TL;DR: Since the initial overview in 2024, the “bottleneck war” has moved from simple KV cache management to architectural revolutions like Multi-Head Latent Attention (MLA) and hardware-native 4-bit floating point (FP4) on Blackwell GPUs.

This post serves as a direct update to my 2024 article on Large Transformer Model inference. While the original discussion established the foundations of I/O awareness and memory fragmentation, the industry has since moved toward a vertically integrated stack where model architecture and hardware work in unison. Below is a high-level summary contrasting the foundational techniques with the breakthroughs that define the 2026 landscape.

Updated Summary Table

Technique Phase Optimized Primary Benefit Applications / Frameworks
FOUNDATIONS (until 2024)      
Quantization (AWQ1 / SmoothQuant2) Both Reduced VRAM TensorRT-LLM3, vLLM4, BitsAndBytes
vLLM (PagedAttention4) Decode Solves Fragmentation Industry Standard (vLLM, TGI, Ray Serve5)
GQA6 / MQA7 Decode Smaller KV Cache Llama 2/38, Mistral 7B9, Falcon 40B10
FlashAttention-1, 2, 311 Prefill IO-Awareness & Asynchrony Native in PyTorch12, JAX13, CUDA kernels14
Speculative Decoding (Draft-Target)15 Decode Lower Latency T5-XXL15, Early GPT-4 Serving 16
NEW FRONTIERS (2025 & 2026)      
MLA (Latent Attention)17 Decode 4-6x KV Cache reduction DeepSeek-V317, SGLang18, Qwen-Reasoning
MTP / Self-Speculation17 Decode Native generation speed DeepSeek-V3, GPT-4o / GPT-519, Qwen320
FP4 (NVFP4)21 Both 2-4x Throughput Llama 422, FLUX.123, Blackwell GPUs
RadixAttention24 Prefill Instant Prefix Reuse SGLang, vLLM (Prefix Caching)25, Snowflake26
P-EAGLE27 Decode Parallel Drafting vLLM, TensorRT-LLM, Qwen3-Coder27

Overview: The 2026 Landscape

In late 2024, the focus was on squeezing efficiency out of standard Transformers using techniques like GQA 6 and vLLM 4. In 2026, we have entered the era of Inference-Aware Architectures. Models are now designed during the pre-training phase to be inherently optimized for low-precision hardware and massive context windows.

New Algorithmic Optimization

Multi-Head Latent Attention (MLA)

Popularized by the DeepSeek-V3 series 17, MLA is the spiritual successor to Grouped-Query Attention (GQA). While GQA reduced the number of heads to save memory, MLA uses low-rank joint compression to “squeeze” Key and Value vectors into a tiny latent vector.

  • Impact: It reduces the KV cache memory footprint by 4–6x compared to GQA.
  • Adoption: Beyond DeepSeek, this architectural shift is seen in the Qwen-Reasoning models and is a core optimization supported in the SGLang inference engine 18.

Multi-Token Prediction (MTP) & Self-Speculation

Moving beyond standard “Next Token Prediction,” 2025/2026 models are increasingly trained with MTP heads. The model is trained to predict $k$ future tokens in parallel. This enables Self-Speculation, where the model drafts its own future tokens in a single forward pass, removing the need for a separate, smaller “draft model” previously required for speculative decoding 15.

  • Usage: This is a defining feature of the DeepSeek-V3 and Qwen3 families 20. Evidence suggests that GPT-4o and GPT-5 transitioned to native multi-token prediction heads to achieve the high throughput observed in production, eliminating the I/O overhead of separate draft models 1928.

System and Hardware Breakthroughs

FP4 & NVFP4 (NVIDIA Blackwell)

With the rollout of the Blackwell architecture 21, FP4 (4-bit Floating Point) has replaced INT8/INT4 as the gold standard for high-speed inference. Unlike INT4, the NVFP4 format handles the dynamic range of activations much better, leading to negligible accuracy degradation. This provides a 2x–4x throughput boost over FP8/FP16.

  • Ecosystem: Llama 4 22 and the FLUX.1 image generation family 23 are among the first to be distributed with native FP4 weights for Blackwell-based clusters.

RadixAttention & Prefix Caching

While vLLM solved physical memory fragmentation, RadixAttention (pioneered in SGLang 24) addresses the “redundant prefill” problem.

