[ROCM][KERNEL] Paged attention for V1 (#15720)

Signed-off-by: Aleksandr Malyshev <maleksan@amd.com>
Signed-off-by: root <root@banff-cyxtera-s65-4.amd.com>
Co-authored-by: Aleksandr Malyshev <maleksan@amd.com>
Co-authored-by: root <root@banff-cyxtera-s65-4.amd.com>
This commit is contained in:
Aleksandr Malyshev
2025-04-02 19:48:00 -07:00
committed by GitHub
co-authored by Aleksandr Malyshev root
parent bd7599d34a
commit e73ff24e31
11 changed files with 219 additions and 109 deletions
+4 -2
View File
@@ -110,6 +110,7 @@ def paged_attention_rocm(
scale: float,
block_tables: torch.Tensor,
seq_lens: torch.Tensor,
query_start_loc: Optional[torch.Tensor],
block_size: int,
max_seq_len: int,
alibi_slopes: Optional[torch.Tensor],
@@ -120,8 +121,9 @@ def paged_attention_rocm(
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,
block_size, max_seq_len, alibi_slopes,
kv_cache_dtype, k_scale, v_scale)
query_start_loc, block_size, max_seq_len,
alibi_slopes, kv_cache_dtype, k_scale,
v_scale)
def mla_decode_kvcache_cpu(