Merge branch 'openvinotoolkit:master' into softmax_cross_entropy_loss_for_onnx_fe

AJThePro99 · web-flow · commit ba96b8f8de58 · 2025-02-11T15:29:01.000+05:30
diff --git a/src/common/transformations/src/transformations/convert_precision.cpp b/src/common/transformations/src/transformations/convert_precision.cpp
@@ -1417,13 +1417,7 @@ bool fuse_type_to_constant(const std::shared_ptr<ov::Node>& node,
         new_const->validate_and_infer_types();
         new_const->set_friendly_name(constant->get_friendly_name());
         ov::copy_runtime_info(constant, new_const);
-
-        const auto& rt_info = node->get_rt_info();
-        auto weightless_caching_attr = rt_info.find(ov::WeightlessCacheAttribute::get_type_info_static());
-        if (weightless_caching_attr != rt_info.end()) {
-            new_const->get_rt_info()[ov::WeightlessCacheAttribute::get_type_info_static()] =
-                weightless_caching_attr->second;
-        }
+        ov::copy_weightless_cache_attr(constant, new_const);
         return true;
     }
     return false;
diff --git a/src/core/dev_api/openvino/core/rt_info/weightless_caching_attributes.hpp b/src/core/dev_api/openvino/core/rt_info/weightless_caching_attributes.hpp
@@ -10,6 +10,8 @@
 
 namespace ov {
 
+OPENVINO_API void copy_weightless_cache_attr(const std::shared_ptr<Node>& from, const std::shared_ptr<Node>& to);
+
 /**
  * @brief Holds weightless caching attributes of a single constant.
  *
diff --git a/src/core/src/op/util/weightless_caching_attributes.cpp b/src/core/src/op/util/weightless_caching_attributes.cpp
@@ -7,3 +7,13 @@
 bool ov::WeightlessCacheAttribute::is_copyable() const {
     return false;
 }
+
+OPENVINO_API void ov::copy_weightless_cache_attr(const std::shared_ptr<ov::Node>& from,
+                                                 const std::shared_ptr<ov::Node>& to) {
+    const auto& rt_info = from->get_rt_info();
+    auto weightless_caching_attr = rt_info.find(ov::WeightlessCacheAttribute::get_type_info_static());
+
+    if (weightless_caching_attr != rt_info.end()) {
+        to->get_rt_info()[ov::WeightlessCacheAttribute::get_type_info_static()] = weightless_caching_attr->second;
+    }
+}
diff --git a/src/core/src/pass/constant_folding.cpp b/src/core/src/pass/constant_folding.cpp
@@ -7,6 +7,7 @@
 #include "openvino/cc/pass/itt.hpp"
 #include "openvino/core/constant_fold_utils.hpp"
 #include "openvino/core/rt_info.hpp"
+#include "openvino/core/rt_info/weightless_caching_attributes.hpp"
 #include "openvino/op/constant.hpp"
 #include "openvino/op/convert.hpp"
 #include "openvino/op/util/op_types.hpp"
@@ -153,6 +154,7 @@ bool ov::pass::ConstantFolding::run_on_model(const std::shared_ptr<ov::Model>& m
                     copy_runtime_info_from_input_values(original_node);
                     // Propagate runtime info attributes to replacement
                     copy_runtime_info(original_node, replacement_ptr);
+                    ov::copy_weightless_cache_attr(original_node, replacement_ptr);
 
                     rewritten = true;
                 }
diff --git a/src/inference/dev_api/openvino/runtime/system_conf.hpp b/src/inference/dev_api/openvino/runtime/system_conf.hpp
@@ -295,14 +295,6 @@ OPENVINO_RUNTIME_API void reserve_available_cpus(const std::vector<std::vector<i
  */
 OPENVINO_RUNTIME_API void set_cpu_used(const std::vector<int>& cpu_ids, const int used);
 
