Skip to content

Add v2 test to paged_attention_decode #764

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
rahulbatra85 wants to merge 1 commit into
base: main_perf
Choose a base branch
from
Open

Add v2 test to paged_attention_decode #764

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 13 additions & 6 deletions python/perf-kernels/paged_attention_decode.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -382,8 +382,8 @@ def paged_attn_decode_v2(output: torch.Tensor, #[num_seqs, num_kv_heads*query_g
_paged_attn_decode_v2_wo_dot_reduce_kernel[grid](output, exp_sums, max_logits, tmp_output, seq_lens,
output.stride(0), output.stride(1), exp_sums.stride(0),
exp_sums.stride(1), tmp_output.stride(0), tmp_output.stride(1),
tmp_output.stride(2), compute_type=compute_type,
HEAD_SZ=head_sz, HEAD_SZ_POW2=head_sz_pow2,
tmp_output.stride(2), HEAD_SZ=head_sz,
HEAD_SZ_POW2=head_sz_pow2,
SEQ_PARTITION_SZ=_SEQ_PARTITION_SIZE,
MAX_NUM_SEQ_PARTITIONS=int(max_num_partitions),
MAX_NUM_SEQ_PARTITIONS_POW2=int(max_num_partitions_pow2))
Expand Down Expand Up @@ -412,8 +412,8 @@ def paged_attn_decode_v2(output: torch.Tensor, #[num_seqs, num_kv_heads*query_g
_paged_attn_decode_v2_w_dot_reduce_kernel[grid](
output, exp_sums, max_logits, tmp_output, seq_lens, output.stride(0), output.stride(1), exp_sums.stride(0),
exp_sums.stride(1), exp_sums.stride(2), tmp_output.stride(0), tmp_output.stride(1), tmp_output.stride(2),
tmp_output.stride(3), compute_type=compute_type, HEAD_SZ=head_sz, HEAD_SZ_POW2=head_sz_pow2,
QUERY_GRP_SZ=query_grp_sz, QUERY_GRP_SZ_POW2=query_grp_sz_pow2, SEQ_PARTITION_SZ=_SEQ_PARTITION_SIZE,
tmp_output.stride(3), HEAD_SZ=head_sz, HEAD_SZ_POW2=head_sz_pow2, QUERY_GRP_SZ=query_grp_sz,
QUERY_GRP_SZ_POW2=query_grp_sz_pow2, SEQ_PARTITION_SZ=_SEQ_PARTITION_SIZE,
MAX_NUM_SEQ_PARTITIONS=int(max_num_partitions),
MAX_NUM_SEQ_PARTITIONS_POW2=int(triton.next_power_of_2(max_num_partitions)))

Expand Down Expand Up @@ -855,21 +855,26 @@ def paged_attention_decode_ref(output, #[num_seqs, num_q_heads, head_sz]
(8, 1, 1, 1, 1, 10),
(16, 1, 1, 1, 1, 10),
(64, 1, 1, 1, 1, 10),
(64, 1, 1, 1, 1, 8192),
(64, 1, 1, 1, 1, 10000),

#H_Q and H_KV > 1
(1, 4, 4, 1, 1, 1),
(1, 4, 4, 1, 1, 10),
(1, 4, 4, 1, 1, 10000),

#Head_dim > 1
(1, 1, 1, 8, 1, 1),
(1, 1, 1, 8, 1, 10),
(1, 1, 1, 8, 1, 10000),

#H_Q and H_KV > 1 and Head_dim > 1
(1, 4, 4, 8, 1, 1),
(1, 4, 4, 8, 1, 10),
(4, 4, 4, 8, 1, 10),
(16, 4, 4, 8, 1, 10),
(32, 4, 4, 8, 1, 10),
(32, 4, 4, 8, 1, 10000),

#H_Q and H_KV > 1 and Head_dim > 1 and KV_BLK_SZ > 1
(1, 1, 1, 1, 1, 1),
Expand All @@ -891,12 +896,12 @@ def paged_attention_decode_ref(output, #[num_seqs, num_q_heads, head_sz]
#GQA Basic
(1, 2, 1, 16, 16, 1),
(1, 2, 1, 16, 16, 10),

#GQA Basic
(1, 2, 1, 16, 16, 10000),
(1, 4, 2, 16, 16, 1),
(1, 4, 2, 16, 16, 10),
(1, 4, 2, 128, 16, 1),
(1, 4, 2, 128, 16, 10),
(1, 4, 2, 128, 16, 16384),
(1, 6, 2, 128, 16, 1),
(1, 6, 2, 128, 16, 10),
(1, 6, 2, 128, 16, 16),
Expand All @@ -906,9 +911,11 @@ def paged_attention_decode_ref(output, #[num_seqs, num_q_heads, head_sz]
(1, 6, 2, 128, 16, 56),
(1, 6, 2, 128, 16, 64),
(1, 6, 2, 128, 16, 128),
(1, 6, 2, 128, 16, 8192),
(1, 8, 2, 128, 16, 128),
(4, 8, 2, 128, 16, 128),
(16, 8, 2, 128, 16, 128),
(16, 8, 2, 128, 16, 10000),
(32, 8, 2, 128, 16, 128),
(64, 8, 2, 128, 16, 200),
])
Expand Down