mlx: drop the empty-handle checks outside the bindings

Models and layers checked optional weights for nil and also for a handle
that no longer refers to an array, and evaluation and weight collection
skipped such handles. No path produces one: a missing tensor is nil, and a
handle only loses its array when its scope frees it, after which using it
is a bug. The nil checks stay; the validity check is internal to the
bindings now.

x/mlxrunner/mlx/array.go

@@ -119,7 +119,9 @@ func (t *Array) Clone() *Array {
 
 // misc. utilities
 
-func (t *Array) Valid() bool {
+// valid reports whether t still refers to an array: false once its scope
+// freed it.
+func (t *Array) valid() bool {
 	return t.ctx.ctx != nil
 }
 
@@ -139,7 +141,7 @@ func (t *Array) LogValue() slog.Value {
 	attrs := []slog.Attr{
 		slog.String("name", t.name),
 	}
-	if t.Valid() {
+	if t.valid() {
 		attrs = append(attrs,
 			slog.Any("dtype", t.DType()),
 			slog.Any("shape", t.Dims()),

x/mlxrunner/mlx/mlx.go

@@ -133,7 +133,7 @@ func doEval(outputs []*Array, async bool) {
 	defer freeVectorArray(vector)
 
 	for _, output := range outputs {
-		if output != nil && output.Valid() {
+		if output != nil {
 			mlxCheck(C.mlx_vector_array_append_value(vector, output.ctx))
 		}
 	}

x/mlxrunner/mlx/nn.go

@@ -8,7 +8,7 @@ type Linear struct {
 // Forward computes the linear transformation: x @ Weight.T + Bias
 func (m *Linear) Forward(x *Array) *Array {
 	w := m.Weight.Transpose(1, 0)
-	if m.Bias.Valid() {
+	if m.Bias != nil {
 		return m.Bias.Addmm(x, w, 1.0, 1.0)
 	}
 

x/mlxrunner/mlx/ops_extra.go

@@ -153,7 +153,7 @@ func Conv1d(x, weight *Array, bias *Array, stride, padding, dilation, groups int
 		C.int(groups),
 		DefaultStream().ctx,
 	))
-	if bias != nil && bias.Valid() {
+	if bias != nil {
 		out = Add(out, bias)
 	}
 	return out
@@ -646,7 +646,7 @@ func collect(v reflect.Value, arrays *[]*Array, seen map[uintptr]bool) {
 		seen[ptr] = true
 
 		if arr, ok := v.Interface().(*Array); ok {
-			if arr != nil && arr.Valid() {
+			if arr != nil {
 				*arrays = append(*arrays, arr)
 			}
 			return
@@ -659,7 +659,7 @@ func collect(v reflect.Value, arrays *[]*Array, seen map[uintptr]bool) {
 	case reflect.Struct:
 		// Check if this struct IS an Array (not a pointer to one)
 		if arr, ok := v.Addr().Interface().(*Array); ok {
-			if arr != nil && arr.Valid() {
+			if arr != nil {
 				*arrays = append(*arrays, arr)
 			}
 			return

x/mlxrunner/mlx/scope.go

@@ -177,7 +177,7 @@ func (s *Scope) take(t *Array) {
 		if from == s {
 			return
 		}
-		if !t.Valid() {
+		if !t.valid() {
 			panic(fmt.Sprintf("mlx: array %q used after its scope ended", t.name))
 		}
 		if from.noEscape {

