Fix and test batched inference/generation, position_ids creation, falcon alibi, gpt_bigcode multi-query,..

[IlyasMoutawwakil]

same as huggingface/optimum#2326

[Copilot]

Pull Request Overview

The PR fixes and tests batched inference/generation, position_ids creation, falcon alibi, and gpt_bigcode multi-query attention handling across the ONNX runtime and exporters.

• Updates decoder modeling to properly handle batched inputs, position_ids generation, and past key values masking
• Refactors test files to use batched inputs by default and improve test coverage for various model architectures
• Simplifies CI workflows and removes deprecated model patchers

Reviewed Changes

Copilot reviewed 12 out of 12 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tests/onnxruntime/testing_utils.py	Adds new model variants for falcon-alibi-True and gpt_bigcode-multi_query-False testing
tests/onnxruntime/test_diffusion.py	Replaces external image loading with random image generation and standardizes tolerance values
tests/onnxruntime/test_decoder.py	Major refactor to use batched inputs, improved tokenizer handling, and better past key values comparison
tests/exporters/onnx/test_export.py	Updates custom MPT config to use correct sequence dimension naming
optimum/onnxruntime/modeling_decoder.py	Comprehensive fixes for position_ids creation, past key values handling, and model-specific logic
optimum/exporters/onnx/utils.py	Adds missing model types to position_ids requirements list
optimum/exporters/onnx/model_patcher.py	Removes deprecated model patchers and adds ONNX-compatible implementations
optimum/exporters/onnx/model_configs.py	Simplifies model configurations and fixes sequence dimension naming
optimum/exporters/onnx/config.py	Fixes attention mask dimension ordering for past key values
optimum/exporters/onnx/base.py	Updates sequence dimension naming for past key values
.github/workflows/test_onnxruntime_slow.yml	Simplifies CI workflow by removing matrix variations
.github/workflows/test_onnxruntime.yml	Adds transformers version matrix testing

Comments suppressed due to low confidence (3)

optimum/exporters/onnx/model_patcher.py:374

• [nitpick] The function name 'onnx_compatible_triu' doesn't follow the existing naming pattern. Consider renaming to 'patched_triu' to match the pattern used by other patched functions like 'patched_forward'.

def onnx_compatible_triu(input_tensor: torch.Tensor, *args, **kwargs) -> torch.Tensor:

optimum/exporters/onnx/model_patcher.py:366

• [nitpick] The function name 'onnx_compatible_tril' doesn't follow the existing naming pattern. Consider renaming to 'patched_tril' to match the pattern used by other patched functions like 'patched_forward'.

def onnx_compatible_tril(input_tensor: torch.Tensor, *args, **kwargs) -> torch.Tensor:

optimum/exporters/onnx/model_patcher.py:406

• [nitpick] The function name 'noop_bfloat16_casting' could be more descriptive. Consider renaming to 'identity_bfloat16_casting' or 'passthrough_bfloat16_casting' to better indicate that it returns the tensor unchanged.

def noop_bfloat16_casting(self):

[echarlaix]

LGTM, thanks a lot