feature: add padding side to allow pad from left

[huangrpablo]

Add configurable padding side for batch processing

Problem

The batch processor only supported right padding, which works for encoder-decoder models but causes issues with decoder-only models (like GPT) that require left padding to maintain proper attention patterns.

Solution

Added a padding_side parameter throughout the inference pipeline to control where padding tokens are added when batching sequences of different lengths.

Changes

• Core: Enhanced stack_features() to support left/right padding
• Sync: Added padding_side parameter to ModelBatchProcessor and @dynamically decorator
• Async: Added padding_side parameter to AsyncModelBatchProcessor and async @dynamically decorator
• Tests: Comprehensive test coverage for both padding modes

Usage

# For decoder-only models (GPT, etc.)
processor = ModelBatchProcessor(
    batch_func,
    batch_size=32,
    pad_tokens={"input_ids": 0},
    padding_side="left"  # New parameter
)

# Works with decorators too
@dynamically(batch_size=32, pad_tokens={"input_ids": 0}, padding_side="left")
def gpt_inference(features):
    return model(**features)

Backward Compatibility

• ✅ Fully backward compatible - defaults to "right" padding
• ✅ No breaking changes to existing code
• ✅ All existing tests continue to pass

Example

# Input: [1, 2, 3] and [4, 5]
# Right padding (default): [[1, 2, 3], [4, 5, 0]]
# Left padding (new):      [[1, 2, 3], [0, 4, 5]]

Fixes issues with decoder-only model batching while maintaining full backward compatibility.

[aamir-s18]

lgtm