Skip to content

Changes for fix flash attention KV cache and prefill issues #617

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

Changes for fix flash attention KV cache and prefill issues #617

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
9 changes: 9 additions & 0 deletions test/unit/flash_attention/flash_attention_prefill/flash_prefill_testbed_3x.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -283,6 +283,10 @@ struct TestbedImpl {
block_O.reset(batch * num_heads_q * seq_len_qo * head_size_vo);
block_ref_O.reset(batch * num_heads_q * seq_len_qo * head_size_vo);

// Zero-initialize output memory to prevent uninitialized values from USM reuse
compat::memset(block_O.get(), 0, block_O.size() * sizeof(ElementOutput));
compat::memset(block_ref_O.get(), 0, block_ref_O.size() * sizeof(ElementOutput));

initialize_block(block_Q, seed + 2023);
initialize_block(block_K, seed + 2022);
initialize_block(block_V, seed + 2021);
Expand Down Expand Up @@ -595,6 +599,11 @@ struct TestbedImpl {
CUTLASS_TRACE_HOST("TestbedImpl::run: Allocating workspace of size " << workspace_size);
#endif
cutlass::device_memory::allocation<uint8_t> workspace(workspace_size);

// Zero-initialize workspace to prevent uninitialized memory issues
if (workspace_size > 0) {
compat::memset(workspace.get(), 0, workspace_size);
}

#if (CUTLASS_DEBUG_TRACE_LEVEL > 1)
CUTLASS_TRACE_HOST("TestbedImpl::run: Calling FlashAttention::can_implement");
Expand Down
2 changes: 1 addition & 1 deletion test/unit/flash_attention/flash_attention_prefill/xe_flash_prefill.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,7 +52,7 @@ TEST(TEST_NAME, noncausal) {
EXPECT_TRUE(test::flash_attention::TestFlashPrefillAll<Kernel>(HEAD_DIM));
}

TEST(GTEST_CONCAT_TOKEN_(DISABLED_, TEST_NAME), varlen_causal) {
TEST(TEST_NAME, varlen_causal) {
using Kernel = test::flash_attention::XE_Flash_Attention_Prefill<INPUT_TYPE, float, OUT_TYPE, typename Shape_h::ShapeQK, typename Shape_h::ShapePV,
typename Shape_h::ShapeOutput, typename Shape_h::SubgroupLayout, MMAOperation, true, true, 2>::Kernel;
EXPECT_TRUE(test::flash_attention::TestFlashPrefillAll<Kernel>(HEAD_DIM));
Expand Down
10 changes: 10 additions & 0 deletions ...it/flash_attention/flash_attention_prefill_cachedkv/flash_prefill_cachedkv_testbed_3x.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -220,6 +220,11 @@ struct TestbedImpl {
block_V_cache.reset(batch * num_heads_kv * seq_len_kv_cache * head_size_vo);
block_O.reset(batch * num_heads_q * seq_len_qo * head_size_vo);
block_ref_O.reset(batch * num_heads_q * seq_len_qo * head_size_vo);

// Zero-initialize output buffers to prevent uninitialized memory issues
if (block_O.size() > 0) {
compat::memset(block_O.get(), 0, block_O.size() * sizeof(ElementOutput));
}

if constexpr (UsePagedKV) {
std::vector<int> num_pages_per_seq{0};
Expand Down Expand Up @@ -632,6 +637,11 @@ struct TestbedImpl {
CUTLASS_TRACE_HOST("TestbedImpl::run: Allocating workspace of size " << workspace_size);
#endif
cutlass::device_memory::allocation<uint8_t> workspace(workspace_size);

// Zero-initialize workspace to prevent uninitialized memory issues
if (workspace_size > 0) {
compat::memset(workspace.get(), 0, workspace_size);
}

#if (CUTLASS_DEBUG_TRACE_LEVEL > 1)
CUTLASS_TRACE_HOST("TestbedImpl::run: Calling FlashPrefillCachedKV::can_implement");
Expand Down
2 changes: 1 addition & 1 deletion ...tention/flash_attention_prefill_cachedkv/xe_flash_prefill_cachedkv_bf16_fp32_fp32_128.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,7 +61,7 @@ TEST(XE_Flash_Attention_Prefill_bf16_128, noncausal) {
EXPECT_TRUE(test::flash_attention::TestFlashPrefillCachedKVAll<Kernel>(128));
}

TEST(DISABLED_XE_Flash_Attention_Prefill_bf16_128, varlen_causal) {
TEST(XE_Flash_Attention_Prefill_bf16_128, varlen_causal) {
constexpr int PipelineStages = 2;
using ShapeQK = Shape<_128, _64, _64>;
using ShapePV = Shape<_128, _32, _64>;
Expand Down
2 changes: 1 addition & 1 deletion ...tention/flash_attention_prefill_cachedkv/xe_flash_prefill_cachedkv_bf16_fp32_fp32_192.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,7 +61,7 @@ TEST(XE_Flash_Attention_Prefill_bf16_192, noncausal) {
EXPECT_TRUE(test::flash_attention::TestFlashPrefillCachedKVAll<Kernel>(192));
}

