Build: CUDA include path, nvcc arch string handling, embedding error text

- xmake: ensure CUDA headers visible for host-compiled operator code.
- nvidia.lua: coerce cuda_arch to string for xmake 3; align infiniop/
  infinirt device link -gencode with selected arch.
- embedding_nvidia: log cudaGetErrorString on launch failure for easier debug.

Made-with: Cursor

Author: Ceng23333

src/infiniop/ops/embedding/nvidia/embedding_nvidia.cu

@@ -4,6 +4,7 @@
 #include "../../../tensor.h"
 #include "../cuda/embedding_kernel.cuh"
 #include "embedding_nvidia.cuh"
+#include <cstdio>
 #include <cuda_runtime.h>
 
 template <typename T, typename IndexType>
@@ -212,9 +213,9 @@ infiniStatus_t Descriptor::calculate(
         return INFINI_STATUS_BAD_TENSOR_DTYPE;
     }
 
-    // Check for kernel launch errors
     cudaError_t err = cudaGetLastError();
     if (err != cudaSuccess) {
+        fprintf(stderr, "[infiniop embedding nvidia] cudaGetLastError: %s\n", cudaGetErrorString(err));
         return INFINI_STATUS_INTERNAL_ERROR;
     }
 

xmake.lua

@@ -54,6 +54,8 @@ option_end()
 
 if has_config("nv-gpu") then
     add_defines("ENABLE_NVIDIA_API")
+    local cuda_home = os.getenv("CUDA_HOME") or "/usr/local/cuda"
+    add_includedirs(path.join(cuda_home, "include"))
     includes("xmake/nvidia.lua")
 end
 

xmake/nvidia.lua

@@ -63,12 +63,17 @@ target("infiniop-nvidia")
 
     add_cuflags("-Xcompiler=-Wno-error=deprecated-declarations", "-Xcompiler=-Wno-error=unused-function")
 
+    -- xmake 3 may surface option values as non-string types; coerce so we do not fall back to "native".
     local arch_opt = get_config("cuda_arch")
-    if arch_opt and type(arch_opt) == "string" then
-        for _, arch in ipairs(arch_opt:split(",")) do
+    local arch_str = arch_opt and tostring(arch_opt):gsub("^%s+", ""):gsub("%s+$", "") or ""
+    if arch_str ~= "" and arch_str ~= "nil" then
+        for _, arch in ipairs(arch_str:split(",")) do
             arch = arch:trim()
             local compute = arch:gsub("sm_", "compute_")
-            add_cuflags("-gencode=arch=" .. compute .. ",code=" .. arch)
+            local gencode = "-gencode=arch=" .. compute .. ",code=" .. arch
+            add_cuflags(gencode)
+            -- Separate nvcc device-link step must use the same arch or linked GPU code defaults wrong.
+            add_culdflags(gencode)
         end
     else
         add_cugencodes("native")
@@ -101,6 +106,20 @@ target("infinirt-nvidia")
         add_cxxflags("-fPIC")
     end
 
+    local arch_opt_rt = get_config("cuda_arch")
+    local arch_str_rt = arch_opt_rt and tostring(arch_opt_rt):gsub("^%s+", ""):gsub("%s+$", "") or ""
+    if arch_str_rt ~= "" and arch_str_rt ~= "nil" then
+        for _, arch in ipairs(arch_str_rt:split(",")) do
+            arch = arch:trim()
+            local compute = arch:gsub("sm_", "compute_")
+            local gencode = "-gencode=arch=" .. compute .. ",code=" .. arch
+            add_cuflags(gencode)
+            add_culdflags(gencode)
+        end
+    else
+        add_cugencodes("native")
+    end
+
     set_languages("cxx17")
     add_files("../src/infinirt/cuda/*.cu")
 target_end()