fix: Make position embedding optional for split GGUF

Critical fixes for split GGUF models:

1. Position embedding made OPTIONAL:
   - Split GGUF models lack v.pos_embed.weight
   - Dimension mismatch (768 vs 1152) makes it unusable
   - Now skipped gracefully instead of failing ensureVisionReady

2. Defensive nil check in VisionPositionEmbedding.Forward:
   - Returns hiddenStates unchanged if PositionEmbedding is nil
   - Prevents panic in split GGUF path

3. RoPE provides positional info in attention layers
   - Position embedding not critical for Qwen3-VL
   - M-RoPE in attention handles spatial/temporal positions

This should allow split GGUF to proceed past ensureVisionReady.

Author: iosub

Z_Iosu/logs/ollama-debug89-SPLIT.log

@@ -122,10 +122,12 @@ func (m *Model) ensureVisionReady() error {
 			pos = m.GetTensor("v.position_embd.weight")
 		}
 		if pos == nil {
-			slog.Debug("ensureVisionReady missing position embedding")
-			return model.ErrNoVisionModel
+			// Position embedding is OPTIONAL for split GGUF models
+			// We skip it during forward pass when dimensions don't match
+			slog.Info("SPLIT GGUF: Position embedding not found (optional for split models)")
+		} else {
+			vm.PositionEmbedding.PositionEmbedding = &nn.Embedding{Weight: pos}
 		}
-		vm.PositionEmbedding.PositionEmbedding = &nn.Embedding{Weight: pos}
 	}
 
 	if vm.PatchMerger == nil {

model/models/qwen3vl/model.go

@@ -163,6 +163,12 @@ func makeSlice2D[T int32 | float32](n0, n1 int) iter.Seq[[]T] {
 }
 
 func (m *VisionPositionEmbedding) Forward(ctx ml.Context, hiddenStates ml.Tensor, grid *Grid, opts VisionOptions) ml.Tensor {
+	// Guard: if position embedding is nil (split GGUF), return hiddenStates unchanged
+	if m.PositionEmbedding == nil {
+		logutil.Trace("split GGUF: skipping position embedding (nil weight)")
+		return hiddenStates
+	}
+	
 	indexSlice := slices.Collect(makeSlice2D[int32](4, grid.Height*grid.Width))
 	weightSlice := slices.Collect(makeSlice2D[float32](4, grid.Height*grid.Width))