.NET AI Fundamentals Feedback / Feature Requests

[luisquintanilla]

Original issue created in microsoft/semantic-kernel#9793.

Sample project: tjwald/high-perf-ML

Related feedback:

• Adds the new System.Numerics.Tensors as an input/output type when using dotnet 8.0 and up. microsoft/onnxruntime#23261
• Tensor<T> and supporting types runtime#100924
• Tensor<T> and supporting types, part 2 runtime#103611
• Batch Tokenization Support #7371
• Context Aware Tokenization #7372
• RunAsync C# API crashes without any error microsoft/onnxruntime#19140
• Docs: HuggingFace (NLP) Migration Guide #7384

[tjwald]

Both issues linked:

dotnet/runtime#100924

dotnet/runtime#103611

Should be replaced with: dotnet/runtime#113068

[svick]

AI-generated analysis: This comparison was produced by Copilot.

After reviewing the full discussion in microsoft/semantic-kernel#9793, I found three requests that do not appear to be cleanly represented by dedicated tracking issues:

1. Hugging Face-style tokenizer factory — an equivalent of AutoTokenizer.from_pretrained(...) that detects the tokenizer type and loads its configuration and resources. Docs: HuggingFace (NLP) Migration Guide #7384 covers migration documentation, and issues such as Tokenizers: read the tokenizer.json 'decoder' section in CreateFromTokenizerJson #7660 cover individual tokenizer.json gaps, but neither tracks the complete factory/API request.
2. High-level inference pipelines — reusable equivalents of Hugging Face task pipelines that combine tokenization, batched model execution, and post-processing. Docs: HuggingFace (NLP) Migration Guide #7384 tracks documenting pipeline migration, but not implementing this product capability.
3. Inference orchestration — dynamic batching, time-window scheduling, GPU/model placement, and memory-aware hosting. This was described in the original issue's “Bonus” section, and I could not find a dedicated tracking issue for it.