Skip to content

Rename parameter "softmax_sink" to "sinks" for flash_attn_with_kvcache kernel #30

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

Merged
merged 1 commit into from
Nov 4, 2025
Merged

Rename parameter "softmax_sink" to "sinks" for flash_attn_with_kvcache kernel #30

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
18 changes: 8 additions & 10 deletions benchmark/bench_flash_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,7 @@ def flash_attn_baseline(
causal,
window_size,
softmax_scale,
softmax_sink,
sinks,
cache_seqlens,
page_table,
cu_seqlens_q,
Expand All @@ -24,7 +24,7 @@ def flash_attn_baseline(
k_cache,
v_cache,
causal=causal,
softmax_sink=softmax_sink,
sinks=sinks,
window_size=window_size,
softmax_scale=softmax_scale,
page_table=page_table,
Expand All @@ -39,7 +39,7 @@ def flash_attn_baseline(
# Benchmark configurations
causal = [True, False]
local = [True, False]
use_softmax_sink = [True, False]
use_sinks = [True, False]
batch_size = [1, 16]
q_seq_length_range = [1, 512, 1024]
kv_seq_length_range = [512, 1024, 2048, 4096, 8192, 16384]
Expand All @@ -50,7 +50,7 @@ def flash_attn_baseline(
product(
causal,
local,
use_softmax_sink,
use_sinks,
batch_size,
q_seq_length_range,
kv_seq_length_range,
Expand All @@ -65,7 +65,7 @@ def flash_attn_baseline(
x_names=[
"causal",
"local",
"use_softmax_sink",
"use_sinks",
"batch_size",
"q_seq_length",
"kv_seq_length",
Expand All @@ -84,7 +84,7 @@ def flash_attn_baseline(
def benchmark(
causal,
local,
use_softmax_sink,
use_sinks,
batch_size,
q_seq_length,
kv_seq_length,
Expand Down Expand Up @@ -127,9 +127,7 @@ def benchmark(
max_seqlen_q = q_seq_length
window_size = (-1, -1) if not local else torch.randint(0, kv_seq_length, (2,))

softmax_sink = (
torch.randn(num_heads, device=device, dtype=dtype) if use_softmax_sink else None
)
sinks = torch.randn(num_heads, device=device, dtype=dtype) if use_sinks else None

softmax_scale = 1.0 / (head_dim**0.5)

Expand All @@ -144,7 +142,7 @@ def benchmark(
causal=causal,
window_size=window_size,
softmax_scale=softmax_scale,
softmax_sink=softmax_sink,
sinks=sinks,
cache_seqlens=cache_seqlens,
page_table=page_table,
cu_seqlens_q=cu_seqlens_q,
Expand Down
2 changes: 1 addition & 1 deletion include/sgl_flash_kernel_ops.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,7 +62,7 @@ std::vector<at::Tensor> mha_fwd(
std::optional<at::Tensor>& k_descale_, // (b, h_k)
std::optional<at::Tensor>& v_descale_, // (b, h_k)
float const softmax_scale,
std::optional<const at::Tensor>& softmax_sink,
std::optional<const at::Tensor>& sinks,
bool is_causal,
int window_size_left,
int window_size_right,
Expand Down
4 changes: 2 additions & 2 deletions python/sgl_kernel/flash_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,7 +49,7 @@ def flash_attn_with_kvcache(
k_descale: Optional[torch.Tensor] = None,
v_descale: Optional[torch.Tensor] = None,
softmax_scale=None,
softmax_sink=None,
sinks=None,
causal=False,
window_size=(-1, -1), # -1 means infinite context window
softcap=0.0, # 0.0 means deactivated
Expand Down Expand Up @@ -205,7 +205,7 @@ def flash_attn_with_kvcache(
k_descale,
v_descale,
softmax_scale,
softmax_sink,
sinks,
causal,
window_size[0],
window_size[1],
Expand Down
6 changes: 3 additions & 3 deletions src/sycl/chunked_prefill.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -489,7 +489,7 @@ std::vector<at::Tensor> mha_fwd(
std::optional<at::Tensor>& k_descale_, // (b, h_k)
std::optional<at::Tensor>& v_descale_, // (b, h_k)
const float softmax_scale_,
std::optional<const at::Tensor>& softmax_sink_,
std::optional<const at::Tensor>& sinks_,
bool is_causal,
int window_size_left,
int window_size_right,
Expand Down Expand Up @@ -643,8 +643,8 @@ std::vector<at::Tensor> mha_fwd(

