Ulysses Sequence Parallelism: 12x Longer Sequences

Hugging Face released TRL / Shared entities / Same source domain / What happened next

Liquid AI's DSpark draft models hit 3.18x throughput and cut function-calling latency 57%.

Linked by a graph relationship (Hugging Face released TRL); both cover GPU, H100; reported by the same outlet (huggingface.co).

Hugging Face released TRL / Shared entity: Transformers / Same source domain / Shared topic / What happened next

Hugging Face closed the Transformers-to-vLLM speed gap.

Linked by a graph relationship (Hugging Face released TRL); both cover Transformers; reported by the same outlet (huggingface.co).

Hugging Face released TRL / Shared entity: GPU / Same source domain / Earlier coverage

HyperNova 60B: Free Quantum-Compressed Model with 5x Agentic Performance

Linked by a graph relationship (Hugging Face released TRL); both cover GPU; reported by the same outlet (huggingface.co).

Hugging Face released TRL / Shared entity: H100 / What happened next / Tension

Claude designed protein binders against 14 of 15 targets at a 26.7% hit rate.

Linked by a graph relationship (Hugging Face released TRL); both cover H100; picks up the H100 thread on 2026-08-20.

Hugging Face released TRL / Shared entity: GPU / What happened next

Qwen3.8-27B is architecturally byte-identical to Qwen3.6-27B. Every gain came from training.

Linked by a graph relationship (Hugging Face released TRL); both cover GPU; picks up the GPU thread on 2026-08-15.

llama.cpp now runs a 295B model on one machine

Linked by a graph relationship (Hugging Face released TRL); both cover GPU; picks up the GPU thread on 2026-07-15.

Hugging Face released TRL / Shared entity: H100 / What happened next

Google open-sourced DiffusionGemma, a text-diffusion model that generates 4x faster, and you can run it locally today

Linked by a graph relationship (Hugging Face released TRL); both cover H100; picks up the H100 thread on 2026-06-11.

NanoRL uses transformers / Shared entities / What happened next

NanoRL fits REINFORCE, PPO, GRPO and RLOO into ~1,800 lines with no Ray, TRL or DeepSpeed.

Linked by a graph relationship (NanoRL uses transformers); both cover GPU, TRL; picks up the GPU thread on 2026-08-14.