Skip to content

MetalPerformanceShadersGraph macOS xcode27.0 b2

Jump to bottom
Alex Soto edited this page Jun 23, 2026 · 1 revision

#MetalPerformanceShadersGraph.framework

diff -ruN /Applications/Xcode_27.0.0-beta.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShadersGraph.framework/Headers/MPSGraph.h /Applications/Xcode_27.0.0-beta2.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShadersGraph.framework/Headers/MPSGraph.h
--- /Applications/Xcode_27.0.0-beta.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShadersGraph.framework/Headers/MPSGraph.h	2026-05-21 09:32:50
+++ /Applications/Xcode_27.0.0-beta2.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShadersGraph.framework/Headers/MPSGraph.h	2026-06-16 03:43:48
@@ -118,8 +118,7 @@
 -(void) disableTypeInference;
 
 /// Turns on Automatic Layout Conversion (for conv like operations) for GPU.
-/// DEPRECATED: Layout conversion is now default, so this function is a no-op.
--(void) convertLayoutToNHWC MPS_AVAILABLE_STARTING(macos(26.4), ios(26.4), tvos(26.4));
+-(void) convertLayoutToNHWC MPS_AVAILABLE_STARTING_BUT_DEPRECATED("Layout Conversion to NHWC is enabled by default on M5 and newer", macos(26.4, 27.0), ios(26.4, 27.0), tvos(26.4, 27.0));
 /// Turns off Automatic Layout Conversion (for conv like operations) for GPU.
 -(void) disableAutoLayoutConversion MPS_AVAILABLE_STARTING(macos(27.0), ios(27.0), tvos(27.0));
 
diff -ruN /Applications/Xcode_27.0.0-beta.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShadersGraph.framework/Headers/MPSGraphMatrixMultiplicationOps.h /Applications/Xcode_27.0.0-beta2.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShadersGraph.framework/Headers/MPSGraphMatrixMultiplicationOps.h
--- /Applications/Xcode_27.0.0-beta.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShadersGraph.framework/Headers/MPSGraphMatrixMultiplicationOps.h	2026-05-23 03:56:39
+++ /Applications/Xcode_27.0.0-beta2.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShadersGraph.framework/Headers/MPSGraphMatrixMultiplicationOps.h	2026-06-13 22:36:00
@@ -18,7 +18,7 @@
 /// Use this descriptor with
 /// ``MPSGraph/scaledDotProductAttentionWithQueryTensor:keyTensor:valueTensor:descriptor:name:``
 /// to specify optional features such as an attention mask, causal masking, and attention sinks.
-MPS_CLASS_AVAILABLE_STARTING(macos(26.0), ios(26.0), macCatalyst(26.0), tvos(26.0))
+MPS_CLASS_AVAILABLE_STARTING(macos(27.0), ios(27.0), macCatalyst(27.0), tvos(27.0))
 @interface MPSGraphSDPADescriptor : MPSGraphObject
 
 /// The scale applied to the result of the query–key matrix multiply before softmax.
@@ -28,7 +28,7 @@
 /// An optional additive mask tensor applied to the scaled QK^T scores before softmax.
 /// Must be broadcast-compatible with shape ``[batch, heads, T_q, T_kv]``.
 /// Mutually exclusive with ``isCausal``.
-@property(nonatomic, nullable) MPSGraphTensor *maskTensor;
+@property(nonatomic, retain, nullable) MPSGraphTensor *maskTensor;
 
 /// When YES, a causal (lower-triangular) mask is applied so that each query position
 /// attends only to key positions at or before it. Mutually exclusive with ``maskTensor``.
@@ -37,7 +37,7 @@
 /// An optional attention-sinks tensor of shape ``[nHeads]``. Each element seeds the
 /// online-softmax accumulator for the corresponding query head with a virtual token logit,
 /// causing real-token attention weights to sum to less than one.
-@property(nonatomic, nullable) MPSGraphTensor *sinksTensor;
+@property(nonatomic, retain, nullable) MPSGraphTensor *sinksTensor;
 
