resolve matmul_nbits.cc

luoyu-intel · Nov 17, 2023 · 4c07742 · 4c07742
1 parent f5ada45
commit 4c07742
Showing 1 changed file with 30 additions and 7 deletions.
diff --git a/onnxruntime/contrib_ops/cpu/quantization/matmul_nbits.cc b/onnxruntime/contrib_ops/cpu/quantization/matmul_nbits.cc
@@ -9,7 +9,6 @@
 #include "core/mlas/inc/mlas_q4.h"
 #include "core/providers/cpu/math/matmul_helper.h"
 #include "core/providers/common.h"
-#include "core/mlas/inc/mlas_q4.h"
 
 namespace onnxruntime {
 namespace contrib {
@@ -26,6 +25,13 @@ class MatMulNBits final : public OpKernel {
                 "Only 4b quantization is supported for MatMulNBits op, additional bits support is planned.");
     info.GetAttrOrDefault<int64_t>("accuracy_level", &accuracy_level_, 0);
     is_asym_ = info.GetInputCount() >= 4;
+    const Tensor* tensor_B = nullptr;
+    const Tensor* tensor_scale = nullptr;
+    const Tensor* tensor_zero_point = nullptr;
+    bool get_B = info.TryGetConstantInput(0, &tensor_B);
+    bool get_scale = info.TryGetConstantInput(1, &tensor_scale);
+    bool get_zero_point = info.TryGetConstantInput(2, &tensor_zero_point);
+    all_constant_ = get_B && get_scale && get_zero_point;
   }
 
   Status Compute(OpKernelContext* context) const override;
@@ -46,22 +52,30 @@ class MatMulNBits final : public OpKernel {
   IAllocatorUniquePtr<void> packed_b_;
   size_t packed_b_size_;
   bool is_asym_;
+  bool all_constant_;
   int64_t accuracy_level_;
 };
 
 Status MatMulNBits::PrePack(const Tensor& tensor, int input_idx, /*out*/ AllocatorPtr alloc,
                             /*out*/ bool& is_packed,
                             /*out*/ PrePackedWeights* prepacked_weights) {
   is_packed = false;
-  auto compt_type = static_cast<MLAS_COMPUTE_TYPE>(accuracy_level_);
+  if (!all_constant_) {
+    return;
+  }
   if (MlasNBitsGemmPackBSupport(N_, K_, block_size_, static_cast<int>(nbits_), is_asym_, compt_type)) {
-    // TODO use threadpool here
+    auto compt_type = static_cast<MLAS_COMPUTE_TYPE>(accuracy_level_);
+    // better to use threadpool here, LLM weight will consume a lot of time
     MLAS_THREADPOOL* pool = NULL;
     if (input_idx == 1) {
       auto qptr = tensor.Data<uint8_t>();
       packed_b_size_ = MlasNBitsGemmPackBSize(N_, K_, block_size_, static_cast<int>(nbits_), is_asym_, compt_type);
       packed_b_ = IAllocator::MakeUniquePtr<void>(alloc, packed_b_size_, true);
-      MlasNBitsGemmPackB(packed_b_.get(), qptr, nullptr, nullptr, N_, K_, K_, block_size_, static_cast<int>(nbits_), is_asym_, false, compt_type, pool);
+      if (packed_b_ == nullptr) {
+        return;
+      }
+      MlasNBitsGemmPackB(packed_b_.get(), qptr, nullptr, nullptr, N_, K_, K_, block_size_, static_cast<int>(nbits_),
+                         is_asym_, false, compt_type, pool);
       if (prepacked_weights) {
         prepacked_weights->buffers_.push_back(std::move(packed_b_));
         prepacked_weights->buffer_sizes_.push_back(packed_b_size_);
@@ -70,7 +84,11 @@ Status MatMulNBits::PrePack(const Tensor& tensor, int input_idx, /*out*/ Allocat
     }
     if (input_idx == 2) {
       auto sptr = tensor.Data<float>();
-      MlasNBitsGemmPackB(packed_b_.get(), nullptr, sptr, nullptr, N_, K_, K_, block_size_, static_cast<int>(nbits_), is_asym_, !is_asym_, compt_type, pool);
+      if (packed_b_ == nullptr) {
+        return;
+      }
+      MlasNBitsGemmPackB(packed_b_.get(), nullptr, sptr, nullptr, N_, K_, K_, block_size_, static_cast<int>(nbits_),
+                         is_asym_, !is_asym_, compt_type, pool);
       if (prepacked_weights) {
         prepacked_weights->buffers_.push_back(std::move(packed_b_));
         prepacked_weights->buffer_sizes_.push_back(packed_b_size_);
@@ -79,7 +97,11 @@ Status MatMulNBits::PrePack(const Tensor& tensor, int input_idx, /*out*/ Allocat
     }
     if (input_idx == 3) {
       auto zptr = tensor.Data<uint8_t>();
-      MlasNBitsGemmPackB(packed_b_.get(), nullptr, nullptr, zptr, N_, K_, K_, block_size_, static_cast<int>(nbits_), is_asym_, is_asym_, compt_type, pool);
+      if (packed_b_ == nullptr) {
+        return;
+      }
+      MlasNBitsGemmPackB(packed_b_.get(), nullptr, nullptr, zptr, N_, K_, K_, block_size_, static_cast<int>(nbits_),
+                         is_asym_, is_asym_, compt_type, pool);
       if (prepacked_weights) {
         prepacked_weights->buffers_.push_back(std::move(packed_b_));
         prepacked_weights->buffer_sizes_.push_back(packed_b_size_);
@@ -140,7 +162,8 @@ Status MatMulNBits::Compute(OpKernelContext* ctx) const {
     AllocatorPtr allocator;
     auto status = ctx->GetTempSpaceAllocator(&allocator);
     ORT_RETURN_IF_ERROR(status);
-    auto ws_ptr = IAllocator::MakeUniquePtr<float>(allocator, SafeInt<size_t>(K) * M);  // workspace for activation process(dynamic quantization and others)
+    // workspace for activation process(dynamic quantization and others)
+    auto ws_ptr = IAllocator::MakeUniquePtr<float>(allocator, SafeInt<size_t>(K) * std::max(M, 32));
     for (size_t i = 0; i < max_len; i++) {
       gemm_params[i].A = a_data + helper.LeftOffsets()[i];
       gemm_params[i].lda = lda;