Open Source Contributions

Merged in three repositories of the vllm-project organization.

vllm : the inference engine

• hadacore_transform: respect inplace parameter : the kernel wrote its output into the input tensor, ignoring the inplace parameter; under torch.compile, QuIP models produced garbage outputs

llm-compressor : quantization toolkit for LLM deployment

Approved and merged by the projects' maintainers.

• iMatrix weighted MSE observer and IMatrixGatherer : importance-weighted (E[x²]) range selection, no full Hessian
• Norm calibration context for unit-offset RMSNorm : fixes garbage AWQ output on Gemma
• MoE calibration module for GlmMoeDsa (GLM-5) : unpacks GLM-5's packed 3D expert weights
• Topological ordering in FX graph cleanup : fixes erase_node crash on Granite4 GPTQ
• Packed weight handling in granite4 to_3d_expert : W4A16 support
• SmoothQuant regex fix for q_a_proj : DeepSeek and GLM-5
• SmoothQuant mapping for GLM-5
• AWQ mapping for GLM-5
• MR-GPTQ example : QuIP + GPTQ + NVFP4A16

compressed-tensors : safetensors extension for sparse and quantized tensors

Approved and merged by the projects' maintainers.

• N-dimensional tensor support in pack/unpack_int32 : fixes 3D MoE expert weight packing
• cast_to_fp4 per-rank torch.compile recompilation fix : removes a Dynamo recompile storm on NVFP4 MoE generation

GPU Kernels

• gemm-ladder : an fp16 GEMM built step by step on a T4 and an A100, from NVIDIA's tensor-core sample, adapted (0.62x and 0.27x), to 1.05x and 1.01x the judge's fastest cuBLASLt configuration, with the tools and raw outputs in the repository.

Applied AI Systems

• finsight : visual RAG over 10 French annual reports (ColQwen2.5 + Qdrant). 5,982 pages indexed, 92% Recall@10, 84% strict answer accuracy (LLM-judge on non-circular ground truth), 92% citation faithfulness. Async FastAPI with SSE streaming, React frontend, 184 tests, CI.
• reasonforge : iterative self-improvement for Text-to-SQL on Ministral-8B. Spider dev: 60.1% base, 68.8% SFT, 78.0% looped with k=16 self-consistency (vLLM + LoRA + H200).
• ai-watch : autonomous AI news agent on LangGraph. Daily briefings in production since February 2026 via GitHub Actions.