dylanyunlong/project_6
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

fix: block_copy.cu remove utils.h (glog), CHECK→TORCH_CHECK

Browse Source 
3/4 kernels now compile:
  ✓ xllm_norm.so      (rms_norm, fused_add_rms_norm)
  ✓ xllm_activation.so (silu_and_mul, gelu_and_mul, act_and_mul)
  ✓ xllm_rope.so       (rotary_embedding)
  → xllm_cache.so      block_copy.cu had utils.h→glog — fixed
This commit is contained in:
claude
2026-08-14 11:02:57 +00:00
parent a206fc1d43
commit 900ae0b1ef
1 changed files with 10 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
ex_engine/xllm_kernels/cuda/block_copy.cu
View File

@@ -24,7 +24,6 @@ limitations under the License.
#include <type_traits>
#include "device_utils.cuh"
#include "utils.h"
namespace xllm::kernel::cuda {
namespace {
@@ -133,11 +132,11 @@ void block_copy(torch::Tensor key_cache_ptrs,
return;
}
CHECK(key_cache_ptrs.is_cuda());
CHECK(value_cache_ptrs.is_cuda());
CHECK(src_block_indices.is_cuda());
CHECK(dst_block_indices.is_cuda());
CHECK(cum_sum.is_cuda());
TORCH_CHECK(key_cache_ptrs.is_cuda());
TORCH_CHECK(value_cache_ptrs.is_cuda());
TORCH_CHECK(src_block_indices.is_cuda());
TORCH_CHECK(dst_block_indices.is_cuda());
TORCH_CHECK(cum_sum.is_cuda());
CHECK_EQ(key_cache_ptrs.scalar_type(), torch::kInt64);
CHECK_EQ(value_cache_ptrs.scalar_type(), torch::kInt64);
CHECK_EQ(src_block_indices.scalar_type(), torch::kInt32);
@@ -148,11 +147,11 @@ void block_copy(torch::Tensor key_cache_ptrs,
CHECK_EQ(src_block_indices.dim(), 1);
CHECK_EQ(dst_block_indices.dim(), 1);
CHECK_EQ(cum_sum.dim(), 1);
CHECK(key_cache_ptrs.is_contiguous());
CHECK(value_cache_ptrs.is_contiguous());
CHECK(src_block_indices.is_contiguous());
CHECK(dst_block_indices.is_contiguous());
CHECK(cum_sum.is_contiguous());
TORCH_CHECK(key_cache_ptrs.is_contiguous());
TORCH_CHECK(value_cache_ptrs.is_contiguous());
TORCH_CHECK(src_block_indices.is_contiguous());
TORCH_CHECK(dst_block_indices.is_contiguous());
TORCH_CHECK(cum_sum.is_contiguous());
CHECK_EQ(key_cache_ptrs.size(0), value_cache_ptrs.size(0));
CHECK_EQ(src_block_indices.size(0), cum_sum.size(0));
CHECK_GT(numel_per_block, 0);