-/**
- * @brief      Get socket id by current numa node id
- * @ingroup    ov_dev_api_system_conf
- * @param[in]  numa_node_id numa node id
- * @return     socket id
- */
-OPENVINO_RUNTIME_API int get_socket_by_numa_node(int numa_node_id);
-
 /**
  * @brief      Get original socket id by current socket id, the input socket id is recalculated after filtering (like
  * numactl), while the original socket id is the original id before filtering
diff --git a/src/inference/dev_api/openvino/runtime/threading/cpu_streams_executor_internal.hpp b/src/inference/dev_api/openvino/runtime/threading/cpu_streams_executor_internal.hpp
@@ -42,6 +42,7 @@ void get_cur_stream_info(const int stream_id,
                          int& concurrency,
                          int& core_type,
                          int& numa_node_id,
+                         int& socket_id,
                          int& max_threads_per_core);
 
 /**
diff --git a/src/inference/src/dev/threading/cpu_streams_executor.cpp b/src/inference/src/dev/threading/cpu_streams_executor.cpp
@@ -104,10 +104,11 @@ struct CPUStreamsExecutor::Impl {
                                    const int concurrency,
                                    const int core_type,
                                    const int numa_node_id,
+                                   const int socket_id,
                                    const int max_threads_per_core) {
             auto stream_processors = _impl->_config.get_stream_processor_ids();
-            _numaNodeId = std::max(0, numa_node_id);
-            _socketId = get_socket_by_numa_node(_numaNodeId);
+            _numaNodeId = numa_node_id;
+            _socketId = socket_id;
             if (stream_type == STREAM_WITHOUT_PARAM) {
                 _taskArena.reset(new custom::task_arena{custom::task_arena::constraints{}
                                                             .set_max_concurrency(concurrency)
@@ -159,6 +160,7 @@ struct CPUStreamsExecutor::Impl {
             int concurrency;
             int cpu_core_type;
             int numa_node_id;
+            int socket_id;
             int max_threads_per_core;
             StreamCreateType stream_type;
             const auto org_proc_type_table = get_org_proc_type_table();
@@ -173,6 +175,7 @@ struct CPUStreamsExecutor::Impl {
                                 concurrency,
                                 cpu_core_type,
                                 numa_node_id,
+                                socket_id,
                                 max_threads_per_core);
             if (concurrency <= 0) {
                 return;
@@ -182,6 +185,7 @@ struct CPUStreamsExecutor::Impl {
                                   concurrency,
                                   cpu_core_type,
                                   numa_node_id,
+                                  socket_id,
                                   max_threads_per_core);
         }
 #endif
diff --git a/src/inference/src/dev/threading/cpu_streams_executor_internal.cpp b/src/inference/src/dev/threading/cpu_streams_executor_internal.cpp
@@ -21,6 +21,7 @@ void get_cur_stream_info(const int stream_id,
                          int& concurrency,
                          int& core_type,
                          int& numa_node_id,
+                         int& socket_id,
                          int& max_threads_per_core) {
     int stream_total = 0;
     size_t stream_info_id = 0;
@@ -36,6 +37,7 @@ void get_cur_stream_info(const int stream_id,
     concurrency = streams_info_table[stream_info_id][THREADS_PER_STREAM];
     core_type = streams_info_table[stream_info_id][PROC_TYPE];
     numa_node_id = streams_info_table[stream_info_id][STREAM_NUMA_NODE_ID];
+    socket_id = streams_info_table[stream_info_id][STREAM_SOCKET_ID];
     max_threads_per_core = 1;
     if (core_type == ALL_PROC) {
         for (size_t i = stream_info_id + 1; i < streams_info_table.size(); i++) {
diff --git a/src/inference/src/system_conf.cpp b/src/inference/src/system_conf.cpp
@@ -285,10 +285,6 @@ void reserve_available_cpus(const std::vector<std::vector<int>> streams_info_tab
                             const int cpu_status) {}
 void set_cpu_used(const std::vector<int>& cpu_ids, const int used) {}
 
-int get_socket_by_numa_node(int numa_node_id) {
-    return -1;
-};
-
 int get_org_socket_id(int socket_id) {
     return -1;
 }
@@ -352,16 +348,6 @@ void reserve_available_cpus(const std::vector<std::vector<int>> streams_info_tab
                             const int cpu_status) {}
 void set_cpu_used(const std::vector<int>& cpu_ids, const int used) {}
 
