feat: pybind11 bindings for xllm CUDA kernels

norm.cu compiled successfully on BI-V100 (only warning: fp8 __host__ attr).
Failed at import because no PYBIND11_MODULE — now fixed.

New bindings/ directory with 4 binding files:
  xllm_norm_bind.cpp      → rms_norm, fused_add_rms_norm
  xllm_activation_bind.cpp → silu_and_mul, gelu_and_mul, act_and_mul
  xllm_rope_bind.cpp       → rotary_embedding
  xllm_cache_bind.cpp      → reshape_paged_cache, block_copy

Build script updated: each .so = kernel .cu + binding .cpp

ex_engine/xllm_kernels/cuda/bindings/xllm_activation_bind.cpp

@@ -0,0 +1,18 @@
+// xllm_activation_bind.cpp
+#include <torch/extension.h>
+
+namespace xllm::kernel::cuda {
+void act_and_mul(torch::Tensor out, torch::Tensor input,
+                 const std::string& act_mode);
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("silu_and_mul", [](torch::Tensor out, torch::Tensor input) {
+        xllm::kernel::cuda::act_and_mul(out, input, "silu");
+    }, "SiLU and Mul", py::arg("out"), py::arg("input"));
+    m.def("gelu_and_mul", [](torch::Tensor out, torch::Tensor input) {
+        xllm::kernel::cuda::act_and_mul(out, input, "gelu");
+    }, "GELU and Mul", py::arg("out"), py::arg("input"));
+    m.def("act_and_mul", &xllm::kernel::cuda::act_and_mul,
+          "Activation and Mul", py::arg("out"), py::arg("input"), py::arg("act_mode"));
+}

ex_engine/xllm_kernels/cuda/bindings/xllm_cache_bind.cpp

@@ -0,0 +1,19 @@
+// xllm_cache_bind.cpp
+#include <torch/extension.h>
+
+namespace xllm::kernel::cuda {
+void reshape_paged_cache(torch::Tensor slot_ids, torch::Tensor keys,
+                         torch::Tensor values, torch::Tensor key_cache,
+                         torch::Tensor value_cache);
+void block_copy(torch::Tensor key_cache_ptrs, torch::Tensor value_cache_ptrs,
+                torch::Tensor src_block_indices, torch::Tensor dst_block_indices,
+                torch::Tensor cum_sum, int64_t numel_per_block,
+                torch::ScalarType cache_dtype);
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("reshape_paged_cache", &xllm::kernel::cuda::reshape_paged_cache,
+          "Reshape Paged KV Cache");
+    m.def("block_copy", &xllm::kernel::cuda::block_copy,
+          "Block Copy for KV Cache");
+}

ex_engine/xllm_kernels/cuda/bindings/xllm_norm_bind.cpp

@@ -0,0 +1,24 @@
+// xllm_norm_bind.cpp — pybind11 entry point for xllm norm kernels
+// Compiled together with norm.cu to produce xllm_norm.so
+//
+// Exports: rms_norm, fused_add_rms_norm
+
+#include <torch/extension.h>
+
+namespace xllm::kernel::cuda {
+void rms_norm(torch::Tensor output, torch::Tensor input,
+              torch::Tensor weight, double eps);
+void fused_add_rms_norm(torch::Tensor& input, torch::Tensor& residual,
+                        torch::Tensor& weight, double epsilon);
+}  // namespace xllm::kernel::cuda
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("rms_norm", &xllm::kernel::cuda::rms_norm,
+          "RMS Norm (xllm CUDA kernel)",
+          py::arg("output"), py::arg("input"),
+          py::arg("weight"), py::arg("eps") = 1e-6);
+    m.def("fused_add_rms_norm", &xllm::kernel::cuda::fused_add_rms_norm,
+          "Fused Add + RMS Norm (xllm CUDA kernel)",
+          py::arg("input"), py::arg("residual"),
+          py::arg("weight"), py::arg("epsilon") = 1e-6);
+}

