Apply clang-format formatting fixes

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: TimDettmers

csrc/compat.cuh

@@ -25,8 +25,8 @@
 
 #if BNB_HIP
 
-#include <hip/hip_fp16.h>
 #include <hip/hip_bfloat16.h>
+#include <hip/hip_fp16.h>
 #include <hip/hip_math_constants.h>
 #include <hip/hip_runtime.h>
 #include <hipblas/hipblas.h>
@@ -94,7 +94,9 @@ using bnb_error_t = cudaError_t;
 // ============================================================================
 
 #if BNB_HIP
-#define __syncwarp() do {} while(0)
+#define __syncwarp()                                                                                                   \
+    do {                                                                                                               \
+    } while (0)
 #endif
 
 // ============================================================================

csrc/kernels.cu

@@ -331,19 +331,16 @@ __global__ void kQuantizeBlockwise(
     // block_dim=32 (from BLOCK_SIZE=64/NUM_PER_TH=2) is too small. Fall back
     // to DIRECT load/store in that case.
     static constexpr int THREADS = BLOCK_SIZE / NUM_PER_TH;
-    static constexpr auto LOAD_ALGO = (THREADS >= BNB_WARP_SIZE)
-        ? bnb_cub::BLOCK_LOAD_WARP_TRANSPOSE : bnb_cub::BLOCK_LOAD_DIRECT;
-    static constexpr auto STORE_ALGO = (THREADS >= BNB_WARP_SIZE)
-        ? bnb_cub::BLOCK_STORE_WARP_TRANSPOSE : bnb_cub::BLOCK_STORE_DIRECT;
+    static constexpr auto LOAD_ALGO =
+        (THREADS >= BNB_WARP_SIZE) ? bnb_cub::BLOCK_LOAD_WARP_TRANSPOSE : bnb_cub::BLOCK_LOAD_DIRECT;
+    static constexpr auto STORE_ALGO =
+        (THREADS >= BNB_WARP_SIZE) ? bnb_cub::BLOCK_STORE_WARP_TRANSPOSE : bnb_cub::BLOCK_STORE_DIRECT;
 
     typedef bnb_cub::BlockLoad<T, THREADS, NUM_PER_TH, LOAD_ALGO> LoadT;
-    typedef bnb_cub::BlockStore<
-        unsigned char, THREADS, (DATA_TYPE > 0) ? NUM_PER_TH / 2 : NUM_PER_TH,
-        STORE_ALGO>
+    typedef bnb_cub::BlockStore<unsigned char, THREADS, (DATA_TYPE > 0) ? NUM_PER_TH / 2 : NUM_PER_TH, STORE_ALGO>
         StoreChar;
     typedef bnb_cub::BlockReduce<float, THREADS> BlockReduce;
-    typedef bnb_cub::BlockLoad<float, THREADS, NUM_PER_TH, LOAD_ALGO>
-        LoadFloat;
+    typedef bnb_cub::BlockLoad<float, THREADS, NUM_PER_TH, LOAD_ALGO> LoadFloat;
 
     __shared__ typename LoadT::TempStorage loadt;
     __shared__ typename LoadFloat::TempStorage loadf;

csrc/ops.cuh

@@ -13,9 +13,9 @@
 #include <stdio.h>
 #include <vector>
 
+#include "common.cuh"
 #include "compat.cuh"
 #include <common.h>
-#include "common.cuh"
 
 // ============================================================================
 // Error checking helpers