-int get_socket_by_numa_node(int numa_node_id) {
-    CPU& cpu = cpu_info();
-    for (size_t i = 0; i < cpu._proc_type_table.size(); i++) {
-        if (cpu._proc_type_table[i][PROC_NUMA_NODE_ID] == numa_node_id) {
-            return cpu._proc_type_table[i][PROC_SOCKET_ID];
-        }
-    }
-    return -1;
-};
-
 int get_org_socket_id(int socket_id) {
     CPU& cpu = cpu_info();
     auto iter = cpu._socketid_mapping_table.find(socket_id);
@@ -522,16 +508,6 @@ void set_cpu_used(const std::vector<int>& cpu_ids, const int used) {
     }
 }
 
-int get_socket_by_numa_node(int numa_node_id) {
-    CPU& cpu = cpu_info();
-    for (int i = 0; i < cpu._processors; i++) {
-        if (cpu._cpu_mapping_table[i][CPU_MAP_NUMA_NODE_ID] == numa_node_id) {
-            return cpu._cpu_mapping_table[i][CPU_MAP_SOCKET_ID];
-        }
-    }
-    return -1;
-}
-
 int get_number_of_logical_cpu_cores(bool bigCoresOnly) {
     int logical_cores = parallel_get_max_threads();
 #    if (OV_THREAD == OV_THREAD_TBB || OV_THREAD == OV_THREAD_TBB_AUTO)
diff --git a/src/inference/tests/unit/cpu_stream_info_test.cpp b/src/inference/tests/unit/cpu_stream_info_test.cpp
@@ -63,6 +63,7 @@ class LinuxCpuStreamTypeTests : public ov::test::TestsCommon,
             int test_concurrency;
             int test_core_type;
             int test_numa_node_id;
+            int test_socket_id = 0;
             int test_max_threads_per_core;
             get_cur_stream_info(i,
                                 test_data._cpu_pinning,
@@ -72,6 +73,7 @@ class LinuxCpuStreamTypeTests : public ov::test::TestsCommon,
                                 test_concurrency,
                                 test_core_type,
                                 test_numa_node_id,
+                                test_socket_id,
                                 test_max_threads_per_core);
             test_stream_types.push_back(test_stream_type);
             test_concurrencys.push_back(test_concurrency);
diff --git a/src/plugins/intel_cpu/src/compiled_model.cpp b/src/plugins/intel_cpu/src/compiled_model.cpp
@@ -153,7 +153,7 @@ CompiledModel::GraphGuard::Lock CompiledModel::get_graph() const {
     auto streamsExecutor = std::dynamic_pointer_cast<IStreamsExecutor>(m_task_executor);
     if (nullptr != streamsExecutor) {
         streamId = streamsExecutor->get_stream_id();
-        socketId = streamsExecutor->get_socket_id();
+        socketId = std::max(0, streamsExecutor->get_socket_id());
     }
     auto graphLock = GraphGuard::Lock(m_graphs[streamId % m_graphs.size()]);
     if (!graphLock._graph.IsReady()) {
diff --git a/src/plugins/intel_cpu/src/graph_context.cpp b/src/plugins/intel_cpu/src/graph_context.cpp
@@ -29,7 +29,7 @@ GraphContext::GraphContext(Config config,
       m_memoryControl(m_auxiliaryNetworkMemoryControl->createMemoryControlUnit()) {
     if (m_streamExecutor) {
