GatedDeltaNet: Move import

Author: turboderp

exllamav3/exllamav3_ext/bindings.cpp

@@ -14,6 +14,7 @@
 #include "softcap.cuh"
 #include "routing.cuh"
 #include "gdn.cuh"
+#include "causal_conv1d.cuh"
 
 #include "quant/quantize.cuh"
 #include "quant/pack.cuh"

exllamav3/exllamav3_ext/causal_conv1d.cu

@@ -0,0 +1,34 @@
+#include <cuda_fp16.h>
+#include <cuda_fp16.hpp>
+#include "activation.cuh"
+#include <c10/cuda/CUDAGuard.h>
+#include <ATen/cuda/CUDAContext.h>
+#include "util.h"
+#include "util.cuh"
+#include "compat.cuh"
+#include <cmath>
+
+//__global__ __launch_bounds__(MAX_HEAD_DIM)
+//void causal_conv1d_kernel
+//(
+//
+//)
+//{
+//
+//}
+//
+
+void causal_conv1d
+(
+
+
+
+
+
+
+)
+{
+
+
+}
+

exllamav3/exllamav3_ext/causal_conv1d.cuh

@@ -0,0 +1,4 @@
+void causal_conv1d
+(
+
+);

exllamav3/modules/gated_delta_net.py

@@ -79,6 +79,11 @@ def causal_conv1d_fwd_function_cu(
     causal_conv1d_update_function = causal_conv1d_update_function_torch
     causal_conv1d_fwd_function = causal_conv1d_fwd_function_torch
 
+try:
+    from fla.ops.gated_delta_rule import chunk_gated_delta_rule
+except ModuleNotFoundError:
+    chunk_gated_delta_rule = None
+
 """
 fla wrapper, reduce overhead by bypassing input_guard and torch custom ops stuff
 """
@@ -465,8 +470,6 @@ def forward(
         out_dtype: torch.dtype | None = None
     ) -> torch.Tensor:
 
-        from fla.ops.gated_delta_rule import chunk_gated_delta_rule
-
         bsz, seqlen, _ = x.shape
 
         # Previous state