speculative
speculative_decoding_benchmarksspeculative-decoding-bench-rtx4090
Speculative Decoding Benchmark — RTX 4090
TL;DR: 4,576 benchmark runs measuring speculative decoding speedup / acceptance rate
across llama.cpp and LM Studio, Qwen3 (8B/14B) and Llama-3.1-8B target models, on a
single consumer RTX 4090 (24GB). Best observed case: the draft-free ngram-mod
self-speculative mode on structured tasks (JSON extraction 2.81x, code 2.76x,
global-median aggregation at temp=0). Open-ended tasks (creative writing, translation)
with a traditional draft… See the full description on the dataset page: https://huggingface.co/datasets/steven0226/speculative-decoding-bench-rtx4090.speculative-reasoning-matheval-4b-sweep
MathEval sweep — speculative reasoning on Qwen3-4B
Per-sample generations and grading for four arms of a MathEval run, measuring what
speculative reasoning costs and saves against a base model that does not speculate.
Code and write-up: yurun-yuan/speculative-reasoning
— see docs/05-rl-4b.md.
The arms
All four answer the same 1,547 MathEval problems under a 20,000-token response budget.
split
model
runtime
base_plain
Qwen/Qwen3-4B
plain — no speculation… See the full description on the dataset page: https://huggingface.co/datasets/yyuan244/speculative-reasoning-matheval-4b-sweep.eagle3-speculative-decoding-energy-sweep
EAGLE3 Speculative Decoding Energy Sweep
Per-config energy/throughput/latency measurements for EAGLE3 speculative decoding
(speculative_num_steps, speculative_eagle_topk, speculative_num_draft_tokens)
served with sglang, across batch sizes. Collected for an RL project that learns to
pick speculative-decoding parameters to hold GPU energy utilization in a target band.
Model: unsloth/Llama-3.2-1B-Instruct + rescommons/SpecForge-EAGLE3-Llama-3.2-1B-Instruct draft head.
Hardware:… See the full description on the dataset page: https://huggingface.co/datasets/Pradheep1647/eagle3-speculative-decoding-energy-sweep.speculative-ar-video-results
Speculative AR Video Generation Results
Curated qualitative comparisons for versioned speculative autoregressive
video-generation methods.
Initial contents
2 method versions
16 fixed cases, one from each assigned VBench dimension
64 MP4 files: method generations plus their matched targets
Each row in data/metadata.jsonl pairs one method generation with the target
produced by that exact experiment version. Targets are method-versioned because
validation found… See the full description on the dataset page: https://huggingface.co/datasets/CCChristine2/speculative-ar-video-results.speculative-decoding-benchmark-results