 /// Creates a descriptor with the given scale and all other properties set to their defaults
 /// (no mask, isCausal = NO, no sinks).
@@ -139,7 +139,7 @@
                                                   descriptor:(MPSGraphSDPADescriptor *)descriptor
                                                         name:(NSString *_Nullable)name
     MPS_SWIFT_NAME(scaledDotProductAttention(query:key:value:descriptor:name:))
-        MPS_AVAILABLE_STARTING(macos(26.0), ios(26.0), macCatalyst(26.0), tvos(26.0));
+        MPS_AVAILABLE_STARTING(macos(27.0), ios(27.0), macCatalyst(27.0), tvos(27.0));
 @end
 
 
diff -ruN /Applications/Xcode_27.0.0-beta.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShadersGraph.framework/Headers/MPSGraphQuantizationOps.h /Applications/Xcode_27.0.0-beta2.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShadersGraph.framework/Headers/MPSGraphQuantizationOps.h
--- /Applications/Xcode_27.0.0-beta.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShadersGraph.framework/Headers/MPSGraphQuantizationOps.h	2026-05-23 03:56:38
+++ /Applications/Xcode_27.0.0-beta2.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/System/Library/Frameworks/MetalPerformanceShadersGraph.framework/Headers/MPSGraphQuantizationOps.h	2026-06-16 03:43:47
@@ -17,14 +17,16 @@
 
 /// Creates a Quantize operation and returns the result tensor.
 ///
-/// Convert the float `tensor` to an i8 or u8 tensor by applying a scale + bias transform: 
-/// result = (tensor / scale) + zeroPoint
+/// Convert the float `tensor` to a quantized tensor by applying a scale + bias transform:
+/// result = round(tensor / scale) + zeroPoint
 ///
 /// - Parameters:
 ///   - tensor: Input tensor to be quantized
 ///   - scale: Scale scalar parameter
 ///   - zeroPoint: Bias scalar parameter (converted to dataType of resultTensor)
-///   - dataType: Integer data type of the result tensor.
+///   - dataType: Data type of the result tensor. Supports ``MPSDataTypeInt8``, ``MPSDataTypeUInt8``,
+///     ``MPSDataTypeFloat8E4M3``, and ``MPSDataTypeFloat8E5M2``. Float8 output requires symmetric
+///     quantization (zeroPoint = 0).
 ///   - name: The name for the operation.
 /// - Returns: A valid MPSGraphTensor array of datatype dataType
 -(MPSGraphTensor*) quantizeTensor:(MPSGraphTensor*)tensor
@@ -37,9 +39,15 @@
 
 /// Creates Dequantize operation and returns the result tensor.
 ///
-/// Convert the i8 or u8 `tensor` to a float tensor by applying a scale + bias transform: 
-/// result = scale(tensor - zeroPoint)
+/// Convert the quantized `tensor` to a float tensor by applying a scale + bias transform:
+/// result = scale * (tensor - zeroPoint)
 ///
+/// Supported input types: ``MPSDataTypeInt4``, ``MPSDataTypeUInt4``,
+/// ``MPSDataTypeInt8``, ``MPSDataTypeUInt8``,
+/// ``MPSDataTypeFloat8E4M3``, ``MPSDataTypeFloat8E5M2``.
+/// The scale type and `dataType` must match (``MPSDataTypeFloat16``, ``MPSDataTypeFloat32``,
+/// or ``MPSDataTypeBFloat16``).
+///
 /// - Parameters:
 ///   - tensor: Input tensor to be dequantized
 ///   - scale: Scale scalar parameter
@@ -57,14 +65,16 @@
 
 /// Creates a Quantize operation and returns the result tensor.
 ///
-/// Convert the float `tensor` to an i8 or u8 tensor by applying a scale + bias transform: 
-/// result = (tensor / scaleTensor) + zeroPoint
+/// Convert the float `tensor` to a quantized tensor by applying a scale + bias transform:
+/// result = round(tensor / scaleTensor) + zeroPoint
 ///
 /// - Parameters:
 ///   - tensor: Input tensor to be quantized
 ///   - scaleTensor: Scale 1D Tensor parameter with size == tensor.shape[axis]
 ///   - zeroPoint: Bias scalar parameter (converted to dataType of resultTensor)
-///   - dataType: Integer data type of the result tensor.
+///   - dataType: Data type of the result tensor. Supports ``MPSDataTypeInt8``, ``MPSDataTypeUInt8``,
+///     ``MPSDataTypeFloat8E4M3``, and ``MPSDataTypeFloat8E5M2``. Float8 output requires symmetric
+///     quantization (zeroPoint = 0).
 ///   - axis: Axis on which the scale 1D value is being broadcasted
 ///   - name: The name for the operation.
 /// - Returns: A valid MPSGraphTensor array of datatype dataType
@@ -79,9 +89,15 @@
 
