issue/695 c++ infinicore::nn::module支持bf16

* issue/695 c++ infinicore::nn::module支持bf16
Signed-off-by: Ceng23333 <441651826@qq.com>

* embedding support f16
Signed-off-by: Ceng23333 <441651826@qq.com>

* resolve comments
Signed-off-by: Ceng23333 <441651826@qq.com>

* fix format
Signed-off-by: Ceng23333 <441651826@qq.com>

---------
Signed-off-by: Ceng23333 <441651826@qq.com>

src/infinicore/nn/embedding.cc

 #include "infinicore/nn/embedding.hpp"
 #include "infinicore/context/context.hpp"
 #include "infinicore/ops.hpp"
+#include <limits>
 #include <spdlog/spdlog.h>
 #include <stdexcept>
 
@@ -55,7 +56,6 @@ Tensor Embedding::forward(const Tensor &indices) const {
     // Flatten indices for sequential row copies
     auto cpu_device = Device(Device::Type::CPU, 0);
     auto indices_cpu = indices->to(cpu_device)->contiguous();
-    const auto *indices_data = reinterpret_cast<const int64_t *>(indices_cpu->data());
 
     // Calculate total number of lookups
     size_t num_lookups = 1;
@@ -63,17 +63,41 @@ Tensor Embedding::forward(const Tensor &indices) const {
         num_lookups *= dim;
     }
 
-    const size_t row_bytes = embedding_dim_ * (weight_->dtype() == DataType::F32 ? sizeof(float) : weight_->dtype() == DataType::BF16 ? sizeof(uint16_t)
-                                                                                                                                      : sizeof(float));
+    const size_t row_bytes = embedding_dim_ * dsize(weight_->dtype());
 
     // Source and destination base pointers
     auto *weight_base = weight_->data();
     auto *out_base = out->data();
 
+    // Helper lambda to read index based on dtype with bounds checking
+    auto read_index = [&](size_t i) -> int64_t {
+        auto dtype = indices_cpu->dtype();
+        if (dtype == DataType::I32) {
+            const auto *data = reinterpret_cast<const int32_t *>(indices_cpu->data());
+            return static_cast<int64_t>(data[i]);
+        } else if (dtype == DataType::I64) {
+            const auto *data = reinterpret_cast<const int64_t *>(indices_cpu->data());
+            return data[i];
+        } else if (dtype == DataType::U32) {
+            const auto *data = reinterpret_cast<const uint32_t *>(indices_cpu->data());
+            return static_cast<int64_t>(data[i]);
+        } else if (dtype == DataType::U64) {
+            const auto *data = reinterpret_cast<const uint64_t *>(indices_cpu->data());
+            uint64_t val = data[i];
+            // Check if value can fit in int64_t
+            if (val > static_cast<uint64_t>(std::numeric_limits<int64_t>::max())) {
+                throw std::out_of_range("Index value out of range for int64_t: " + std::to_string(val));
+            }
+            return static_cast<int64_t>(val);
+        } else {
+            throw std::runtime_error("Embedding indices must be integer type, got dtype=" + std::to_string(static_cast<int>(dtype)));
+        }
+    };
+
     if (weight_->device().getType() == Device::Type::CPU) {
         // CPU path: memcpy row by row
         for (size_t i = 0; i < num_lookups; ++i) {
-            int64_t idx = indices_data[i];
+            int64_t idx = read_index(i);
             if (idx < 0 || idx >= static_cast<int64_t>(num_embeddings_)) {
                 throw std::out_of_range(
                     "Index out of range: " + std::to_string(idx) + " (num_embeddings=" + std::to_string(num_embeddings_) + ")");
@@ -83,7 +107,7 @@ Tensor Embedding::forward(const Tensor &indices) const {
     } else {
         // Device path: use stream-ordered D2D copies
         for (size_t i = 0; i < num_lookups; ++i) {
-            int64_t idx = indices_data[i];
+            int64_t idx = read_index(i);
             if (idx < 0 || idx >= static_cast<int64_t>(num_embeddings_)) {
                 throw std::out_of_range(
                     "Index out of range: " + std::to_string(idx) + " (num_embeddings=" + std::to_string(num_embeddings_) + ")");

src/infinicore/nn/rope.cc

 #include "infinicore/nn/rope.hpp"
+#include "../../utils.h"
 #include "../utils.hpp"
 #include "infinicore/ops.hpp"
 #include <algorithm>
 #include <cmath>
 #include <functional>
 #include <stdexcept>
+#include <utility>
+#include <vector>
 
 namespace infinicore::nn {
 
@@ -50,36 +53,63 @@ void RoPE::initialize_cache() {
     for (size_t pos = 0; pos < max_seq_len_; pos++) {
         for (size_t j = 0; j < cache_dim; j++) {
             // GPT-J style inverse frequency: theta^(-2j/head_dim)
-            double inv_freq = 1.0 / std::pow(theta_, 2.0 * static_cast<double>(j) / static_cast<double>(head_dim_));
+            // Compute directly in float to avoid double->float casting
+            float inv_freq = 1.0f / std::pow(static_cast<float>(theta_), 2.0f * static_cast<float>(j) / static_cast<float>(head_dim_));
 
             // Compute angle: position * inverse_frequency
-            double angle = static_cast<double>(pos) * inv_freq;
+            float angle = static_cast<float>(pos) * inv_freq;
 
