TensorRT-LLM v1.3.0rc22 breaks torch.compile for DeepSeek-V3 & Llama-3.1-8B FP8

🚀 Models & Releases

• NVIDIA TensorRT-LLM v1.3.0rc22 introduces a breaking change: torch.compile crashes in the PyTorch backend, and multi-GPU accuracy paths fail with KeyError in the remove_copy pass for configurations like DeepSeek-V3-Lite (bf16/FP8/NVFP4) and Llama-3.1-8B-Instruct (FP8 with FlashInfer). Workaround: disable torch.compile for affected setups. DeepSeek-V3.2 FP8 block-scale also hits OOM issues. (NVIDIA/TensorRT-LLM)

• Ling-3.0-flash, a 124B-parameter MoE model (5.1B activated/token), launches as a free tier on OpenRouter. Prioritizes token efficiency and agentic inference for production-scale workflows. (OpenRouter)

• Poolside Laguna S 2.1 (118B total params, 8B active) achieves 70.2% on Terminal-Bench 2.1, targeting coding agent use cases. (OpenRouter)

• Google Gemini 3.5 Flash Lite upgrades agentic capabilities for subagents in multi-agent workflows. (OpenRouter)

🛠️ Outils & Produits

• llama.cpp b10099 improves NVFP4 W4A4 activation quantization with 32-byte loads, per-block amax, and fused per-channel quantization kernels. (ggml-org/llama.cpp)

• goose v1.44.0 adds support for latest Gemini models, declarative Sakana AI provider (OpenAI-compatible Fugu API), and working directory context for Stop hooks. (block/goose)

• cline Desktop v0.0.4 enables workspace-free chat sessions, drag-and-drop file attachments, inline image display, and one-time routine scheduling. (cline/cline)

• AutoGPT Platform v0.6.69 introduces app sidebar redesign, proactive Slack/Telegram posting, DM delivery for copilot bots, and Agent-Building Mode. (Significant-Gravitas/AutoGPT)

• pydantic-ai v2.17.0 adds arbitrary field support to RequestUsage/RunUsage for upcoming genai-prices, and caches per-message OTel serialization to avoid O(n²) overhead. (pydantic/pydantic-ai)

⚙️ API & Dépréciations

• Anthropic API now supports effort levels for Claude Managed Agents (configure via model object at agent creation). Webhooks expand to cover environment/memory store lifecycle (environment.*, memory_store.* events). (Anthropic Docs)

• langchain-openai v1.4.1 and langchain-fireworks v1.5.1 add LangSmith Gateway support via env vars, and fix gpt-5.3-chat-latest profile. (langchain-ai/langchain)

⚖️ Régulation & Société

• US Congress Democrats pressure USTR over reduced AI regulation powers in the USMCA update. (MLex)

• House Energy & Commerce Committee discusses AI policy and the Ratepayer Protection Act, signaling legislative focus on AI governance. (House.gov)

• MIT Tech Review explores how AI accelerates drug discovery, reducing costs and failure rates in biologic medicine development. (MIT TR)

• TechCrunch reports on AI guardrails hindering offensive cybersecurity research, with researchers citing OpenAI/Anthropic restrictions. (TechCrunch)

🔬 Recherche

• VLM-IE3D enhances 3D spatial awareness in vision-language models (VLMs) for fine-grained reasoning tasks. (arXiv)

• Expanding Generative Flows (EFlows) introduces dynamic dimension/sequence length support for flow-based generative models. (arXiv)

• GraphVid enables interactive graph-controllable video generation, addressing multi-object interaction precision challenges. (arXiv)

• AREX proposes a recursively self-improving agent for deep research, leveraging discovery-verification asymmetry. (Hugging Face)

💡 À retenir

TensorRT-LLM v1.3.0rc22 breaks torch.compile for DeepSeek-V3 and Llama-3.1-8B FP8 — immediate action required to avoid production crashes. Disable torch.compile for affected configs until NVIDIA releases a fix.