[ROCm][Kernel][V1] Enable AMD Radeon GPU Custom Paged Attention on v1 (#17004)
Signed-off-by: Hosang Yoon <hosang.yoon@amd.com>
This commit is contained in:
@@ -861,7 +861,8 @@ class ROCmFlashAttentionImpl(AttentionImpl):
|
||||
gqa_ratio = num_heads // self.num_kv_heads
|
||||
use_custom = use_rocm_custom_paged_attention(
|
||||
decode_query.dtype, head_size, block_size, gqa_ratio,
|
||||
decode_meta.max_decode_seq_len, self.sliding_window)
|
||||
decode_meta.max_decode_seq_len, self.sliding_window,
|
||||
self.kv_cache_dtype, self.alibi_slopes)
|
||||
if use_custom:
|
||||
max_seq_len = (decode_meta.max_decode_seq_len if self.attn_type
|
||||
!= AttentionType.ENCODER_DECODER else
|
||||
|
||||
@@ -283,7 +283,8 @@ def chunked_prefill_paged_decode(
|
||||
use_custom = use_rocm_custom_paged_attention(query.dtype, head_size,
|
||||
block_size,
|
||||
num_queries_per_kv,
|
||||
max_seq_len, sliding_window)
|
||||
max_seq_len, sliding_window,
|
||||
kv_cache_dtype, alibi_slopes)
|
||||
if use_custom:
|
||||
_PARTITION_SIZE_ROCM = 256
|
||||
max_num_partitions = ((max_seq_len + _PARTITION_SIZE_ROCM - 1) //
|
||||
|
||||
Reference in New Issue
Block a user