issue/158/fix: 支持天数的特殊 BLOCK SIZE

Signed-off-by: YdrMaster <ydrml@hotmail.com>

src/infiniop/devices/cuda/cuda_kernel_common.cuh

@@ -6,6 +6,7 @@
 
 // Posible maximum number of threads per block for CUDA architectures
 // Used for picking correct kernel launch configuration
+#define CUDA_BLOCK_SIZE_4096 4096
 #define CUDA_BLOCK_SIZE_1024 1024
 #define CUDA_BLOCK_SIZE_512 512
 

src/infiniop/ops/causal_softmax/cuda/causal_softmax_cuda.cu

@@ -69,6 +69,10 @@ infiniStatus_t Descriptor::calculate(void *workspace, size_t workspace_size,
         CHECK_STATUS(launchKernel<CUDA_BLOCK_SIZE_512>(
             y, x, _info.dtype, _info.batch_size, _info.seq_len, _info.total_seq_len,
             _info.y_stride_b, _info.y_stride_i, _info.x_stride_b, _info.x_stride_i, stream));
+    } else if (_opaque->internal->maxThreadsPerBlock() == CUDA_BLOCK_SIZE_4096) {
+        CHECK_STATUS(launchKernel<CUDA_BLOCK_SIZE_4096>(
+            y, x, _info.dtype, _info.batch_size, _info.seq_len, _info.total_seq_len,
+            _info.y_stride_b, _info.y_stride_i, _info.x_stride_b, _info.x_stride_i, stream));
     } else {
         return INFINI_STATUS_DEVICE_ARCHITECTURE_NOT_SUPPORTED;
     }

src/infiniop/ops/rms_norm/cuda/rms_norm_cuda.cu

@@ -95,6 +95,8 @@ infiniStatus_t Descriptor::calculate(
         CHECK_STATUS(launchKernel<CUDA_BLOCK_SIZE_1024>(batch_size, dim, y, _info.atype, stride_y, x, stride_x, w, _info.wtype, _info.epsilon, cuda_stream));
     } else if (_opaque->internal->maxThreadsPerBlock() == CUDA_BLOCK_SIZE_512) {
         CHECK_STATUS(launchKernel<CUDA_BLOCK_SIZE_512>(batch_size, dim, y, _info.atype, stride_y, x, stride_x, w, _info.wtype, _info.epsilon, cuda_stream));
+    } else if (_opaque->internal->maxThreadsPerBlock() == CUDA_BLOCK_SIZE_4096) {
+        CHECK_STATUS(launchKernel<CUDA_BLOCK_SIZE_4096>(batch_size, dim, y, _info.atype, stride_y, x, stride_x, w, _info.wtype, _info.epsilon, cuda_stream));
     } else {
         return INFINI_STATUS_DEVICE_ARCHITECTURE_NOT_SUPPORTED;
     }