#speculative-decoding articles

GAGraham Andanjefata-fanaka.hashnode.devSep 9 · 5 min read

Speculative Decoding

Speculative Decoding is an inference optimization technique that pairs a large target model with a lightweight draft mechanism that quickly proposes several next tokens. The target model verifies thos

0

MSManu Shuklaecorpit.hashnode.devAug 21 · 14 min read

LFM2.5-DSpark ships 3.2x on an H100 and 1.18x on a MacBook: the 20 August 2026 numbers mobile teams should read

LFM2.5-DSpark ships 3.2x on an H100 and 1.18x on a MacBook: the 20 August 2026 numbers mobile teams should read Summary. Liquid AI published LFM2.5-DSpark on 20 August 2026: three speculative-decoding

0

MSManu Shuklaecorpit.hashnode.devAug 2 · 15 min read

vLLM Speculative Decoding in 2026: P-EAGLE vs DFlash vs DSpark for Faster, Cheaper LLM Serving

vLLM speculative decoding in 2026: P-EAGLE vs DFlash vs DSpark for faster, cheaper LLM serving Summary. On July 28, 2026, the vLLM project and its Speculators library shipped open-source support for t

0

SLShaun Liewshaunliew.hashnode.devMay 18 · 15 min read

Qwen3 Speculative Decoding on the DGX Spark: Two Models, Four Methods, One Surprising Lesson

The Problem: One Token at a Time When you ask a large language model a question, it does not write the whole answer in a single step. It generates one small piece of text at a time. That piece is call

0

JKJangwook Kimeffloow.hashnode.devMay 11 · 5 min read

PARSE: Faster LLM Inference via Parallel Prefix Speculative Decoding

Speculative decoding became the standard inference speedup technique through 2024 and 2025. The idea: a small draft model generates a sequence of candidate tokens, and a larger target model verifies them in parallel — accepting the longest valid pref...

0

JKJangwook Kimeffloow.hashnode.devMay 9 · 6 min read

Gemma 4 MTP Drafters: How Multi-Token Prediction Delivers 2x+ Faster Local Inference

On May 5, 2026, Google released Multi-Token Prediction (MTP) drafters for the Gemma 4 family. The headline claim — up to 3x inference speedup — is technically accurate on specific hardware. The more realistic number for most developer setups is 1.7x ...

0

JKJangwook Kimeffloow.hashnode.devMay 9 · 10 min read

SpecKV: Adaptive Speculative Decoding with Dynamic Gamma

Every production LLM deployment using speculative decoding is likely running a fixed speculation length of γ=4. That number comes from early benchmarks, it has been copy-pasted across blog posts and framework defaults, and almost nobody questions it....

0

강강문규devsnack.hashnode.devMay 6 · 10 min read

Qwen3.6 on DGX Spark: vLLM + NVFP4 + DFlash vs llama.cpp — 2x Faster at 88–104 tok/s

TL;DR — I was happily running Qwen3.6 on llama.cpp. Then I saw claims of 2× speed with vLLM + NVFP4 + DFlash. So I installed it, fought through crashes, and measured it myself. Verdict: it's real. 88–

0

강강문규devsnack.hashnode.devMay 6 · 8 min read

Gemma 4 MTP Drafter on DGX Spark: 2.89x Speedup for Dense 31B — No Quality Loss

An 870 MB drafter model turned Dense 31B from 6.5 → 18.8 tok/s. No model swap, no training, no quality degradation. If you have a DGX Spark, there's no reason not to use this. Key Results Model Fra

0

AYAdam Yanggradient.networkNov 24, 2025 · 6 min read

Turning Latency into Throughput: Speculative Decoding for the Decentralized Inference

https://arxiv.org/abs/2511.11733 The Latency Wall In centralized inference, speed is mostly a function of compute. You optimize by saturating HBM bandwidth, fusing kernels, and keeping GPUs close to their roofline. In decentralized inference, where...

0

#speculative-decoding

#speculative-decoding

Explore Hashnode

Trending tags this week

Speculative Decoding

LFM2.5-DSpark ships 3.2x on an H100 and 1.18x on a MacBook: the 20 August 2026 numbers mobile teams should read

vLLM Speculative Decoding in 2026: P-EAGLE vs DFlash vs DSpark for Faster, Cheaper LLM Serving

Qwen3 Speculative Decoding on the DGX Spark: Two Models, Four Methods, One Surprising Lesson

PARSE: Faster LLM Inference via Parallel Prefix Speculative Decoding

Gemma 4 MTP Drafters: How Multi-Token Prediction Delivers 2x+ Faster Local Inference

SpecKV: Adaptive Speculative Decoding with Dynamic Gamma

Qwen3.6 on DGX Spark: vLLM + NVFP4 + DFlash vs llama.cpp — 2x Faster at 88–104 tok/s

Gemma 4 MTP Drafter on DGX Spark: 2.89x Speedup for Dense 31B — No Quality Loss

Turning Latency into Throughput: Speculative Decoding for the Decentralized Inference

#speculative-decoding

Search Hashnode

#speculative-decoding

Explore Hashnode

Trending tags this week

Speculative Decoding

LFM2.5-DSpark ships 3.2x on an H100 and 1.18x on a MacBook: the 20 August 2026 numbers mobile teams should read

vLLM Speculative Decoding in 2026: P-EAGLE vs DFlash vs DSpark for Faster, Cheaper LLM Serving

Qwen3 Speculative Decoding on the DGX Spark: Two Models, Four Methods, One Surprising Lesson

PARSE: Faster LLM Inference via Parallel Prefix Speculative Decoding

Gemma 4 MTP Drafters: How Multi-Token Prediction Delivers 2x+ Faster Local Inference

SpecKV: Adaptive Speculative Decoding with Dynamic Gamma

Qwen3.6 on DGX Spark: vLLM + NVFP4 + DFlash vs llama.cpp — 2x Faster at 88–104 tok/s

Gemma 4 MTP Drafter on DGX Spark: 2.89x Speedup for Dense 31B — No Quality Loss

Turning Latency into Throughput: Speculative Decoding for the Decentralized Inference