recipes doc

docs/source/guides/1_quantization.rst

@@ -19,6 +19,8 @@ Below, you can find the documentation for the quantization toolkit in ModelOpt:
     ./_basic_quantization.rst
     ./_choosing_quant_methods.rst
     ./_pytorch_quantization.rst
+    ./_quant_cfg.rst
+    ./_recipes.rst
     ./_customized_model_quantization.rst
     ./_compress_quantized_models.rst
     ./_onnx_quantization.rst

docs/source/guides/_pytorch_quantization.rst

@@ -237,14 +237,16 @@ For debugging purposes or simple customizations, you can modify an existing conf
 
 .. code-block:: python
 
-    # Create a copy of the default INT8 configuration
-    config = mtq.INT8_DEFAULT_CFG.copy()
+    import copy
 
-    # Disable input quantizers for all layers
-    config["quant_cfg"]["*input_quantizer"]["enable"] = False
+    # Create a deep copy of the default INT8 configuration
+    config = copy.deepcopy(mtq.INT8_DEFAULT_CFG)
+
+    # Disable input quantizers for all layers (appended last, so it takes precedence)
+    config["quant_cfg"].append({"quantizer_path": "*input_quantizer", "enable": False})
 
     # Disable all quantizers for layers matching the pattern "layer1.*"
-    config["quant_cfg"]["*layer1.*"] = {"enable": False}
+    config["quant_cfg"].append({"quantizer_path": "*layer1.*", "enable": False})
 
 Advanced Configuration Creation
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
@@ -253,18 +255,23 @@ For exploring new quantization recipes, you can compose a completely new configu
 
 .. code-block:: python
 
+    from modelopt.torch.quantization.config import _default_disabled_quantizer_cfg
+
     # Custom configuration for INT4 block-wise weights and INT8 dynamic activations
     MY_CUSTOM_CONFIG = {
-        "quant_cfg": {
+        "quant_cfg": [
+            # Disable all quantizers by default, then enable selectively
+            {"quantizer_path": "*", "enable": False},
+
             # Configure weight quantizers with 4-bit precision and 128-element blocks
-            "*weight_quantizer": {"num_bits": 4, "block_sizes": {-1: 128}, "enable": True},
+            {"quantizer_path": "*weight_quantizer", "cfg": {"num_bits": 4, "block_sizes": {-1: 128}}, "enable": True},
 
             # Configure input quantizers with 8-bit dynamic quantization
-            "*input_quantizer": {"num_bits": 8, "type": "dynamic", "block_sizes": {-1: None}},
+            {"quantizer_path": "*input_quantizer", "cfg": {"num_bits": 8, "type": "dynamic", "block_sizes": {-1: None}}},
 
             # Include default disabled quantizer configurations
-            **_default_disabled_quantizer_cfg,
-        },
+            *_default_disabled_quantizer_cfg,
+        ],
         "algorithm": "max",
     }
 
@@ -394,8 +401,10 @@ You can specify ``custom_calib`` as ``algorithm`` in ``quant_cfg`` to use it. He
 
     # create quantization configuration with "custom_calib" method
     quant_cfg = {
-        'quant_cfg': {'*weight_quantizer': ..},
-        'algorithm':  {"method": 'custom_calib'},
+        'quant_cfg': [
+            {"quantizer_path": "*weight_quantizer", "cfg": {...}},
+        ],
+        'algorithm': {"method": 'custom_calib'},
     }