Skip to content
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

[Kernel] Triton Paged Attn Decode Kernel #11033

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

[Kernel] Triton Paged Attn Decode Kernel #11033

Show file tree
Hide file tree
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
94 changes: 70 additions & 24 deletions tests/kernels/test_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,12 @@

from tests.kernels.utils import opcheck
from vllm import _custom_ops as ops
from vllm.attention.ops.triton_paged_attn_decode import (
_SEQ_PARTITION_SIZE as TRITON_PAGED_ATTN_DECODE_PARTITION_SIZE)
from vllm.attention.ops.triton_paged_attn_decode import (
paged_attn_decode_v1 as triton_paged_attn_decode_v1)
from vllm.attention.ops.triton_paged_attn_decode import (
paged_attn_decode_v2 as triton_paged_attn_decode_v2)
from vllm.platforms import current_platform
from vllm.utils import get_max_shared_memory_bytes

Expand Down Expand Up @@ -117,7 +123,8 @@ def ref_single_query_cached_kv_attention(

@pytest.mark.parametrize(
"version",
["v1", "v2"] if not current_platform.is_rocm() else ["v1", "v2", "rocm"])
["v1", "v2", "triton_v1", "triton_v2"] if not current_platform.is_rocm()
else ["v1", "v2", "rocm", "triton_v1", "triton_v2"])
@pytest.mark.parametrize("num_seqs", NUM_GEN_SEQS)
@pytest.mark.parametrize("num_heads", NUM_HEADS)
@pytest.mark.parametrize("head_size", HEAD_SIZES)
Expand Down Expand Up @@ -146,6 +153,7 @@ def test_paged_attention(

current_platform.seed_everything(seed)
torch.set_default_device(device)
torch.cuda.set_device(device)
scale = float(1.0 / (head_size**0.5))
num_query_heads, num_kv_heads = num_heads
query = torch.empty(num_seqs, num_query_heads, head_size, dtype=dtype)
Expand All @@ -157,7 +165,13 @@ def test_paged_attention(
if use_alibi:
alibi_slopes = torch.randn(num_query_heads, dtype=torch.float)

seq_lens = [random.randint(1, MAX_SEQ_LEN) for _ in range(num_seqs)]
if version == "triton_v2":
seq_lens = [
random.randint(TRITON_PAGED_ATTN_DECODE_PARTITION_SIZE + 1,
MAX_SEQ_LEN) for _ in range(num_seqs)
]
else:
seq_lens = [random.randint(1, MAX_SEQ_LEN) for _ in range(num_seqs)]
seq_lens[-1] = MAX_SEQ_LEN
max_seq_len = max(seq_lens)
seq_lens = torch.tensor(seq_lens, dtype=torch.int)
Expand All @@ -180,38 +194,58 @@ def test_paged_attention(
kv_cache_dtype, dtype, seed,
device)
key_cache, value_cache = key_caches[0], value_caches[0]

# Using default kv_scale
k_scale = v_scale = 1.0

# Call the paged attention kernel.
output = torch.empty_like(query)
if version == "v1":
ops.paged_attention_v1(
output,
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,
)
if version in ("v1", "triton_v1"):
if version == "v1":
ops.paged_attention_v1(
output,
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,
)

opcheck(torch.ops._C.paged_attention_v1,
opcheck(
torch.ops._C.paged_attention_v1,
(output, 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, 0, 0, 0, 64, 0),
kv_cache_dtype, k_scale, v_scale, 0, 0, 0, 64, 0, 1024),
cond=(head_size == HEAD_SIZES[0]
and block_size == BLOCK_SIZES[0]))
else:
key_cache_tri = key_cache.permute(0, 1, 3, 2,
4).flatten(3, 4).contiguous()
value_cache_tri = value_cache.permute(0, 1, 3, 2).contiguous()
triton_paged_attn_decode_v1(
output,
query,
key_cache_tri,
value_cache_tri,
block_tables,
seq_lens,
max_seq_len,
kv_cache_dtype,
num_kv_heads,
scale,
alibi_slopes,
k_scale,
v_scale,
)

elif version in ("v2", "rocm"):
elif version in ("v2", "rocm", "triton_v2"):
num_partitions = ((max_seq_len + PARTITION_SIZE - 1) // PARTITION_SIZE)
assert PARTITION_SIZE % block_size == 0
num_seqs, num_heads, head_size = output.shape
Expand Down Expand Up @@ -253,6 +287,19 @@ def test_paged_attention(
cond=(head_size == HEAD_SIZES[0]
and block_size == BLOCK_SIZES[0]))

