#lm_head
7 related articles
·2 min
Unsloth Dynamic v3 Released: 10% Accuracy Gain at Same Size, 1-bit Quantization in Just 6GB
Unsloth releases Dynamic v3 quantization: Qwen3.8-27B GGUF models achieve 10% top-1% accuracy gain at same size, plus 6-8GB 1-bit extreme quantization. New Divergence-300 metric for realistic evaluation.
Read more →
·3 min
nanoGPT Speedrun Techniques: How Delayed Untying Solves the Sparse Gradient Problem in Embedding Layers
Deep dive into the Delayed Untying technique in nanoGPT speedruns: why tying embed and lm_head weights early then untying later solves both sparse gradients and limited expressiveness.
Read more →
·3 min
Latent Space Reasoning: How DeepSeek-V4 Lets AI Think in Hidden Layers
Deep dive into DeepSeek-V4's latent space reasoning technology — how AI shifts from explicit chain-of-thought to implicit vector space reasoning, its efficiency gains, and challenges in interpretability.
Read more →
·3 min
D-Flash: The Secret Behind 3.5x LLM Inference Speedup with Diffusion Drafting
D-Flash uses fast diffusion parallel drafting and target feature KV injection to solve the autoregressive Drafter latency bottleneck in speculative decoding. 16 tokens in just 6ms, up to 3.5x speedup on HumanEval, surpassing EAGLE3 and MTP.
Read more →
·2 min
D-Flash: How Diffusion Drafting Delivers 3.5x Faster LLM Inference
D-Flash solves the autoregressive drafter latency bottleneck in speculative decoding via fast diffusion parallel drafting and target feature KV injection. 16 tokens in just 6ms, up to 3.5x speedup on HumanEval, beating EAGLE3 and MTP.
Read more →
·3 min
Running a 27B Model on iPhone: How 1-bit Quantization Achieves Extreme Compression
PrismML's Bonsai compresses a 27B model from 54GB to 3.9GB, running at ~11 tokens/sec on iPhone. A deep dive into QAT, knowledge distillation, and speculative decoding.
Read more →
