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

[CPU] Enable u8 kv cache by default #27454

Merged
merged 5 commits into from
Nov 13, 2024
Merged

[CPU] Enable u8 kv cache by default #27454

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
0f67bd8
default enable u8 kv cache
luo-cheng2021 Nov 7, 2024
f37b571
disable u8 kvcache default config on non-x86
luo-cheng2021 Nov 8, 2024
81d3411
apply review comment: fix u8 set_state failure
luo-cheng2021 Nov 11, 2024
16578fe
Merge remote-tracking branch 'upstream/master' into luocheng/enable_i…
luo-cheng2021 Nov 11, 2024
45eb699
cover bf16/f32/u8 for get/set_state
luo-cheng2021 Nov 12, 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
4 changes: 4 additions & 0 deletions src/plugins/intel_cpu/src/config.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -358,6 +358,7 @@ void Config::readProperties(const ov::AnyMap& prop, const ModelType modelType) {
}
} else if (key == ov::hint::kv_cache_precision.name()) {
try {
kvCachePrecisionSetExplicitly = true;
auto const prec = val.as<ov::element::Type>();
if (one_of(prec, ov::element::f32, ov::element::f16, ov::element::bf16, ov::element::u8)) {
kvCachePrecision = prec;
Expand Down Expand Up @@ -411,6 +412,9 @@ void Config::readProperties(const ov::AnyMap& prop, const ModelType modelType) {
if (!fcDynamicQuantizationGroupSizeSetExplicitly) {
fcDynamicQuantizationGroupSize = 0;
}
if (!kvCachePrecisionSetExplicitly) {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

why not set kvCachePrecision to u8 here ? To make the kvCachePrecision compatible in ACL platform ? If so, it's better to left a comment here to explain this.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Setting the kvCachePrecision in constructor should be more clearer and consistent with other properties such as DynamicQuantizationGroup, so my suggestion is to keep them in current style.

kvCachePrecision = ov::element::f32;
}
}

if (!prop.empty())
Expand Down
4 changes: 3 additions & 1 deletion src/plugins/intel_cpu/src/config.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -51,14 +51,16 @@ struct Config {
std::string device_id = {};
float fcSparseWeiDecompressionRate = 1.0f;
uint64_t fcDynamicQuantizationGroupSize = 32;
ov::element::Type kvCachePrecision = ov::element::f16;
bool fcDynamicQuantizationGroupSizeSetExplicitly = false;
bool kvCachePrecisionSetExplicitly = false;
#if defined(OV_CPU_WITH_ACL)
bool aclFastMath = false;
#endif
#if defined(OPENVINO_ARCH_X86_64)
ov::element::Type kvCachePrecision = ov::element::u8;
size_t rtCacheCapacity = 5000ul;
#else
ov::element::Type kvCachePrecision = ov::element::f16;
// TODO: Executor cache may leads to incorrect behavior on oneDNN ACL primitives
size_t rtCacheCapacity = 0ul;
#endif
Expand Down
9 changes: 5 additions & 4 deletions src/plugins/intel_cpu/src/memory_state.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -297,18 +297,19 @@ void VariableStateKVcache::set_state_impl(const ov::SoPtr<ov::ITensor>& state) {
auto S = internal.size(3);
auto nthr = parallel_get_max_threads();
std::vector<PlainTensor> buffers(nthr);
m_scale_zp.resize<float>({L0, B, H, 2});
zhangYiIntel marked this conversation as resolved.
Show resolved Hide resolved
parallel_for3d(B, H, L0, [&](size_t ithr, size_t b, size_t h, size_t m) {
buffers[ithr].resize<float>({S});
cpu_convert(external.ptr_v(b, h, m),
cpu_convert(external.ptr_v(m, b, h),
buffers[ithr].ptr<float>(),
external.m_dt,
element::f32,
S);
attn_quant_u8(buffers[ithr].ptr<float>(),
internal.ptr<uint8_t>(b, h, m),
internal.ptr<uint8_t>(m, b, h),
S,
m_scale_zp.at<float>({b, h, m, size_t{0}}),
m_scale_zp.at<float>({b, h, m, size_t{1}}));
m_scale_zp.at<float>({m, b, h, size_t{0}}),
m_scale_zp.at<float>({m, b, h, size_t{1}}));
});
} else {
m_internal_mem->load(external_mem);
Expand Down
11 changes: 11 additions & 0 deletions src/plugins/intel_cpu/tests/functional/custom/behavior/ov_executable_network/properties.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -194,6 +194,17 @@ TEST_F(OVClassConfigTestCPU, smoke_CpuExecNetworkCheckAccuracyModeDynamicQuantiz
ASSERT_EQ(groupSize, 0);
}

TEST_F(OVClassConfigTestCPU, smoke_CpuExecNetworkCheckAccuracyModeKVCachePrecision) {
ov::Core core;

ASSERT_NO_THROW(core.set_property(deviceName, ov::hint::execution_mode(ov::hint::ExecutionMode::ACCURACY)));
ov::CompiledModel compiledModel = core.compile_model(model, deviceName);

auto kv_cache_precision_value = ov::element::undefined;
ASSERT_NO_THROW(kv_cache_precision_value = compiledModel.get_property(ov::hint::kv_cache_precision));
ASSERT_EQ(kv_cache_precision_value, ov::element::f32);
}

const auto bf16_if_can_be_emulated = ov::with_cpu_x86_avx512_core() ? ov::element::bf16 : ov::element::f32;

TEST_F(OVClassConfigTestCPU, smoke_CpuExecNetworkCheckExecutionModeIsAvailableInCoreAndModel) {
Expand Down
Loading