x/mlxrunner/mlx/scope_test.go

@@ -22,17 +22,17 @@ func TestScopeFreesWhatIsNotReturned(t *testing.T) {
 				return []*Array{FromValue(3), nil, kept}
 			})
 			returned = out[0]
-			if !returned.Valid() {
+			if !returned.valid() {
 				t.Fatal("returned array was freed with the scope that created it")
 			}
-			if dropped.Valid() {
+			if dropped.valid() {
 				t.Fatal("array not returned survived its scope")
 			}
 		})
-		if returned.Valid() {
+		if returned.valid() {
 			t.Fatal("returned array survived the scope it was returned into")
 		}
-		if !kept.Valid() {
+		if !kept.valid() {
 			t.Fatal("returning a held array moved it out of its scope")
 		}
 	})
@@ -48,10 +48,10 @@ func TestScopedEvalEvaluatesAfterBuild(t *testing.T) {
 				tmp = FromValue(2)
 				return []*Array{FromValue(1).Add(tmp)}
 			})
-			if tmp.Valid() {
+			if tmp.valid() {
 				t.Fatal("intermediate survived the build scope")
 			}
-			if !out[0].Valid() || out[0].Int() != 3 {
+			if !out[0].valid() || out[0].Int() != 3 {
 				t.Fatal("returned array was not evaluated after the build scope")
 			}
 		})
@@ -72,11 +72,11 @@ func TestHeldScope(t *testing.T) {
 			kept, discarded, detached = FromValue(1), FromValue(2), FromValue(3)
 			held.Attach(kept, discarded, detached)
 			held.Discard(discarded)
-			if discarded.Valid() {
+			if discarded.valid() {
 				t.Fatal("discarded array survived")
 			}
 			Scoped(func() { held.Detach(detached) })
-			if detached.Valid() {
+			if detached.valid() {
 				t.Fatal("detached array survived the scope it was detached into")
 			}
 			if !panics(func() { other.Discard(kept) }) {
@@ -92,11 +92,11 @@ func TestHeldScope(t *testing.T) {
 				t.Fatal("no panic holding an array twice")
 			}
 		})
-		if !kept.Valid() {
+		if !kept.valid() {
 			t.Fatal("held array was freed with the scope that created it")
 		}
 		held.Close()
-		if kept.Valid() {
+		if kept.valid() {
 			t.Fatal("held array survived its scope's close")
 		}
 	})
@@ -115,7 +115,7 @@ func TestScopeEndsOnPanic(t *testing.T) {
 				panic("build failed")
 			})
 		}()
-		if a.Valid() {
+		if a.valid() {
 			t.Fatal("array survived the scope that panicked")
 		}
 		if currentScope != start {

x/models/cohere2_moe/cohere2_moe.go

@@ -398,7 +398,7 @@ func supportsGatherQMM(mode string, bits int) bool {
 // weights to the [E, in, out] layout GatherMM consumes, materialized once at
 // load so the forward path avoids per-call transposes.
 func transposeExpertWeightForGatherMM(w *mlx.Array) *mlx.Array {
-	if w == nil || !w.Valid() || w.NumDims() != 3 {
+	if w == nil || w.NumDims() != 3 {
 		return w
 	}
 	return mlx.Transpose(w, 0, 2, 1).Clone()

x/models/gemma4/gemma4.go

@@ -186,7 +186,7 @@ func sliceAxis1(a *mlx.Array, start, stop int32) *mlx.Array {
 // transposeForGatherMM transposes stacked expert weights from [experts, out, in]
 // to [experts, in, out] for use with GatherMM (which computes a @ b[group]).
 func transposeForGatherMM(w *mlx.Array) *mlx.Array {
-	if w == nil || !w.Valid() || w.NumDims() != 3 {
+	if w == nil || w.NumDims() != 3 {
 		return w
 	}
 	return mlx.Transpose(w, 0, 2, 1).Clone()

x/models/laguna/laguna.go

@@ -594,14 +594,14 @@ func stackAndClone(parts []*mlx.Array) *mlx.Array {
 }
 
 func transposeExpertWeightForGatherMM(w *mlx.Array) *mlx.Array {
-	if w == nil || !w.Valid() || w.NumDims() != 3 {
+	if w == nil || w.NumDims() != 3 {
 		return w
 	}
 	return mlx.Transpose(w, 0, 2, 1).Clone()
 }
 
 func transposeExpertWeightViewForGatherMM(w *mlx.Array) *mlx.Array {
-	if w == nil || !w.Valid() || w.NumDims() != 3 {
+	if w == nil || w.NumDims() != 3 {
 		return w
 	}
 	return mlx.Transpose(w, 0, 2, 1)
@@ -637,7 +637,7 @@ func denseExpertWeightForGatherMM(w *stackedExpertWeights) *mlx.Array {
 }
 
 func denseExpertWeightSupportsSourceLayout(w *stackedExpertWeights) bool {
-	return w != nil && w.Weight != nil && w.Weight.Valid() && w.Scales == nil && w.Weight.DType() == mlx.DTypeBFloat16
+	return w != nil && w.Weight != nil && w.Scales == nil && w.Weight.DType() == mlx.DTypeBFloat16
 }
 
