.com>
This commit is contained in:
xiao-llm
2025-09-15 10:43:26 -04:00
committed by GitHub
co-authored by Xiao Yu Xiao Yu Bowen Bao
parent 0e219cd50b
commit 01413e0cf5
5 changed files with 257 additions and 79 deletions
+2 -1
View File
@@ -117,13 +117,14 @@ def paged_attention_rocm(
k_scale: torch.Tensor,
v_scale: torch.Tensor,
fp8_out_scale: Optional[torch.Tensor] = None,
mfma_type: str = "fp8" if envs.VLLM_ROCM_FP8_MFMA_PAGE_ATTN else "f16",
) -> None:
torch.ops._rocm_C.paged_attention(out, exp_sum, max_logits, tmp_out, query,
key_cache, value_cache, num_kv_heads,
scale, block_tables, seq_lens,
query_start_loc, block_size, max_seq_len,
alibi_slopes, kv_cache_dtype, k_scale,
v_scale, fp8_out_scale)
v_scale, fp8_out_scale, mfma_type)
def mla_decode_kvcache_cpu(