DeepQuant Quantized Linear Support

.github/workflows/CI.yml

@@ -75,6 +75,7 @@ jobs:
         testFloatMul
         Quant
         Dequant
+        QuantizedLinear
 
 
   generic-models:
@@ -270,6 +271,7 @@ jobs:
         testFloatMul
         Quant
         Dequant
+        QuantizedLinear
       num-cores: 8
 
   siracusa-models:

CHANGELOG.md

@@ -206,4 +206,11 @@ Change main.c to use OUTPUTTYPE instead of float
 - Custom `DequantPatternPass` class to replace matched patterns with a single `Dequant` operator
 - Parser implementation in `Parsers.py` to extract dequantization parameters
 - C template implementation in `DequantTemplate.py` for efficient dequantization
-- Type checker implementation in `TypeCheckers.py` to handle bit-width and signedness
+- Type checker implementation in `TypeCheckers.py` to handle bit-width and signedness
+
+## Implemented Updates for handling Quantized Linear DNN
+
+### Added
+- New `_sanitizeGraphNames` function to sanitize the names of the nodes and tensors of the graph
+- Implementation for both Generic and Siracusa targets in the Deeploy framework
+- Modified the binding of dequant in `Bindings.py` to handle int32 after GEMM operation

Deeploy/DeeployTypes.py

@@ -3129,6 +3129,18 @@ def _foldConstants(self, graph: gs.Graph):
         graph.fold_constants()
         graph.cleanup().toposort()
 
+    def _sanitizeGraphNames(self, graph: gs.Graph):
+
+        def sanitize(name: str) -> str:
+            # Remove illegal characters: anything not a letter, digit, or underscore.
+            sanitized_name = re.sub(r'[^a-zA-Z0-9_]', '', name)
+            return sanitized_name
+
+        for node in graph.nodes:
+            node.name = sanitize(node.name)
+            for tensor in node.inputs + node.outputs:
+                tensor.name = sanitize(tensor.name)
+
     # Don't override this
     # Duplicate constants with multiple users
     def _removeEmptyInputs(self, graph: gs.Graph):
@@ -3147,6 +3159,8 @@ def frontEnd(self):
         for idx, outputNode in enumerate(self.graph.outputs):
             outputNode.name = "output_" + str(idx)
 
+        self._sanitizeGraphNames(self.graph)
+
         self._removeEmptyInputs(self.graph)
 
         self._duplicateConstants(self.graph)

Deeploy/Targets/Generic/Bindings.py

@@ -271,4 +271,7 @@
 BasicDequantBindings = [
     NodeBinding(DequantChecker([PointerClass(int8_t)], [PointerClass(float32_t)]), DequantTemplate.referenceTemplate,
                 BasicTransformer),
+] + [
+    NodeBinding(DequantChecker([PointerClass(int32_t)], [PointerClass(float32_t)]), DequantTemplate.referenceTemplate,
+                BasicTransformer),
 ]

Deeploy/Targets/PULPOpen/Bindings.py

@@ -372,4 +372,7 @@
 BasicDequantBindings = [
     NodeBinding(DequantChecker([PointerClass(int8_t)], [PointerClass(float32_t)]), DequantTemplate.referenceTemplate,
                 ForkTransformer),
+] + [
+    NodeBinding(DequantChecker([PointerClass(int32_t)], [PointerClass(float32_t)]), DequantTemplate.referenceTemplate,
+                ForkTransformer),
 ]

Deeploy/Targets/PULPOpen/Platform.py

@@ -32,7 +32,8 @@
     NodeTemplate, StructBuffer, TopologyOptimizer, TransientBuffer, VariableBuffer
 from Deeploy.MemoryLevelExtension.MemoryLevels import MemoryHierarchy, MemoryLevel
 from Deeploy.MemoryLevelExtension.NetworkDeployers.MemoryLevelDeployer import MemoryPlatform, MemoryPlatformWrapper
-from Deeploy.Targets.Generic.Bindings import BasicPad1DBindings, BasicPad2DBindings, BasicRQIntegerDivBinding
+from Deeploy.Targets.Generic.Bindings import BasicGEMMBindings, BasicPad1DBindings, BasicPad2DBindings, \
+    BasicRQIntegerDivBinding
 from Deeploy.Targets.Generic.Layers import AddLayer, ConcatLayer, ConvLayer, DequantLayer, GatherLayer, GELULayer, \
     GEMMLayer, LayerNormLayer, MatMulLayer, MaxPoolLayer, MulLayer, PadLayer, QuantLayer, ReduceMeanLayer, ReluLayer, \
     RequantShiftLayer, ReshapeLayer, RQIntegerDivLayer, RQSiGELULayer, RQSiHardswishLayer, SliceLayer, SoftmaxLayer, \
@@ -110,12 +111,13 @@
 
 QuantMapper = NodeMapper(QuantParser(), BasicQuantBindings)
 DequantMapper = NodeMapper(DequantParser(), BasicDequantBindings)
+GEMMDequantMapper = NodeMapper(PULPGEMMParser(), BasicGEMMBindings)
 
 PULPMapping = {
     'Conv': ConvLayer([FPConv2DMapper]),
     'RequantizedConv': PULPRQSConvLayer([Conv2DMapper, DWConv2DMapper, Conv1DMapper, DWConv1DMapper]),
     'RequantizedGemm': PULPRQSGEMMLayer([MatrixVecMapper, TallGEMMMapper, GEMMMapper]),
-    'Gemm': GEMMLayer([FloatGEMMMapper]),
+    'Gemm': GEMMLayer([FloatGEMMMapper, GEMMDequantMapper]),
     'Gelu': GELULayer([GELUMapper]),
     'LayerNormalization': LayerNormLayer([LayerNormMapper]),
     'MaxPool': MaxPoolLayer([MaxPool2DMapper]),