         m_cpuStreamExecutor = std::dynamic_pointer_cast<ov::threading::CPUStreamsExecutor>(m_streamExecutor);
-        m_numaNodeId = m_cpuStreamExecutor ? m_cpuStreamExecutor->get_numa_node_id() : 0;
+        m_numaNodeId = m_cpuStreamExecutor ? std::max(0, m_cpuStreamExecutor->get_numa_node_id()) : 0;
         auto nNumaNodes = get_num_numa_nodes();
         if (m_numNumaNodes < nNumaNodes) {
             m_numNumaNodes = nNumaNodes;
diff --git a/src/plugins/intel_cpu/src/nodes/kernels/scaled_attn/executor_pa.cpp b/src/plugins/intel_cpu/src/nodes/kernels/scaled_attn/executor_pa.cpp
@@ -1048,7 +1048,7 @@ static void pack_32xK_kernel(T* dst, T* src, size_t dst_stride, size_t src_strid
     static const uint64_t idx[8] = {0, 4, 1, 5, 2, 6, 3, 7};
     auto midx = _mm512_loadu_si512(idx);
     __mmask16 mask = (1 << K) - 1;
-    for (size_t i = 0; i < K; i++) {
+    for (size_t i = 0; i < 16; i++) {
         auto x = _mm256_maskz_loadu_epi16(mask, src);               // [a1  a2  a3 a4]   total 256-bits in 4 64bits unit
         auto y = _mm256_maskz_loadu_epi16(mask, src + src_stride);  // [b1  b2  b3 b4]   total 256-bits
         auto a = _mm512_castsi256_si512(x);
@@ -2356,6 +2356,28 @@ struct AttentionExecutor : public PagedAttentionExecutor {
                 _slot_mapping.ptr<int32_t>()[idx++] =
                     block_number * _helper._block_size + block_offset % _helper._block_size;
             }
+            // To simplify tails of the kernels for Q*K and W*V:
+            // for first token kernels:
+            //    Q*K aka [m, k0] * [n0, k0]', there is already tails logic for m, k0, but no(always assume n0 is
+            //    block_size) for n0 which means the tail of k_cache need to be set to zero.
+            //    W*V aka [m, k1] * [n1, k1]', there is no tails handing for n1, so tails of v_cache need to be set to
+            //    zero.
+            // for second token, the kernels have tails handling logic
+            if (q_len != 1 && kv_len % _helper._block_size != 0) {
+                // block no. which contains tails
+                auto block_number = block_indices.ptr<int32_t>()[block_number_start + kv_len / _helper._block_size];
+                // tails start position
+                auto block_offset = kv_len % _helper._block_size;
+                // tails number
+                auto zero_tokens = _helper._block_size - block_offset;
+                auto S = k_cache.m_dims[3];
+                auto SV = v_cache.m_dims[3];
+                auto Hk = k_cache.m_dims[1];    // shape: [block, H, 32, S]
+                parallel_for2d(Hk, zero_tokens, [&](size_t h, size_t l) {
+                    std::memset(k_cache.ptr_v(block_number, h, block_offset + l, 0), 0, S * k_cache.m_element_size);
+                    std::memset(v_cache.ptr_v(block_number, h, block_offset + l, 0), 0, SV * v_cache.m_element_size);
+                });
+            }
         }
 
