Gahow Wang 6485c87c5b docs: Phase 25 — three speculative-decoding paradigms compared ...

Contrast Small-Model (Phase 22-24, done), EAGLE3 (this phase's target),
and Multi-Token Prediction (DeepSeek-V3 style, not applicable here).

Includes the actual EAGLE3-Qwen3-8B weight tensor listing pulled from
AngelSlim/Qwen3-8B_eagle3 on dash5:
- 1 midlayer (attention + mlp) with hidden_size=4096
- fc.weight (4096, 12288) fusing 3 target hidden-state levels
- q_proj (4096, 8192) taking concat(embed, fused_h) as input
- lm_head only over draft_vocab_size=32000, mapped back with d2t table
- ~750 MB total (vs 1.2 GB for Qwen3-0.6B), draft cost ~1/10 of target

Also captures Qwen3-8B + EAGLE3 speedup benchmark on vLLM: ~1.97-2.02x
across MT-bench/HumanEval/GSM8K/Alpaca. That's the number to beat.

Next commits will implement Eagle3Head in xserv-model + hook target
hidden states out of Qwen3::decode_core.

2026-07-01 16:53:37 +08:00

..

benchmarks

docs: Phase 21 — decode CUDA graph + GPU argmax results

2026-06-12 20:12:37 +08:00

00-roadmap.md

docs: Phase 21 — decode CUDA graph + GPU argmax results

2026-06-12 20:12:37 +08:00

01-cuda-ffi.md

docs: add design docs + takeaways for Phase 2 and Phase 3

2026-05-21 20:59:45 +08:00

02-tensor.md

docs: add design docs + takeaways for Phase 2 and Phase 3

2026-05-21 20:59:45 +08:00

03-gemm.md

docs: add design docs + takeaways for Phase 2 and Phase 3

2026-05-21 20:59:45 +08:00

04-transformer-kernels.md

phase 4: transformer core kernels

2026-05-21 21:07:24 +08:00

05-attention.md

phase 5: naive multi-head attention

2026-05-21 21:17:23 +08:00

06-model-loading.md

phase 6+7+8: model loading, BPE tokenizer, GPT-2 inference (Milestone ①)

2026-05-21 22:04:00 +08:00

07-tokenizer.md

phase 6+7+8: model loading, BPE tokenizer, GPT-2 inference (Milestone ①)

2026-05-21 22:04:00 +08:00

08-gpt2.md

phase 6+7+8: model loading, BPE tokenizer, GPT-2 inference (Milestone ①)

2026-05-21 22:04:00 +08:00

09-kv-cache.md

phase 9: KV cache + autoregressive generation

2026-05-21 23:39:41 +08:00

10-qwen3.md

fix: comprehensive review + 14 bug fixes + Phase 12/14 overhaul

2026-05-22 17:53:28 +08:00

11-paged-attention.md

docs: Phase 14 design doc + benchmark, fix Phase 11/12 honesty

2026-05-22 18:51:29 +08:00

12-continuous-batching.md

docs: Phase 14 design doc + benchmark, fix Phase 11/12 honesty

2026-05-22 18:51:29 +08:00

13-http-api.md

docs: split Phase 12 and Phase 13 into separate design documents

2026-05-22 13:15:27 +08:00

14-flash-attention.md

docs: Phase 14 design doc + benchmark, fix Phase 11/12 honesty

2026-05-22 18:51:29 +08:00

15-performance.md

docs: Phase 15 design doc + benchmark report

2026-05-23 00:39:27 +08:00

16-llama-cpp-comparison.md

docs: update llama.cpp comparison with 8192 results (OOM fixed)

2026-05-28 21:32:14 +08:00

17-tensor-parallelism.md

docs: Phase 17 tensor parallelism design

2026-05-29 11:10:03 +08:00

18-pipeline-parallelism.md

docs: Phase 18 pipeline parallelism — design + benchmark results

2026-05-29 18:57:09 +08:00

19-gpt-oss-moe.md

docs: fill the Phase 19 gap, refresh README/roadmap to actual state

2026-06-12 17:02:59 +08:00

20-sparse-moe.md

moe: sparse top-k decode — compute only routed experts (1.8x, beats llama TP=2)

2026-06-12 16:29:10 +08:00

21-cuda-graph-decode.md

docs: Phase 21 — decode CUDA graph + GPU argmax results

2026-06-12 20:12:37 +08:00

22-speculative-decoding.md

speculative: Qwen3 draft-model v0 with paged verify parity

2026-07-01 14:16:30 +08:00

23-speculative-verify-parity.md

speculative: Qwen3 draft-model v0 with paged verify parity

2026-07-01 14:16:30 +08:00

24-speculative-batched-verify.md

docs: Phase 24 investigation notes and revised speedup plan

2026-07-01 15:35:11 +08:00

25-speculative-methods-comparison.md

docs: Phase 25 — three speculative-decoding paradigms compared

2026-07-01 16:53:37 +08:00