 func denseExpertWeightsSupportSourceLayout(weights ...*stackedExpertWeights) bool {
@@ -755,7 +755,7 @@ func splitLastDim(x *mlx.Array, first int32) (*mlx.Array, *mlx.Array) {
 }
 
 func fuseExpertStacks(a, b *mlx.Array, axis int) *mlx.Array {
-	if a == nil || !a.Valid() || b == nil || !b.Valid() {
+	if a == nil || b == nil {
 		return nil
 	}
 	return mlx.Concatenate([]*mlx.Array{a, b}, axis).Clone()

x/models/nemotron_h/nemotron_h.go

@@ -422,7 +422,7 @@ func applyExpertWeightGlobalScale(weight, scale *mlx.Array) *mlx.Array {
 }
 
 func transposeExpertWeightForGatherMM(w *mlx.Array) *mlx.Array {
-	if w == nil || !w.Valid() || w.NumDims() != 3 {
+	if w == nil || w.NumDims() != 3 {
 		return w
 	}
 	return mlx.Transpose(w, 0, 2, 1).Clone()

x/models/nn/nn.go

@@ -61,7 +61,7 @@ type Linear struct {
 }
 
 func NewLinear(weight *mlx.Array, bias *mlx.Array) *Linear {
-	if bias != nil && bias.Valid() && bias.DType() != weight.DType() {
+	if bias != nil && bias.DType() != weight.DType() {
 		bias = bias.AsType(weight.DType())
 	}
 	return &Linear{Weight: weight, Bias: bias}
@@ -69,7 +69,7 @@ func NewLinear(weight *mlx.Array, bias *mlx.Array) *Linear {
 
 func (l *Linear) Forward(x *mlx.Array) *mlx.Array {
 	w := l.Weight.Transpose(1, 0)
-	if l.Bias != nil && l.Bias.Valid() {
+	if l.Bias != nil {
 		return l.Bias.Addmm(x, w, 1.0, 1.0)
 	}
 	return x.Matmul(w)
@@ -98,7 +98,7 @@ func NewQuantizedLinear(weight *mlx.Array, bias *mlx.Array, groupSize, bits int,
 	} else {
 		mlx.Eval(qw, scales)
 	}
-	if bias != nil && bias.Valid() && bias.DType() != weight.DType() {
+	if bias != nil && bias.DType() != weight.DType() {
 		bias = bias.AsType(weight.DType())
 	}
 	return &QuantizedLinear{
@@ -130,7 +130,7 @@ func (ql *QuantizedLinear) Forward(x *mlx.Array) *mlx.Array {
 		// coverage for this path.
 		out = quantizedLinearOutputScale(out, ql.GlobalScale)
 	}
-	if ql.Bias != nil && ql.Bias.Valid() {
+	if ql.Bias != nil {
 		bias := ql.Bias
 		if bias.DType() != out.DType() {
 			bias = bias.AsType(out.DType())
@@ -194,7 +194,7 @@ func (qe *QuantizedEmbedding) Forward(indices *mlx.Array) *mlx.Array {
 	weight := qe.Weight.TakeAxis(indices, 0)
 	scales := qe.Scales.TakeAxis(indices, 0)
 	var qbiases *mlx.Array
-	if qe.QBiases != nil && qe.QBiases.Valid() {
+	if qe.QBiases != nil {
 		qbiases = qe.QBiases.TakeAxis(indices, 0)
 	}
 	return mlx.Dequantize(weight, scales, qbiases, qe.GroupSize, qe.Bits, qe.Mode, qe.GlobalScale)

x/models/qwen3_5/qwen3_5.go

@@ -483,14 +483,14 @@ func stackAndClone(parts []*mlx.Array) *mlx.Array {
 }
 
 func transposeExpertWeightForGatherMM(w *mlx.Array) *mlx.Array {
-	if w == nil || !w.Valid() || w.NumDims() != 3 {
+	if w == nil || w.NumDims() != 3 {
 		return w
 	}
 	return mlx.Transpose(w, 0, 2, 1).Clone()
 }
 
 func fuseExpertStacks(a, b *mlx.Array, axis int) *mlx.Array {
-	if a == nil || !a.Valid() || b == nil || !b.Valid() {
+	if a == nil || b == nil {
 		return nil
 	}
 	return mlx.Concatenate([]*mlx.Array{a, b}, axis).Clone()