change sgl_kernel WARP_SIZE to 64

sgl-kernel/csrc/gemm/qserve_w4a8_per_chn_gemm.cu

@@ -25,7 +25,7 @@
 #define INTRIN_M 16
 #define INTRIN_N 16
 #define INTRIN_K 32
-#define WARP_SIZE 32
+#define WARP_SIZE 64
 #define SMEM_PAD_A 0
 #define SMEM_PAD_B 0
 #define PACK_SIZE 16

sgl-kernel/csrc/gemm/qserve_w4a8_per_group_gemm.cu

@@ -25,7 +25,7 @@
 #define INTRIN_M 16
 #define INTRIN_N 16
 #define INTRIN_K 32
-#define WARP_SIZE 32
+#define WARP_SIZE 64
 #define SMEM_PAD_A 0
 #define SMEM_PAD_B 0
 #define PACK_SIZE 16

sgl-kernel/csrc/kvcacheio/transfer.cu

@@ -5,7 +5,7 @@
 #include <cstdint>
 
 #ifndef USE_ROCM
-#define WARP_SIZE 32
+#define WARP_SIZE 64
 #include "pytorch_extension_utils.h"
 #else
 #include "pytorch_extension_utils_rocm.h"

sgl-kernel/csrc/quantization/gguf/ggml-common.h

@@ -3,7 +3,7 @@
 // copied from https://github.com/ggerganov/llama.cpp/blob/b2899/ggml-common.h
 #define QK_K 256
 #define K_QUANTS_PER_ITERATION 2
-#define WARP_SIZE_GGUF 32
+#define WARP_SIZE_GGUF 64
 #define K_SCALE_SIZE 12
 #define CUDA_DEQUANTIZE_BLOCK_SIZE 256
 #define CUDA_QUANTIZE_BLOCK_SIZE 256

sgl-kernel/include/utils.h

@@ -340,7 +340,7 @@ inline bool getEnvEnablePDL() {
 #define CEILDIV(x, y) (((x) + (y) - 1) / (y))
 
 #ifndef USE_ROCM
-#define WARP_SIZE 32
+#define WARP_SIZE 64
 #else
 #if defined(__GFX9__) || !defined(__HIP_DEVICE_COMPILE__)
 #define WARP_SIZE 64