  • Mechanism: It treats the KV cache as a Radix Tree. If multiple queries share a common system prompt or document, the engine “hits” the cache and skips the prefill phase entirely.
  • Impact: This has been adopted as “Prefix Caching” in vLLM 25 and is extensively used by companies like Snowflake 26 and Anyscale to optimize TTFT for long-context RAG pipelines.


Figure 1: RadixAttention tree structure for efficient prefix reuse

Parallel Speculative Decoding (P-EAGLE)

Standard speculative decoding was often bottlenecked by sequential verification. P-EAGLE 27 allows the drafter model to generate a tree of possible future tokens in a single parallel step, pushing generational speedups from 2x up to 3.5x in high-concurrency environments.

  • Serving: Now a staple in vLLM and TensorRT-LLM v1.0, especially for coding models like Qwen3-Coder 27 and GPT-OSS 29 where structured syntax makes parallel drafting highly effective.

References

  1. Lin, Ji, et al. “AWQ: Activation-aware Weight Quantization for LLM Compression.” (2024). ↩

  2. Xiao, Guangxuan, et al. “SmoothQuant: Accurate and Efficient Post-Training Quantization for LLMs.” (2023). ↩

  3. NVIDIA. “Optimizing LLM Inference Performance with NVIDIA TensorRT-LLM.” (2024). ↩

  4. Kwon, Woosuk, et al. “vLLM: Efficient Memory Management for LLM Serving.” (2023). ↩ ↩2 ↩3

  5. Anyscale. “Scaling LLM Workloads with Ray Serve and vLLM.” (2024). ↩

  6. Ainslie, Joshua, et al. “GQA: Training generalized multi-query transformer models.” (2023). ↩ ↩2

  7. Shazeer, Noam. “Fast transformer decoding: One write-head is all you need.” (2019). ↩

  8. Dubey, Abhimanyu, et al. “The Llama 3 Herd of Models.” (2024). ↩

  9. Jiang, Albert Q., et al. “Mistral 7B.” (2023). ↩

  10. Almazrouei, Ebtesam, et al. “The Falcon Series of Language Models.” (2023). ↩

  11. Shah, Jay, et al. “FlashAttention-3: Fast and accurate attention with asynchrony.” (2024). ↩

  12. PyTorch Foundation. “PyTorch 2.2: FlashAttention-v2 integration.” (2024). ↩

  13. nshepperd. “JAX bindings for Flash Attention v2.” (2024). ↩

  14. Dao, Tri, et al. “Flashattention: Fast and memory-efficient exact attention.” (2022). ↩

  15. Leviathan, Yaniv, et al. “Fast inference from transformers via speculative decoding.” (2023). ↩ ↩2 ↩3

  16. Artificial Analysis. “GPT-4o API Provider Benchmarking & Analysis.” (2024). ↩

  17. DeepSeek-AI. “DeepSeek-V3 Technical Report.” (2025). ↩ ↩2 ↩3 ↩4

  18. SGLang Project. “DeepSeek-V3 Support in SGLang.” (2025). ↩ ↩2

  19. OpenAI. “Introducing GPT-5.4 mini and nano.” (2026). ↩ ↩2

  20. Alibaba Group. “Alibaba Open-Sources Qwen3.5.” (2026). ↩ ↩2

  21. NVIDIA. “3 Ways NVFP4 Accelerates AI Training and Inference.” (2026). ↩ ↩2

  22. Wikipedia. “Llama (Language Model): Llama 4 Release.” (2025). ↩ ↩2

  23. NVIDIA Blog. “Scaling NVFP4 Inference for FLUX.1 on NVIDIA Blackwell GPUs.” (2026). ↩ ↩2

  24. LMSYS. “SGLang: Efficient Execution of Structured Language Model Programs.” (2024). ↩ ↩2

  25. vLLM Project. “Automatic Prefix Caching Design.” (2026). ↩ ↩2

  26. Agarwal et al. “From Prefix Cache to Fusion RAG Cache.” (2026). ↩ ↩2

  27. AWS Machine Learning Blog. “P-EAGLE: Faster LLM inference with Parallel Speculative Decoding.” (2026). ↩ ↩2 ↩3 ↩4

  28. Xu, Jiawei, et al. “A Comprehensive Survey on Large Language Models: From Pre-training to Autonomous Agents.” (2025). ↩

  29. Hugging Face Blog. “GPT-OSS Model Evaluation.” (2025). ↩

<
Previous Post
Large Transformer Model - Inference Optimization
>
Next Post
MTP & MTP-D Deep Dive: Beyond Next-Token Prediction