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][Core][WIP] Tree attention and parallel decoding #4325

Open
wants to merge 8 commits into
base: main
Choose a base branch
from
Open

[Kernel][Core][WIP] Tree attention and parallel decoding #4325

Changes from 1 commit
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
bdc863d
merge different seqs in seqs group in to once attention inference wit…
Apr 16, 2024
dadbed1
tree width = 2 have been tested, but there is error when width >2
Apr 19, 2024
7596ad8
temp
Apr 21, 2024
1534c5c
tested
Apr 23, 2024
a3849d1
fix early stop
Apr 24, 2024
af98a27
fix code style
Apr 24, 2024
92ebde5
fix bug
Apr 24, 2024
ce76cc7
add duration check
May 7, 2024
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
Prev Previous commit
add duration check
  • Loading branch information
kavioyu committed May 7, 2024
commit ce76cc7672feda5c53df3d56f4d53c08afba2d0f
58 changes: 34 additions & 24 deletions tests/kernels/test_tree_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -136,16 +136,16 @@ def ref_query_cached_kv_attention(output: torch.Tensor, query: torch.Tensor,
output.reshape(-1, num_kv_heads, head_size)


@pytest.mark.parametrize("num_seqs", NUM_GEN_SEQS)
@pytest.mark.parametrize("num_heads", NUM_HEADS)
@pytest.mark.parametrize("head_size", HEAD_SIZES)
@pytest.mark.parametrize("use_alibi", USE_ALIBI)
@pytest.mark.parametrize("block_size", BLOCK_SIZES)
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("kv_cache_dtype", KV_CACHE_DTYPE)
@pytest.mark.parametrize("seed", SEEDS)
@pytest.mark.parametrize("device", CUDA_DEVICES)
@pytest.mark.parametrize("tree_width", TREEWIDTH)
# @pytest.mark.parametrize("num_seqs", NUM_GEN_SEQS)
# @pytest.mark.parametrize("num_heads", NUM_HEADS)
# @pytest.mark.parametrize("head_size", HEAD_SIZES)
# @pytest.mark.parametrize("use_alibi", USE_ALIBI)
# @pytest.mark.parametrize("block_size", BLOCK_SIZES)
# @pytest.mark.parametrize("dtype", DTYPES)
# @pytest.mark.parametrize("kv_cache_dtype", KV_CACHE_DTYPE)
# @pytest.mark.parametrize("seed", SEEDS)
# @pytest.mark.parametrize("device", CUDA_DEVICES)
# @pytest.mark.parametrize("tree_width", TREEWIDTH)
def test_paged_attention(
kv_cache_factory,
num_seqs: int,
Expand Down Expand Up @@ -206,22 +206,33 @@ def test_paged_attention(
kv_cache_dtype, dtype, seed,
device)
key_cache, value_cache = key_caches[0], value_caches[0]

# #key_cache = torch.randint_like(key_cache, -1, 2)
# #value_cache = torch.randint_like(value_cache, -1, 2)
# value_cache = torch.arange(value_cache.numel()).reshape(value_cache.shape).to(value_cache.dtype)
# #query = torch.randint_like(query, -1, 2)
# #value_cache = torch.ones_like(value_cache)
# key_cache = torch.ones_like(key_cache)
# query = torch.ones_like(query)


torch.cuda.synchronize()
start_time = time.time()
output = torch.empty_like(query)
ops.paged_attention_v1(
output,
query,
key_cache,
value_cache,
num_kv_heads,
scale,
block_tables,
context_lens,
block_size,
max_context_len,
alibi_slopes,
kv_cache_dtype,
1.0,
)
torch.cuda.synchronize()
paged_attn_duration = (time.time()-start_time)*tree_width

start_time = time.time()
tree_attention_fwd(query, output, key_cache, value_cache, block_tables,
context_lens, prompt_lens, tree_width, alibi_slopes)
torch.cuda.synchronize()
#print("tree attention duration:", time.time()-start_time)
tree_attn_duration = time.time()-start_time

ref_output = torch.empty_like(query)
ref_query_cached_kv_attention(ref_output, query, num_queries_per_kv,
Expand All @@ -235,8 +246,7 @@ def test_paged_attention(
atol = 1e-4
rtol = 2e-2

def diff(a, b):
print(((a - b).abs() / (b + 1e-8)).mean())

diff(output, ref_output)
assert torch.allclose(output, ref_output, atol=atol, rtol=rtol)


test_paged_attention(create_kv_caches_with_random, 78, (32, 32), 128, False, 32, torch.half, "auto", 1, 8, 'cuda')