implement propagate_error

Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>

Author: kylesayrs

src/llmcompressor/args/dataset_arguments.py

@@ -223,10 +223,19 @@ class DatasetArguments(CustomDatasetArguments):
     quantization_aware_calibration: bool = field(
         default=True,
         metadata={
-            "help": "Whether to enable quantization-aware calibration in the pipeline. "
-            "When True, quantization is applied during forward pass in calibration. "
-            "When False, quantization is disabled during forward pass in calibration. "
-            "Default is set to True."
+            "help": "Only relevant for the sequential pipeline. "
+            "If True, quantization is applied during forward pass in calibration. "
+            "If False, quantization is disabled during forward pass in calibration. "
+            "Default is True."
+        },
+    )
+    propagate_error: bool = field(
+        default=True,
+        metadata={
+            "help": "Only relevant for the sequential pipeline. If True, use quantized "
+            "layer outputs as the inputs to the next sequential layer. If False, use "
+            "unquantized layer outputs as the inputs to the next sequential layer. "
+            "Default is True"
         },
     )
 

src/llmcompressor/pipelines/sequential/pipeline.py

@@ -97,7 +97,11 @@ def __call__(
                     # do a preliminary pass to trigger modifier hooks
                     for batch_idx in tqdm(range(len(dataloader)), desc=calib_desc):
                         inputs = activations.fetch(batch_idx, subgraph.input_names)
-                        subgraph.forward(model, **inputs)
+                        outputs = subgraph.forward(model, **inputs)
+
+                        if not dataset_args.propagate_error:
+                            activations.update(batch_idx, outputs)
+                            activations.delete(batch_idx, subgraph.consumed_names)
 
                     LifecycleCallbacks.sequential_epoch_end(subgraph)
 
@@ -106,10 +110,10 @@ def __call__(
                     with HooksMixin.disable_hooks():
                         for batch_idx in tqdm(range(len(dataloader)), desc=prop_desc):
                             inputs = activations.fetch(batch_idx, subgraph.input_names)
-                            output = subgraph.forward(model, **inputs)
+                            outputs = subgraph.forward(model, **inputs)
 
-                            if subgraph_index < num_subgraphs - 1:
-                                activations.update(batch_idx, output)
+                            if dataset_args.propagate_error:
+                                activations.update(batch_idx, outputs)
                                 activations.delete(batch_idx, subgraph.consumed_names)
 
             # redundant, finish any remaining compression