TEST(DISABLED_XE_Flash_Attention_Prefill_bf16_192, varlen_causal) {
TEST(XE_Flash_Attention_Prefill_bf16_192, varlen_causal) {
constexpr int PipelineStages = 2;
using ShapeQK = Shape<_256, _64, _64>;
using ShapePV = Shape<_256, _32, _64>;
Expand Down
2 changes: 1 addition & 1 deletion ...ttention/flash_attention_prefill_cachedkv/xe_flash_prefill_cachedkv_bf16_fp32_fp32_64.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,7 +61,7 @@ TEST(XE_Flash_Attention_Prefill_bf16_64, noncausal) {
EXPECT_TRUE(test::flash_attention::TestFlashPrefillCachedKVAll<Kernel>(64));
}

TEST(DISABLED_XE_Flash_Attention_Prefill_bf16_64, varlen_causal) {
TEST(XE_Flash_Attention_Prefill_bf16_64, varlen_causal) {
constexpr int PipelineStages = 2;
using ShapeQK = Shape<_128, _64, _64>;
using ShapePV = Shape<_128, _32, _64>;
Expand Down
2 changes: 1 addition & 1 deletion ...ttention/flash_attention_prefill_cachedkv/xe_flash_prefill_cachedkv_bf16_fp32_fp32_96.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,7 +61,7 @@ TEST(XE_Flash_Attention_Prefill_bf16_96, noncausal) {
EXPECT_TRUE(test::flash_attention::TestFlashPrefillCachedKVAll<Kernel>(96));
}

TEST(DISABLED_XE_Flash_Attention_Prefill_bf16_96, varlen_causal) {
TEST(XE_Flash_Attention_Prefill_bf16_96, varlen_causal) {
constexpr int PipelineStages = 2;
using ShapeQK = Shape<_128, _64, _32>;
using ShapePV = Shape<_128, _32, _64>;
Expand Down
2 changes: 1 addition & 1 deletion ...tention/flash_attention_prefill_cachedkv/xe_flash_prefill_cachedkv_fp16_fp32_fp32_128.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,7 +61,7 @@ TEST(XE_Flash_Attention_Prefill_fp16_128, noncausal) {
EXPECT_TRUE(test::flash_attention::TestFlashPrefillCachedKVAll<Kernel>(128));
}

TEST(DISABLED_XE_Flash_Attention_Prefill_fp16_128, varlen_causal) {
TEST(XE_Flash_Attention_Prefill_fp16_128, varlen_causal) {
constexpr int PipelineStages = 2;
using ShapeQK = Shape<_128, _64, _64>;
using ShapePV = Shape<_128, _32, _64>;
Expand Down
2 changes: 1 addition & 1 deletion ...tention/flash_attention_prefill_cachedkv/xe_flash_prefill_cachedkv_fp16_fp32_fp32_192.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,7 +61,7 @@ TEST(XE_Flash_Attention_Prefill_fp16_192, noncausal) {
EXPECT_TRUE(test::flash_attention::TestFlashPrefillCachedKVAll<Kernel>(192));
}

TEST(DISABLED_XE_Flash_Attention_Prefill_fp16_192, varlen_causal) {
TEST(XE_Flash_Attention_Prefill_fp16_192, varlen_causal) {
constexpr int PipelineStages = 2;
using ShapeQK = Shape<_256, _64, _64>;
using ShapePV = Shape<_256, _32, _64>;
Expand Down
2 changes: 1 addition & 1 deletion ...ttention/flash_attention_prefill_cachedkv/xe_flash_prefill_cachedkv_fp16_fp32_fp32_64.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,7 +61,7 @@ TEST(XE_Flash_Attention_Prefill_fp16_64, noncausal) {
EXPECT_TRUE(test::flash_attention::TestFlashPrefillCachedKVAll<Kernel>(64));
}

TEST(DISABLED_XE_Flash_Attention_Prefill_fp16_64, varlen_causal) {
TEST(XE_Flash_Attention_Prefill_fp16_64, varlen_causal) {
constexpr int PipelineStages = 2;
using ShapeQK = Shape<_128, _64, _64>;
using ShapePV = Shape<_128, _32, _64>;
Expand Down
2 changes: 1 addition & 1 deletion ...ttention/flash_attention_prefill_cachedkv/xe_flash_prefill_cachedkv_fp16_fp32_fp32_96.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,7 +61,7 @@ TEST(XE_Flash_Attention_Prefill_fp16_96, noncausal) {
EXPECT_TRUE(test::flash_attention::TestFlashPrefillCachedKVAll<Kernel>(96));
}

TEST(DISABLED_XE_Flash_Attention_Prefill_fp16_96, varlen_causal) {
TEST(XE_Flash_Attention_Prefill_fp16_96, varlen_causal) {
constexpr int PipelineStages = 2;
using ShapeQK = Shape<_128, _64, _32>;
using ShapePV = Shape<_128, _32, _64>;
Expand Down