 /// Creates Dequantize operation and returns the result tensor.
 ///
-/// Convert the i8 or u8 `tensor` to a float tensor by applying a scale + bias transform: 
-/// result = scaleTensor(tensor - zeroPoint)
+/// Convert the quantized `tensor` to a float tensor by applying a scale + bias transform:
+/// result = scaleTensor * (tensor - zeroPoint)
 ///
+/// Supported input types: ``MPSDataTypeInt4``, ``MPSDataTypeUInt4``,
+/// ``MPSDataTypeInt8``, ``MPSDataTypeUInt8``,
+/// ``MPSDataTypeFloat8E4M3``, ``MPSDataTypeFloat8E5M2``.
+/// The `scaleTensor` type and `dataType` must match (``MPSDataTypeFloat16``,
+/// ``MPSDataTypeFloat32``, or ``MPSDataTypeBFloat16``).
+///
 /// - Parameters:
 ///   - tensor: Input tensor to be dequantized
 ///   - scaleTensor: Scale scalar or 1D Tensor parameter with size == tensor.shape[axis]
@@ -101,14 +117,16 @@
 
 /// Creates a Quantize operation and returns the result tensor.
 ///
-/// Convert the float `tensor` to an i8 or u8 tensor by applying a scale + bias transform: 
-/// result = (tensor / scaleTensor) + zeroPointTensor
+/// Convert the float `tensor` to a quantized tensor by applying a scale + bias transform:
+/// result = round(tensor / scaleTensor) + zeroPointTensor
 ///
 /// - Parameters:
 ///   - tensor: Input tensor to be quantized
 ///   - scaleTensor: Scale scalar or 1D Tensor parameter with size == tensor.shape[axis]
 ///   - zeroPointTensor: Bias scalar or 1D Tensor parameter with size == tensor.shape[axis]
-///   - dataType: Integer data type of the result tensor.
+///   - dataType: Data type of the result tensor. Supports ``MPSDataTypeInt8``, ``MPSDataTypeUInt8``,
+///     ``MPSDataTypeFloat8E4M3``, and ``MPSDataTypeFloat8E5M2``. Float8 output requires symmetric
+///     quantization (zeroPointTensor = 0).
 ///   - axis: Axis on which the scale 1D value is being broadcasted
 ///   - name: The name for the operation.
 /// - Returns: A valid MPSGraphTensor array of datatype dataType
@@ -123,9 +141,15 @@
 
 /// Creates a dequantize operation and returns the result tensor.
 ///
-/// Convert the i8 or u8 `tensor` to a float tensor by applying a scale + bias transform: 
-/// result = scaleTensor(tensor - zeroPointTensor)
+/// Convert the quantized `tensor` to a float tensor by applying a scale + bias transform:
+/// result = scaleTensor * (tensor - zeroPointTensor)
 ///
+/// Supported input types: ``MPSDataTypeInt4``, ``MPSDataTypeUInt4``,
+/// ``MPSDataTypeInt8``, ``MPSDataTypeUInt8``,
+/// ``MPSDataTypeFloat8E4M3``, ``MPSDataTypeFloat8E5M2``.
+/// The `scaleTensor` type and `dataType` must match (``MPSDataTypeFloat16``,
+/// ``MPSDataTypeFloat32``, or ``MPSDataTypeBFloat16``).
+///
 /// - Parameters:
 ///   - tensor: Input tensor to be dequantized
 ///   - scaleTensor: Scale scalar or 1D Tensor parameter with size == tensor.shape[axis]
@@ -143,14 +167,31 @@
 MPS_AVAILABLE_STARTING(macos(13.1), ios(16.2), tvos(16.2))
 MPS_SWIFT_NAME( dequantize(_:scaleTensor:zeroPointTensor:dataType:axis:name:) );
 