         if (k_cache.m_dt == ov::element::Type_t::u8) {
diff --git a/src/plugins/intel_cpu/src/nodes/kernels/x64/brgemm_kernel.cpp b/src/plugins/intel_cpu/src/nodes/kernels/x64/brgemm_kernel.cpp
@@ -273,6 +273,7 @@ void BrgemmKernel::init_brgemm_copy_a(
                         : DnnlExtensionUtils::sizeOfDataType(static_cast<dnnl::memory::data_type>(dt_in0));
     brgCopyKernelConf.transposed_A = transpose;
     brgCopyKernelConf.isa = is_avx_f16_only ? avx512_core_fp16 : avx512_core_amx;
+    brgCopyKernelConf.orig_wei_dt = static_cast<dnnl_data_type_t>(DnnlExtensionUtils::ElementTypeToDataType(inType));
 
     create_brgemm_matmul_copy_a(brgCopyKernel, &brgCopyKernelConf);
 }
diff --git a/src/plugins/intel_cpu/tests/functional/shared_tests_instances/skip_tests_config.cpp b/src/plugins/intel_cpu/tests/functional/shared_tests_instances/skip_tests_config.cpp
@@ -479,18 +479,12 @@ std::vector<std::string> disabledTestPatterns() {
         R"(.*smoke_MatMulCompressedWeights_corner_cases_basic/MatmulWeightsDecompression.CompareWithRefs/data_shape=\[\?.\?.\?\]_\(\[1,1,4096\]\)_weights_shape=\[4096,4096\]_group_size=128_weights_precision=nf4_decompression_precision=f16_scale_precision=undefined_transpose_weights=0_decompression_subtract=full_reshape_on_decompression=1_config=\(\).*)");
     retVector.emplace_back(R"(.*smoke_RDFT_CPU_1D/RDFTTestCPU.CompareWithRefs/prec=f32_IS0=\[\]_TS0=\(\(126\)\)_constAxes=true_axes=\(\(0\)\)_isInverse=false.*)");
     retVector.emplace_back(R"(.*smoke_RDFT_CPU_2D/RDFTTestCPU.CompareWithRefs/prec=f32_IS0=\[\]_TS0=\(\(16.38\)\)_constAxes=true_axes=\(\(0.1\)\)_isInverse=false.*)");
-    // Issue: MFDNN-12818
-    retVector.emplace_back(R"(.*smoke_LPT/RecurrentCellTransformation.CompareWithRefImpl/f32_\[1,1,3\]_CPU_f32FQ_X_level=256_.*_FQ_W_level=255.*)");
-    retVector.emplace_back(R"(.*smoke_static/ConvertFqRnnToQuantizedRnn.CompareWithRefs/Type=GRUSequence.*2.5.10.*2.1.4.*2.1.4.*)");
 #endif
     if (!ov::with_cpu_x86_avx512_core()) {
         // on platforms which do not support bfloat16, we are disabling bf16 tests since there are no bf16 primitives,
         // tests are useless on such platforms
         retVector.emplace_back(R"(.*(BF|bf)16.*)");
         retVector.emplace_back(R"(.*bfloat16.*)");
-        // Issue: MFDNN-12818
-        retVector.emplace_back(R"(.*smoke_LPT/RecurrentCellTransformation.CompareWithRefImpl/f32_\[1,1,3\]_CPU_f32FQ_X_level=256_.*_FQ_W_level=255.*)");
-        retVector.emplace_back(R"(.*smoke_static/ConvertFqRnnToQuantizedRnn.CompareWithRefs/Type=GRUSequence.*2.5.10.*2.1.4.*2.1.4.*)");
     }
     if (!ov::with_cpu_x86_avx2()) {
         // MatMul in Snippets uses BRGEMM that is supported only on AVX2 (and newer) platforms
diff --git a/src/plugins/intel_cpu/thirdparty/onednn b/src/plugins/intel_cpu/thirdparty/onednn
@@ -1 +1 @@
-Subproject commit 157fb8153d8c54cf37311d95a63ed63e66b14ac5
+Subproject commit 0bbc7a77db0a9eb60c97cb96fb60a08185575bdb
diff --git a/src/plugins/intel_gpu/include/intel_gpu/primitives/data.hpp b/src/plugins/intel_gpu/include/intel_gpu/primitives/data.hpp
diff --git a/src/plugins/intel_gpu/src/graph/graph_optimizer/propagate_constants.cpp b/src/plugins/intel_gpu/src/graph/graph_optimizer/propagate_constants.cpp
diff --git a/src/plugins/intel_gpu/src/graph/include/pass_manager.h b/src/plugins/intel_gpu/src/graph/include/pass_manager.h
diff --git a/src/plugins/intel_gpu/src/plugin/transformations/convert_fc_to_compressed.cpp b/src/plugins/intel_gpu/src/plugin/transformations/convert_fc_to_compressed.cpp
diff --git a/src/plugins/intel_gpu/tests/functional/behavior/model_cache.cpp b/src/plugins/intel_gpu/tests/functional/behavior/model_cache.cpp

Original file line number	Diff line number	Diff line change
`@@ -10,6 +10,8 @@`
`10`	`10`
`11`	`11`	`namespace ov {`
`12`	`12`
	`13`	`+OPENVINO_API void copy_weightless_cache_attr(const std::shared_ptr<Node>& from, const std::shared_ptr<Node>& to);`
	`14`	`+`
`13`	`15`	`/**`
`14`	`16`	`* @brief Holds weightless caching attributes of a single constant.`
`15`	`17`	`*`