model/qwen3next: avoid crash in in DeltaNet when offloading (#14541)

Co-authored-by: Yossi Ovadia <jabadia@gmail.com>

model/models/qwen3next/deltanet.go

@@ -454,6 +454,10 @@ func (gdn *GatedDeltaNet) deltaNetChunked(
 	vT := v.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx, chunkSize, headVDim, nChunks, numVHeads*nSeqs)
 	stateT := state.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx, headVDim, headVDim, 1, numVHeads*nSeqs)
 
+	// Collect chunk outputs and concatenate at the end.
+	// Avoids SET on buffer-less intermediates under partial offload.
+	chunks := make([]ml.Tensor, nChunks)
+
 	for chunk := range nChunks {
 		qChunk := q.Slice(ctx, 2, chunk, chunk+1, 1)
 		vTChunk := vT.Slice(ctx, 2, chunk, chunk+1, 1)
@@ -475,14 +479,7 @@ func (gdn *GatedDeltaNet) deltaNetChunked(
 		vAttn := vTNewChunk.Mulmat(ctx, attnChunk)
 		coreAttnOutChunk := attnInter.Add(ctx, vAttn)
 
-		v = v.SetInplace(
-			ctx,
-			coreAttnOutChunk,
-			v.Stride(1),
-			v.Stride(2),
-			v.Stride(3),
-			chunk*v.Stride(2),
-		)
+		chunks[chunk] = coreAttnOutChunk
 
 		// Update state for next chunk
 		gExpLastChunk := gLastExp.Slice(ctx, 2, chunk, chunk+1, 1)
@@ -495,6 +492,20 @@ func (gdn *GatedDeltaNet) deltaNetChunked(
 		stateT = stateT.Add(ctx, kgdMulVNew)
 	}
 
+	// Use a balanced concat tree so concat work does not balloon on long prompts.
+	for len(chunks) > 1 {
+		merged := make([]ml.Tensor, 0, (len(chunks)+1)/2)
+		for i := 0; i < len(chunks); i += 2 {
+			if i+1 < len(chunks) {
+				merged = append(merged, chunks[i].Concat(ctx, chunks[i+1], 2))
+			} else {
+				merged = append(merged, chunks[i])
+			}
+		}
+		chunks = merged
+	}
+	v = chunks[0]
+
 	// Final reshape
 	coreAttnOut := v.Contiguous(ctx, headVDim, chunkSize*nChunks, numVHeads, nSeqs)