6arXiv cs.LG (Machine Learning)·
FlashRT: Agent harness that converts reference implementations into optimized multi-GPU deployments for real-time multimodal apps
FlashRT is an agent harness that guides coding agents to transform developer-written reference implementations of real-time multimodal pipelines (voice agents, video generation, multimodal LLMs) into optimized multi-GPU deployments. Using a 'chain-of-program' paradigm, the system directs an agent through IR construction, static analysis, and a measurement-gated optimization loop to tune placement, streaming, and parallelism. Across benchmarks on NVIDIA B200 and AMD MI355X GPUs, FlashRT achieves up to ~70x latency reduction and 3.6x throughput improvement, and outperforms the expert vLLM-Omni implementation for Qwen3-Omni text-to-audio inference on AMD hardware. The result suggests agent-driven optimization may be especially valuable on platforms lacking mature expert tooling.
Related guides (3)
Related events (8)
Rhetor: Multi-agent system for rehearsed live product demos with real-time voice Q&A
Researchers introduce Rhetor, a multi-agent system that automates live software product demonstrations by taking a running web application and its source code as input, then producing a rehearsed demo with synchronized narration and real-time voice question answering. The system combines UI exploration with source-code analysis, uses semantic locators for browser action dispatch, and includes a pre-presentation rehearsal loop with graceful degradation. Evaluated across six pipeline sessions on four deployed applications, the system achieves high locator-firing rates (sigma-bar ~0.92 on a 53-action workload) and converges to perfect locator resolution on a public-domain reference app. The paper also proposes a ten-metric benchmark protocol for evaluating demo automation systems.
20x Faster TRL Fine-tuning with RapidFire AI
RapidFire AI claims to achieve 20x faster fine-tuning throughput using TRL (Transformer Reinforcement Learning library) compared to standard configurations. The announcement appears on the Hugging Face blog, suggesting integration or compatibility with the HF ecosystem. No additional technical details are available from the body of the post, but the claim targets a significant pain point in LLM post-training workflows.
DreamForge-World 0.1 Preview: real-time controllable world model on consumer GPUs
Researchers present DreamForge-World 0.1 Preview, a world model for real-time interactive simulation that runs on a single RTX 4090 at 14-15 FPS at 480p resolution. The system adapts the LongLive 1 autoregressive video stack (derived from Wan2.1-T2V-1.3B) with a residual action pathway from the Matrix-Game family, supporting keyboard/mouse control, multimodal initialization, mid-stream reprompting, and dual-view operation. The work targets a low-compute niche distinct from frontier-scale world simulators, demonstrating a cost-efficient route to consumer-GPU-deployable interactive world models.
Mach-Mind-4-Flash: 35B MoE agentic model matching 100B-class performance via post-training optimization
Mach-Mind-4-Flash is a 35B-parameter Mixture-of-Experts model with only 3B activated parameters that achieves performance comparable to 100B-class models through post-training techniques alone. The pipeline combines a unified RL/OPD training infrastructure with multi-teacher scheduling, parallel domain-specific RL experts fused via Multi-Teacher On-Policy Distillation (MOPD), and Hybrid Median-length Policy Optimization (HMPO) which compresses reasoning chains 19-46% with minimal accuracy loss. Benchmark results include 92.70 on AIME'26, 82.82 on IFBench, and 75.80 on BFCL-v4, claiming to lead or match models 10-30x its activated size at a fraction of inference cost. The work is notable for demonstrating that post-training optimization can close large gaps in activated parameter count for agentic tasks.
RRFP: A Readiness-Driven Runtime for Pipeline-Parallel Training Under Runtime Variability
The paper introduces Runtime-Readiness-First Pipeline (RRFP), a new runtime for pipeline-parallel large-model training that treats schedules as non-binding hint orders rather than strict execution sequences. By combining message-driven asynchronous communication, lightweight tensor-parallel coordination, and ready-set arbitration, RRFP dynamically dispatches work based on actual task readiness, reducing idle bubbles and stage misalignment. Implemented on a Megatron-based framework and evaluated at up to 128 GPUs, RRFP achieves up to 1.77× speedup on language-only workloads and 2.77× on multimodal workloads versus fixed-order baselines, and outperforms the fastest comparable external system by up to 1.84×.
Optimum-NVIDIA: One-Line LLM Inference Acceleration via TensorRT-LLM
Hugging Face's Optimum-NVIDIA integration wraps NVIDIA's TensorRT-LLM backend to enable high-performance LLM inference with minimal code changes. The library targets developers who want near-peak GPU throughput without manually configuring TensorRT-LLM pipelines. It positions as a bridge between the Hugging Face ecosystem and NVIDIA's optimized inference stack.
OmniAgent: POMDP-based active perception agent for long video understanding with test-time scaling
Researchers introduce OmniAgent, a multimodal agent that reformulates long video understanding as a POMDP-based iterative Observation-Thought-Action cycle, selectively distilling audio-visual cues into persistent textual memory rather than processing all frames uniformly. The system uses Agentic Supervised Fine-Tuning and a novel reinforcement learning method (TAURA) with turn-level entropy for credit assignment. OmniAgent demonstrates positive test-time scaling and achieves state-of-the-art open-source results across ten benchmarks, with its 7B model outperforming Qwen2.5-VL-72B on LVBench (50.5% vs. 47.3%).
Fine-tuning 20B LLMs with RLHF on a 24GB consumer GPU
Hugging Face demonstrates a method for running RLHF fine-tuning on 20-billion-parameter language models using a single 24GB consumer GPU by combining TRL and PEFT (parameter-efficient fine-tuning). The approach uses techniques like LoRA and quantization to dramatically reduce memory requirements. This lowers the hardware barrier for RLHF experimentation from multi-GPU server setups to consumer-grade hardware.