-
 /// Creates a dequantize operation and returns the result tensor.
 ///
-/// Convert the i8, u8, i4 or u4 `tensor` to a float tensor by applying a scale and bias transform:
+/// Convert the quantized `tensor` to a float tensor by applying a scale and bias transform:
 /// ```md
-/// result = scaleTensor(tensor - zeroPointTensor).
+/// result = scaleTensor * (tensor - zeroPointTensor)
 /// ```
 ///
+/// The quantization mode is determined by the element type of `scaleTensor`:
+///
+/// **Regular blockwise** (`scaleTensor` type is an MPSGraph float type):
+/// - Supported input types: ``MPSDataTypeInt4``, ``MPSDataTypeUInt4``,
+///   ``MPSDataTypeInt8``, ``MPSDataTypeUInt8``,
+///   ``MPSDataTypeFloat8E4M3``, ``MPSDataTypeFloat8E5M2``.
+/// - `zeroPointTensor` type must match the input `tensor` type.
+/// - `scaleTensor` shape: leading dims of `tensor` with the last dim replaced by
+///   `tensor.shape[last] / blockSize`.
+///
+/// **MX mode** (`scaleTensor` type is ``MPSDataTypeFloat8E8M0``):
+/// - Supported input types: ``MPSDataTypeFloat4E2M1``, ``MPSDataTypeFloat8E4M3``, ``MPSDataTypeFloat8E5M2``.
+/// - Block size 32, symmetric (zero point must be 0).
+///
+/// **F4 with F8E4M3 scale** (`scaleTensor` type is ``MPSDataTypeFloat8E4M3``,
+/// input type is ``MPSDataTypeFloat4E2M1``):
+/// - Block size 16, symmetric (zero point must be 0).
+///
 /// - Parameters:
 ///   - tensor: Input tensor to be dequantized.
 ///   - scaleTensor: The scale tensor with groups support.
@@ -166,14 +207,31 @@
 MPS_AVAILABLE_STARTING(macos(15.0), ios(18.0), macCatalyst(18.0), tvos(18.0))
 MPS_SWIFT_NAME( dequantize(_:scaleTensor:zeroPointTensor:dataType:name:) );
 
-
 /// Creates a dequantize operation and returns the result tensor.
 ///
-/// Converts the i8, u8, i4 or u4 `tensor` to a float tensor by applying a scale and bias transform:
+/// Converts the quantized `tensor` to a float tensor by applying a scale transform
+/// (symmetric — zero point is implicitly 0):
 /// ```md
-/// result = scaleTensor * tensor.
+/// result = scaleTensor * tensor
 /// ```
 ///
+/// The quantization mode is determined by the element type of `scaleTensor`:
+///
+/// **Regular blockwise** (`scaleTensor` type is an MPSGraph float type):
+/// - Supported input types: ``MPSDataTypeInt4``, ``MPSDataTypeUInt4``,
+///   ``MPSDataTypeInt8``, ``MPSDataTypeUInt8``,
+///   ``MPSDataTypeFloat8E4M3``, ``MPSDataTypeFloat8E5M2``.
+/// - `scaleTensor` shape: leading dims of `tensor` with the last dim replaced by
+///   `tensor.shape[last] / blockSize`.
+///
+/// **MX mode** (`scaleTensor` type is ``MPSDataTypeFloat8E8M0``):
+/// - Supported input types: ``MPSDataTypeFloat4E2M1``, ``MPSDataTypeFloat8E4M3``, ``MPSDataTypeFloat8E5M2``.
+/// - Block size 32.
+///
+/// **F4 with F8E4M3 scale** (`scaleTensor` type is ``MPSDataTypeFloat8E4M3``,
+/// input type is ``MPSDataTypeFloat4E2M1``):
+/// - Block size 16.
+///
 /// - Parameters:
 ///   - tensor: Input tensor to be dequantized.
 ///   - scaleTensor: Scale Tensor parameter with groups support.
@@ -187,17 +245,25 @@
 MPS_AVAILABLE_STARTING(macos(15.0), ios(18.0), macCatalyst(18.0), tvos(18.0))
 MPS_SWIFT_NAME( dequantize(_:scaleTensor:dataType:name:) );
 
