fmt

Author: qinyiqun

include/infinicore/quantization/gptq_qy.hpp

@@ -83,9 +83,6 @@ inline std::vector<uint16_t> float_to_fp16_bits(const std::vector<float> &values
 }
 } // anonymous namespace
 
-// ============================================================================
-// GPTQ_QY Class (最终修复版)
-// ============================================================================
 namespace infinicore::quantization {
 
 class GPTQ_QY : public BaseQuantization {
@@ -123,9 +120,6 @@ class GPTQ_QY : public BaseQuantization {
         const int bits = weight_bits();
         const int values_per_int32 = 32 / bits;
 
-        // --------------------------------------------------------------------
-        // 1. qweight 转换（保持不变）
-        // --------------------------------------------------------------------
         {
             const auto &shape = original_qweight->shape();
             assert(shape.size() == 2);
@@ -148,9 +142,6 @@ class GPTQ_QY : public BaseQuantization {
                 target_device);
         }
 
-        // --------------------------------------------------------------------
-        // 2. qzeros 转换（保持 int32 -> fp32 -> fp16 逻辑）
-        // --------------------------------------------------------------------
         {
             const auto &shape = original_qzeros->shape();
             assert(shape.size() == 2);
@@ -167,24 +158,15 @@ class GPTQ_QY : public BaseQuantization {
                 target_device);
         }
 
-        // --------------------------------------------------------------------
-        // 3. scales 转换（核心修复：如果输入已是 FP16，直接内存拷贝）
-        // --------------------------------------------------------------------
         {
             auto scales_cpu = original_scales->to(Device::Type::CPU);
             size_t num_elements = scales_cpu->numel();
             const void *raw_data = scales_cpu->data();
 
             std::vector<uint16_t> scales_fp16(num_elements);
-
-            // 关键：根据输入 dtype 决定处理方式
             if (scales_cpu->dtype() == DataType::F16) {
-                // 输入已经是 FP16，直接 memcpy，不做任何转换！
-                // spdlog::info("Scales is already FP16, performing direct memory copy");
                 std::memcpy(scales_fp16.data(), raw_data, num_elements * sizeof(uint16_t));
             } else if (scales_cpu->dtype() == DataType::F32) {
-                // 输入是 FP32，才需要转换
-                // spdlog::info("Scales is FP32, converting to FP16");
                 std::vector<float> scales_fp32(num_elements);
                 std::memcpy(scales_fp32.data(), raw_data, num_elements * sizeof(float));
                 scales_fp16 = ::float_to_fp16_bits(scales_fp32);
@@ -201,19 +183,15 @@ class GPTQ_QY : public BaseQuantization {
                 target_device);
         }
 
-        // --------------------------------------------------------------------
-        // 4. g_idx 处理
-        // --------------------------------------------------------------------
         if (g_idx->numel() > 0) {
             g_idx_ = g_idx->to(target_device);
         }
 
         converted_ = true;
-        // spdlog::info("GPTQ_QY conversion completed successfully");
     }
 
     void release_buffers() {
-        converted_weight_ = Tensor(); // 赋值为空 Tensor，释放显存
+        converted_weight_ = Tensor();
         converted_zeros_ = Tensor();
         converted_scales_ = Tensor();
         g_idx_ = Tensor();
@@ -226,16 +204,12 @@ class GPTQ_QY : public BaseQuantization {
             return;
         }
 
-        // 1. 执行转换（只读传入的原始数据）
         convert_from_gptq_w4a16(weight, zeros, scales, g_idx, target_device);
 
-        // 2. 转移所有权（Move 语义：converted_weight_ 的 impl_ 指针会置为 nullptr）
-        //    原 weight/zeros/scales 持有的旧 shared_ptr 会被自动析构，释放显存
         weight = std::move(converted_weight_);
         zeros = std::move(converted_zeros_);
         scales = std::move(converted_scales_);
 
-        // 3. 清理内部状态
         converted_ = false;
         spdlog::debug("GPTQ_QY: Ownership transferred, internal buffers cleared.");
     }
@@ -343,4 +317,4 @@ class GPTQ_QY : public BaseQuantization {
     bool converted_ = false;
 };
 
-} // namespace infinicore::quantization
+} // namespace infinicore::quantization

src/infinicore/nn/linear.cc

@@ -6,7 +6,6 @@
 #include "infinicore/ops/linear_w4a16_awq.hpp"
 #include "infinicore/ops/linear_w4a16_gptq_qy.hpp"
 #include "infinicore/ops/linear_w8a8i8.hpp"
-#include <iostream>
 #include <optional>
 #include <spdlog/spdlog.h>
 
@@ -130,22 +129,6 @@ void BaseLinear::process_weights_after_loading() {
         assert(quantization_->get_quant_scheme() == infinicore::quantization::QuantScheme::GPTQ_W4A16_QY);
     }
 }
-// void BaseLinear::process_weights_after_loading() {
-//     if (quantization_->get_quant_scheme() == infinicore::quantization::QuantScheme::GPTQ_W4A16 && device_.getType() == Device::Type::QY) {
-
-//         // 1. 创建目标量化器
-//         auto config = quantization_->get_config();
-//         auto gptq_qy = std::make_shared<infinicore::quantization::GPTQ_QY>(config);
-
-//         // 2. 🚀 一行完成：转换 + 所有权转移 + 旧显存释放
-//         gptq_qy->convert_and_take_ownership(weight_, weight_zeros_, weight_scale_, gidx_, device_);
-
-//         // 3. 替换量化策略对象
-//         quantization_ = std::move(gptq_qy);
-
-//         assert(quantization_->get_quant_scheme() == infinicore::quantization::QuantScheme::GPTQ_W4A16_QY);
-//     }
-// }
 } // namespace infinicore::nn
 
 namespace infinicore::nn {

src/infinicore/pybind11/tensor.hpp

@@ -81,4 +81,4 @@ inline void bind(py::module &m) {
         pybind11::arg("raw_ptr"), pybind11::arg("shape"), pybind11::arg("strides"), pybind11::arg("dtype"), pybind11::arg("device"));
 }
 
-} // namespace infinicore::tensor
+} // namespace infinicore::tensor