-            // Compute sin and cos
-            sin_data[pos * cache_dim + j] = static_cast<float>(std::sin(angle));
-            cos_data[pos * cache_dim + j] = static_cast<float>(std::cos(angle));
+            // Compute sin and cos directly on float
+            sin_data[pos * cache_dim + j] = std::sin(angle);
+            cos_data[pos * cache_dim + j] = std::cos(angle);
         }
     }
 
-    // Create CPU tensors and copy data
-    auto sin_cpu = Tensor::from_blob(sin_data.data(), {max_seq_len_, cache_dim}, DataType::F32, cpu_device);
-    auto cos_cpu = Tensor::from_blob(cos_data.data(), {max_seq_len_, cache_dim}, DataType::F32, cpu_device);
+    // Convert to target dtype on CPU (matching Python's numpy astype conversion pattern)
+    // Python: np_array.astype(ml_dtypes.bfloat16, copy=True) converts F32 -> BF16
+    if (dtype_ == DataType::F32) {
+        // Direct use of F32 data
+        auto sin_f32_cpu = Tensor::from_blob(sin_data.data(), {max_seq_len_, cache_dim}, DataType::F32, cpu_device);
+        auto cos_f32_cpu = Tensor::from_blob(cos_data.data(), {max_seq_len_, cache_dim}, DataType::F32, cpu_device);
+        sin_cache_->copy_from(sin_f32_cpu);
+        cos_cache_->copy_from(cos_f32_cpu);
+    } else if (dtype_ == DataType::BF16) {
+        // Convert F32 to BF16 using the same conversion as Python's ml_dtypes.bfloat16
+        // This uses round-to-nearest-even (matching _f32_to_bf16 implementation)
+        std::vector<bf16_t> sin_bf16_data(max_seq_len_ * cache_dim);
+        std::vector<bf16_t> cos_bf16_data(max_seq_len_ * cache_dim);
+
+        for (size_t i = 0; i < sin_data.size(); i++) {
+            sin_bf16_data[i] = utils::cast<bf16_t, float>(sin_data[i]);
+            cos_bf16_data[i] = utils::cast<bf16_t, float>(cos_data[i]);
+        }
+
+        auto sin_bf16_cpu = Tensor::from_blob(sin_bf16_data.data(), {max_seq_len_, cache_dim}, DataType::BF16, cpu_device);
+        auto cos_bf16_cpu = Tensor::from_blob(cos_bf16_data.data(), {max_seq_len_, cache_dim}, DataType::BF16, cpu_device);
 
-    // Copy to device
-    // Note: Cache is created with dtype_, but we compute in F32 for precision.
-    // If dtype_ != F32, copy_from will fail. For now, we only support F32 cache.
-    // TODO: Add dtype conversion support when cast operation is available
-    if (dtype_ != DataType::F32) {
+        // copy_from handles cross-device copying to target device
+        sin_cache_->copy_from(sin_bf16_cpu);
+        cos_cache_->copy_from(cos_bf16_cpu);
+    } else if (dtype_ == DataType::F16) {
+        // Convert F32 to F16
+        std::vector<fp16_t> sin_f16_data(max_seq_len_ * cache_dim);
+        std::vector<fp16_t> cos_f16_data(max_seq_len_ * cache_dim);
+
+        for (size_t i = 0; i < sin_data.size(); i++) {
+            sin_f16_data[i] = utils::cast<fp16_t, float>(sin_data[i]);
+            cos_f16_data[i] = utils::cast<fp16_t, float>(cos_data[i]);
+        }
+
+        auto sin_f16_cpu = Tensor::from_blob(sin_f16_data.data(), {max_seq_len_, cache_dim}, DataType::F16, cpu_device);
+        auto cos_f16_cpu = Tensor::from_blob(cos_f16_data.data(), {max_seq_len_, cache_dim}, DataType::F16, cpu_device);
+
+        sin_cache_->copy_from(sin_f16_cpu);
+        cos_cache_->copy_from(cos_f16_cpu);
+    } else {
         throw std::runtime_error(
-            "RoPE cache dtype conversion not yet supported. Please use DataType::F32 for cache. "
-            "Requested dtype: "
+            "RoPE cache dtype conversion not yet supported for dtype: "
             + std::to_string(static_cast<int>(dtype_)));
     }
-
-    // copy_from handles cross-device copying automatically
-    // Direct copy from CPU to target device avoids double copying
-    sin_cache_->copy_from(sin_cpu);
-    cos_cache_->copy_from(cos_cpu);
 }
 
 Tensor RoPE::forward(const Tensor &x, const Tensor &pos) const {

xmake.lua

@@ -376,6 +376,7 @@ target("infinicore_cpp_api")
     add_files("src/infinicore/tensor/*.cc")
     add_files("src/infinicore/nn/*.cc")
     add_files("src/infinicore/ops/*/*.cc")
+    add_files("src/utils/*.cc")
 
     set_installdir(INFINI_ROOT)
     add_installfiles("include/infinicore/(**.h)",    {prefixdir = "include/infinicore"})
@@ -415,6 +416,7 @@ target("_infinicore")
     add_files("src/infinicore/nn/*.cc")
     add_files("src/infinicore/ops/*/*.cc")
     add_files("src/infinicore/pybind11/**.cc")
+    add_files("src/utils/*.cc")
 
     set_installdir("python/infinicore")
 target_end()