ex_engine/xllm_kernels/cuda/bindings/xllm_rope_bind.cpp

@@ -0,0 +1,17 @@
+// xllm_rope_bind.cpp
+#include <torch/extension.h>
+#include <optional>
+
+namespace xllm::kernel::cuda {
+void rotary_embedding(torch::Tensor& positions, torch::Tensor& query,
+                      std::optional<torch::Tensor> key,
+                      torch::Tensor& cos_sin_cache, bool is_neox);
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("rotary_embedding", &xllm::kernel::cuda::rotary_embedding,
+          "Rotary Position Embedding (xllm CUDA kernel)",
+          py::arg("positions"), py::arg("query"),
+          py::arg("key"), py::arg("cos_sin_cache"),
+          py::arg("is_neox") = true);
+}

qwen3_6_scripts/build_xllm_kernels.sh

@@ -1,9 +1,8 @@
 #!/bin/bash
 # build_xllm_kernels.sh — Compile xllm CUDA kernels into .so on BI-V100
 #
-# Uses corex's own CUB (/usr/local/corex/include/cub/) NOT cccl_upstream
-# Source: ex_engine/xllm_kernels/cuda/
-# Output: qwen3_6_scripts/prebuilt/corex-3.2.3-ivcore10/
+# Uses corex CUB (/usr/local/corex/include/cub/) NOT cccl_upstream
+# Each .so = kernel .cu + pybind11 binding .cpp
 #
 # Run: bash qwen3_6_scripts/build_xllm_kernels.sh
 
@@ -12,20 +11,23 @@ SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
 PROJECT_DIR="$(dirname "$SCRIPT_DIR")"
 CUDA_DIR="${PROJECT_DIR}/ex_engine/xllm_kernels/cuda"
 HEADER_DIR="${CUDA_DIR}/headers"
+BIND_DIR="${CUDA_DIR}/bindings"
 PREBUILT_DIR="${SCRIPT_DIR}/prebuilt/corex-3.2.3-ivcore10"
 mkdir -p "$PREBUILT_DIR"
 
 build_kernel() {
     local name="$1"
-    local cu_file="$2"
+    shift
+    local sources="$@"
     echo "=== Building ${name}.so ==="
     python3 -c "
 import os, glob, shutil
 from torch.utils.cpp_extension import load
 
+sources = '${sources}'.split()
 mod = load(
     name='${name}',
-    sources=['${cu_file}'],
+    sources=sources,
     extra_cflags=['-std=c++17'],
     extra_include_paths=['${HEADER_DIR}', '/usr/local/corex/include'],
     verbose=True,
@@ -43,19 +45,24 @@ for f in glob.glob(os.path.join(build_dir, '*.so')):
 fns = [x for x in dir(mod) if not x.startswith('_')]
 print(f'Functions: {fns}')
 "
+    echo ""
 }
 
 echo "Building xllm CUDA kernels for BI-V100 (ivcore10)"
 echo "Using corex CUB: /usr/local/corex/include/cub/"
 echo ""
 
-# Build each kernel
-build_kernel "xllm_norm"       "${CUDA_DIR}/norm.cu"
-build_kernel "xllm_activation" "${CUDA_DIR}/activation.cu"
-build_kernel "xllm_rope"       "${CUDA_DIR}/rope.cu"
-build_kernel "xllm_block_copy" "${CUDA_DIR}/block_copy.cu"
-build_kernel "xllm_cache"      "${CUDA_DIR}/reshape_paged_cache.cu"
+build_kernel "xllm_norm" \
+    "${CUDA_DIR}/norm.cu" "${BIND_DIR}/xllm_norm_bind.cpp"
+
+build_kernel "xllm_activation" \
+    "${CUDA_DIR}/activation.cu" "${BIND_DIR}/xllm_activation_bind.cpp"
+
+build_kernel "xllm_rope" \
+    "${CUDA_DIR}/rope.cu" "${BIND_DIR}/xllm_rope_bind.cpp"
+
+build_kernel "xllm_cache" \
+    "${CUDA_DIR}/reshape_paged_cache.cu" "${CUDA_DIR}/block_copy.cu" "${BIND_DIR}/xllm_cache_bind.cpp"
 
-echo ""
 echo "=== All kernels built ==="
-ls -la "${PREBUILT_DIR}"/xllm_*.so 2>/dev/null
+ls -lh "${PREBUILT_DIR}"/xllm_*.so 2>/dev/null || echo "No .so files found"