// Set the different scale values.
params.scale_softmax = softmax_scale;
bool use_sink = softmax_sink_.has_value();
params.sink_softmax = use_sink ? softmax_sink_.value().data_ptr() : nullptr;
bool use_sink = sinks_.has_value();
params.sink_softmax = use_sink ? sinks_.value().data_ptr() : nullptr;

params.softcap = softcap;

Expand Down
2 changes: 1 addition & 1 deletion src/torch_extension_sycl.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -82,7 +82,7 @@ TORCH_LIBRARY_FRAGMENT(sgl_kernel, m) {
" Tensor? k_descale,"
" Tensor? v_descale,"
" float softmax_scale,"
" Tensor? softmax_sink,"
" Tensor? sinks,"
" bool is_causal,"
" int window_size_left,"
" int window_size_right,"
Expand Down
14 changes: 7 additions & 7 deletions tests/test_flash_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -489,7 +489,7 @@ def generate_qkv(
# )
@pytest.mark.parametrize("causal,local", [(False, True), (False, False), (True, False)])
# @pytest.mark.parametrize("causal,local", [(True, False)])
@pytest.mark.parametrize("use_softmax_sink", [True, False])
@pytest.mark.parametrize("use_sinks", [True, False])
# @pytest.mark.parametrize(
# "seqlen_new_eq_seqlen_q", [True, False] if not DISABLE_APPENDKV else [True]
# )
Expand Down Expand Up @@ -556,7 +556,7 @@ def test_flash_attn_kvcache(
seqlen_new_eq_seqlen_q,
causal,
local,
use_softmax_sink,
use_sinks,
new_kv,
mha_type,
dtype,
Expand Down Expand Up @@ -586,8 +586,8 @@ def test_flash_attn_kvcache(
assert nheads % nheads_k == 0
dtype_ref = torch.bfloat16 if dtype == torch.float8_e4m3fn else dtype
dv_vals = [128, d] if d > 128 and d <= 192 else ([256, 512, d] if d <= 64 else [d])
if use_softmax_sink:
softmax_sink = torch.randn(nheads, device=device, dtype=dtype_ref)
if use_sinks:
sinks = torch.randn(nheads, device=device, dtype=dtype_ref)
if dtype == torch.float8_e4m3fn or not is_hopper():
# for fp8 and ampere arch, we not support v head dim != qk head dim
dv_vals = [d]
Expand Down Expand Up @@ -831,7 +831,7 @@ def test_flash_attn_kvcache(
k_cache_rep,
v_cache_rep,
softmax_scale,
softmax_sink if use_softmax_sink else None,
sinks if use_sinks else None,
query_padding_mask,
key_padding_mask,
causal=causal,
Expand All @@ -844,7 +844,7 @@ def test_flash_attn_kvcache(
k_cache_rep,
v_cache_rep,
softmax_scale,
softmax_sink if use_softmax_sink else None,
sinks if use_sinks else None,
query_padding_mask,
key_padding_mask,
causal=causal,
Expand Down Expand Up @@ -905,7 +905,7 @@ def test_flash_attn_kvcache(
causal=causal,
window_size=window_size,
softmax_scale=softmax_scale,
softmax_sink=softmax_sink if use_softmax_sink else None,
sinks=sinks if use_sinks else None,
rotary_interleaved=rotary_interleaved,
scheduler_metadata=scheduler_metadata,
num_splits=num_splits,
Expand Down