sycl : fix for untransposed GDA recurrent state (ggml-org#20583)

CISC · web-flow · commit ebbf544ed100 · 2026-03-15T19:10:15.000+01:00
diff --git a/ggml/src/ggml-sycl/gated_delta_net.cpp b/ggml/src/ggml-sycl/gated_delta_net.cpp
@@ -55,7 +55,7 @@ void gated_delta_net_sycl(const float *     q,
 #pragma unroll
     for (int r = 0; r < rows_per_lane; r++) {
         const int i = r * warp_size + lane;
-        s_shard[r]  = curr_state[i * S_v + col];
+        s_shard[r]  = curr_state[col * S_v + i];
     }
 
     for (int t = 0; t < n_tokens; t++) {
@@ -137,7 +137,7 @@ void gated_delta_net_sycl(const float *     q,
 #pragma unroll
     for (int r = 0; r < rows_per_lane; r++) {
         const int i          = r * warp_size + lane;
-        state[i * S_v + col] = s_shard[r];
+        state[col * S_v + i] = s_shard[r];
     }
 }
 

Original file line number	Diff line number	Diff line change
`@@ -55,7 +55,7 @@ void gated_delta_net_sycl(const float * q,`
`55`	`55`	`#pragma unroll`
`56`	`56`	`for (int r = 0; r < rows_per_lane; r++) {`
`57`	`57`	`const int i = r * warp_size + lane;`
`58`		`- s_shard[r] = curr_state[i * S_v + col];`
	`58`	`+ s_shard[r] = curr_state[col * S_v + i];`
`59`	`59`	`}`
`60`	`60`
`61`	`61`	`for (int t = 0; t < n_tokens; t++) {`
`@@ -137,7 +137,7 @@ void gated_delta_net_sycl(const float * q,`
`137`	`137`	`#pragma unroll`
`138`	`138`	`for (int r = 0; r < rows_per_lane; r++) {`
`139`	`139`	`const int i = r * warp_size + lane;`
`140`		`- state[i * S_v + col] = s_shard[r];`
	`140`	`+ state[col * S_v + i] = s_shard[r];`
`141`	`141`	`}`
`142`	`142`	`}`
`143`	`143`