Microsoft Releases Flint Visualization Language and Hugging Face Improves vLLM Backend

개발 도구 | Thu Jul 09 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 3 sources

Microsoft released a chart generation language for AI agents while Hugging Face achieved native-speed inference with its vLLM backend.

Analysis

[Microsoft Research] open-sourced the Flint visualization language ^[1][2]

• Open-source visualization language for the AI era
• Provides a middle path between short chart specifications and expressive results
• Supports compact
• human-editable specifications
• Enables AI agents to generate expressive charts

[Hugging Face] achieved native speed with the vLLM transformers modeling backend ^[3]

• Available via a single --model-impl transformers flag
• Matches or exceeds native performance on Qwen3 4B
• 32B
• and 235B FP8 MoE
• Runs 450+ transformers architectures on vLLM without porting
• Combinable with tensor/data/expert parallelism

[Hugging Face] introduced torch.fx-based runtime graph optimization ^[3]

• Performs static analysis of model graphs using torch.fx
• Manipulates source code and rewrites operations via AST
• Maps fused operations to vLLM kernels for Expert Parallelization in MoE
• Infers TP parallel plans via MergedColumnParallelLinear and QKVParallelLinear

Sources

• [1] Flint: A visualization language for the AI era - Microsoft Research Blog
• [2] Show HN: Microsoft releases Flint, a visualization language for AI agents - Hacker News
• [3] Native-speed vLLM transformers modeling backend - Hugging Face Blog