-/// Creates a lookup-table based quantization operation and returns the result tensor.
+/// Creates a lookup-table based dequantize operation and returns the result tensor.
 ///
-/// Converts a u8 or u4 `tensor` to a float tensor by applying a lookup operation:
+/// Converts a `tensor` of integer indices to a float tensor by applying a lookup operation:
 /// ```md
 /// result[i1,...,in] = LUTTensor[i1',...,in',tensor[i1,...,in]].
 /// ```
-/// Note: The operation supports LUT groups up to the last 3 dimensions for `tensor`.
 ///
+/// Supported `tensor` index types and required last-dimension size of `LUTTensor`:
+/// - ``MPSDataTypeUInt4``: 16 entries
+/// - ``MPSDataTypeUInt8``: 256 entries
+///
+/// `LUTTensor` (and result) element types: ``MPSDataTypeFloat16``, ``MPSDataTypeFloat32``,
+/// ``MPSDataTypeBFloat16``, ``MPSDataTypeFloat8E4M3``, ``MPSDataTypeFloat8E5M2``, ``MPSDataTypeInt8``.
+/// The input `tensor` must be a graph constant.
+/// The operation supports LUT groups up to the last 3 dimensions for `tensor`.
+///
 /// - Parameters:
-///   - tensor: Input tensor to be dequantized.
-///   - LUTTensor: The lookup table to use - for u4 the last dimension should have 16 elements, and for u8 256 elements.
+///   - tensor: Input constant integer-index tensor to be dequantized.
+///   - LUTTensor: The lookup table to use.
 ///   - name: The name for the operation.
 /// - Returns: A valid ``MPSGraphTensor`` object.
 -(MPSGraphTensor *) dequantizeTensor:(MPSGraphTensor *) tensor
@@ -206,20 +272,29 @@
 MPS_SWIFT_NAME( dequantize(_:LUTTensor:name:) )
 MPS_AVAILABLE_STARTING(macos(15.0), ios(18.0), macCatalyst(18.0), tvos(18.0));
 
-/// Creates a vector lookup-table based quantization operation and returns the result tensor.
+/// Creates a vector lookup-table based dequantize operation and returns the result tensor.
 ///
-/// Converts a u8 or u4 `tensor` to a float tensor by applying a lookup operation, where each
-/// input index defines a vector of values. The operation reads the vector values from the last dimension of the lookup table
-/// tensor and stores them into the dimension defined by `axis` on the result tensor.
+/// Converts a `tensor` of integer indices to a float tensor by applying a lookup operation,
+/// where each input index defines a vector of values. The operation reads the vector values
+/// from the last dimension of the lookup table tensor and stores them into the dimension
+/// defined by `axis` on the result tensor.
 /// ```md
 /// result[i1, ... , i_axis, ..., in] = LUTTensor[i1', ..., in', tensor[i1, ..., in], i_axis]
 /// ```
-/// Note: The operation supports LUT groups up to the last 2 dimensions for `tensor`.
 ///
+/// Supported `tensor` index types and required second-to-last-dimension size of `LUTTensor`:
+/// - ``MPSDataTypeUInt4``: 16 entries
+/// - ``MPSDataTypeUInt8``: 256 entries
+///
+/// `LUTTensor` (and result) element types: ``MPSDataTypeFloat16``, ``MPSDataTypeFloat32``,
+/// ``MPSDataTypeBFloat16``, ``MPSDataTypeFloat8E4M3``, ``MPSDataTypeFloat8E5M2``, ``MPSDataTypeInt8``.
+/// The input `tensor` must be a graph constant.
+/// The operation supports LUT groups up to the last 2 dimensions for `tensor`.
+///
 /// - Parameters:
-///   - tensor: Input tensor to be dequantized.
-///   - LUTTensor: The lookup table to use - for u4 the second to last dimension should have 16 elements, and for u8 256 elements.
-///   - axis: Axis on which the scale 1D value is being broadcasted.
+///   - tensor: Input constant integer-index tensor to be dequantized.
+///   - LUTTensor: The lookup table to use.
+///   - axis: The result axis into which the LUT vector values are written.
 ///   - name: The name for the operation.
 /// - Returns: A valid ``MPSGraphTensor`` object.
 -(MPSGraphTensor *) dequantizeTensor:(MPSGraphTensor *) tensor

Contributing

Bindings

Tips & tricks

Clone this wiki locally