elif version == "triton_v2":
key_cache_tri = key_cache.permute(0, 1, 3, 2,
4).flatten(3, 4).contiguous()
value_cache_tri = value_cache.permute(0, 1, 3, 2).contiguous()
num_partitions = (
(max_seq_len + TRITON_PAGED_ATTN_DECODE_PARTITION_SIZE - 1) //
TRITON_PAGED_ATTN_DECODE_PARTITION_SIZE)
assert TRITON_PAGED_ATTN_DECODE_PARTITION_SIZE % block_size == 0
triton_paged_attn_decode_v2(output, query, key_cache_tri,
value_cache_tri, block_tables,
seq_lens, max_seq_len, kv_cache_dtype,
num_kv_heads, scale, alibi_slopes,
k_scale, v_scale, num_partitions)
else:
ops.paged_attention_rocm(
output,
Expand Down Expand Up @@ -281,7 +328,6 @@ def test_paged_attention(
kv_cache_dtype, k_scale, v_scale),
cond=(head_size == HEAD_SIZES[0]
and block_size == BLOCK_SIZES[0]))

else:
raise AssertionError(f"Unknown version: {version}")

Expand Down
136 changes: 78 additions & 58 deletions vllm/attention/ops/paged_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,11 +3,14 @@

import torch

import vllm.envs as envs
from vllm import _custom_ops as ops
from vllm.triton_utils import HAS_TRITON

if HAS_TRITON:
from vllm.attention.ops.prefix_prefill import context_attention_fwd
from vllm.attention.ops.triton_paged_attn_decode import (
paged_attn_decode_v1, paged_attn_decode_v2)

# Should be the same as PARTITION_SIZE in `paged_attention_v2_launcher`.
_PARTITION_SIZE = 512
Expand Down Expand Up @@ -125,68 +128,85 @@ def forward_decode(
# For context len > 8192, use V2 kernel to avoid shared memory shortage.
use_v1 = (max_seq_len <= 8192
and (max_num_partitions == 1 or num_seqs * num_heads > 512))

use_triton_pa_decode = envs.VLLM_USE_TRITON_PAGED_ATTN_DECODE
if use_v1:
# Run PagedAttention V1.
ops.paged_attention_v1(
output,
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,
tp_rank,
blocksparse_local_blocks,
blocksparse_vert_stride,
blocksparse_block_size,
blocksparse_head_sliding_step,
)
if use_triton_pa_decode == 1:
paged_attn_decode_v1(
output, query, key_cache, value_cache, block_tables,
seq_lens, max_seq_len, kv_cache_dtype, num_kv_heads, scale,
alibi_slopes, k_scale, v_scale, tp_rank,
blocksparse_local_blocks, blocksparse_vert_stride,
blocksparse_block_size, blocksparse_head_sliding_step)
else:
# Run PagedAttention V1.
ops.paged_attention_v1(
output,
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,
tp_rank,
blocksparse_local_blocks,
blocksparse_vert_stride,
blocksparse_block_size,
blocksparse_head_sliding_step,
)
else:
# Run PagedAttention V2.
assert _PARTITION_SIZE % block_size == 0
tmp_output = torch.empty(
size=(num_seqs, num_heads, max_num_partitions, head_size),
dtype=output.dtype,
device=output.device,
)
exp_sums = torch.empty(
size=(num_seqs, num_heads, max_num_partitions),
dtype=torch.float32,
device=output.device,
)
max_logits = torch.empty_like(exp_sums)
ops.paged_attention_v2(
output,
exp_sums,
max_logits,
tmp_output,
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,
tp_rank,
blocksparse_local_blocks,
blocksparse_vert_stride,
blocksparse_block_size,
blocksparse_head_sliding_step,
)

if use_triton_pa_decode == 1:
paged_attn_decode_v2(
output, query, key_cache, value_cache, block_tables,
seq_lens, max_seq_len, kv_cache_dtype, num_kv_heads, scale,
alibi_slopes, k_scale, v_scale, max_num_partitions,
tp_rank, blocksparse_local_blocks, blocksparse_vert_stride,
blocksparse_block_size, blocksparse_head_sliding_step)
else:
tmp_output = torch.empty(
size=(num_seqs, num_heads, max_num_partitions, head_size),
dtype=output.dtype,
device=output.device,
)
exp_sums = torch.empty(
size=(num_seqs, num_heads, max_num_partitions),
dtype=torch.float32,
device=output.device,
)
max_logits = torch.empty_like(exp_sums)
ops.paged_attention_v2(
output,
exp_sums,
max_logits,
tmp_output,
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,
tp_rank,
blocksparse_local_blocks,
blocksparse_vert_stride,
blocksparse_block_size,
blocksparse_head_sliding_step,
)
return output

@staticmethod
Expand Down
Loading
Loading