Megatron-LM adds async dynamic inference with FP8 and multi-batch CUDA Graphs, introduces HyperCommGrid for flexible model-parallel comms, and lands major MoE performance upgrades.
Now supports Neural Sparse retrieval (SPLADE) with SEISMIC indexing, adds CPU support, and updates reasoning methods for smoother RAG experiments.
Now with video loop mode, a grid view for media, and auto‑resume from your last playback position across all URIs.
Big adds: Qwen 2.5‑VL integrated with vLLM, CCL ops migrated to TT‑Fabric for unified scale-up/scale-out, and Phi‑3‑Mini‑128k now supported on N150/N300.
Huge perf boosts, safer rm/du/chmod/chgrp via secure traversal, and new goodies like auto‑tuned sort buffers, rm --progress, ls -f, realpath -E, plus basenc --base58.
Single‑port deployments are now easy with Next.js API rewrites that proxy /api/* to the FastAPI backend, plus a new INTERNAL_API_URL for multi‑container setups.
Experimental Chinese formula parsing lands alongside cross-page continuation table merging and compatibility with vLLM 0.11.0.
Now with streaming cancellation, guardrails that can rewrite full AI messages, access to AgenticScope inside @Tool methods, custom RAG attributes into ChatMemory, and updated MCP support (registry + tool name mapping).
Now with accessible, PDF/UA‑ready PDFs and full PDF/A support, PDFs-as-images, character‑level justification, a typed HTML API with broader HTML export, and multi‑header tables.
Big update: experimental full lint/format support for HTML, Vue, Svelte & Astro, tsconfig baseUrl resolving, Tailwind v4 parsing, new rdjson/checkstyle reporters, and fresh CLI flags (CSS modules/Tailwind/JSON, format-with-errors).