Skip to content

Kimi K3 DSpark: probabilistic draft + block rejection sampling - #2366

Closed
zixi-qi wants to merge 1 commit into
SemiAnalysisAI:mainfrom
zixi-qi:zixi/kimik3-dspark-accept-rate
Closed

Kimi K3 DSpark: probabilistic draft + block rejection sampling#2366
zixi-qi wants to merge 1 commit into
SemiAnalysisAI:mainfrom
zixi-qi:zixi/kimik3-dspark-accept-rate

Conversation

@zixi-qi

@zixi-qi zixi-qi commented Jul 27, 2026

Copy link
Copy Markdown
Contributor

Optimize Kimi-K3 DSpark speculative decoding acceptance rate.

🤖 Generated with Claude Code

Optimize Kimi-K3 DSpark speculative decoding acceptance rate by setting
draft_sample_method=probabilistic and rejection_sample_method=block in
the speculative-config.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@zixi-qi zixi-qi changed the title [NV] kimik3 DSpark: probabilistic draft + block rejection sampling Kimi K3 DSpark: probabilistic draft + block rejection sampling Jul 27, 2026
functionstackx added a commit that referenced this pull request Jul 28, 2026
…ector variant [skip-sweep]

Duplicate of the kimik3_fp4_b300_vllm.sh collector with PR #2366's
speculative-config change applied — draft_sample_method=probabilistic and
rejection_sample_method=block — as a separate script so the baseline and
optimized-acceptance AL curves can be collected side by side.

中文:复制 kimik3_fp4_b300_vllm.sh 收集器并应用 PR #2366 的投机解码配置改动
(draft_sample_method=probabilistic、rejection_sample_method=block),作为
独立脚本,使基线与优化接受率两条 AL 曲线可并行采集。

Co-authored-by: zixi-qi <zixi@inferact.ai>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
functionstackx added a commit that referenced this pull request Jul 28, 2026
…pling AL collector variant / 新增 Kimi-K3 DSpark probabilistic+block 采样 AL 收集器变体脚本 (#2368)

* feat(speedbench): Kimi-K3 DSpark probabilistic+block sampling AL collector variant [skip-sweep]

Duplicate of the kimik3_fp4_b300_vllm.sh collector with PR #2366's
speculative-config change applied — draft_sample_method=probabilistic and
rejection_sample_method=block — as a separate script so the baseline and
optimized-acceptance AL curves can be collected side by side.

中文:复制 kimik3_fp4_b300_vllm.sh 收集器并应用 PR #2366 的投机解码配置改动
(draft_sample_method=probabilistic、rejection_sample_method=block),作为
独立脚本,使基线与优化接受率两条 AL 曲线可并行采集。

Co-authored-by: zixi-qi <zixi@inferact.ai>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(golden-al): add Kimi-K3 DSpark probabilistic+block golden AL curve [skip-sweep]

Golden acceptance-length curve for the probabilistic draft sampling +
block rejection sampling variant (B300, vLLM DSpark,
Inferact/Kimi-K3-DSpark draft, SPEED-Bench coding, thinking_on). Values
verified byte-identical to the speedbench-reference-al-kimik3 artifact of
actions run 30316471205; uniformly higher AL than the baseline curve
(e.g. 3.00 vs 2.87 at k=3, 4.00 vs 3.88 at k=8).

中文:新增 probabilistic 草稿采样 + block 拒绝采样变体的 Kimi-K3 DSpark 黄金
接受长度(AL)曲线(B300、vLLM DSpark、SPEED-Bench coding、thinking_on)。
数值与 actions run 30316471205 产物逐字节一致;全面高于基线曲线
(如 k=3 时 3.00 对 2.87,k=8 时 4.00 对 3.88)。

Co-authored-by: zixi-qi <zixi@inferact.ai>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

---------

Co-authored-by: zixi-qi <zixi@inferact.ai>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
@functionstackx

Copy link
Copy Markdown
Collaborator

thanks @zixi-qi and inferact team for this!

superceded by fa3f2a2 to have 2 different config yamls
1 for engines that support probabilistic draft + block rejection sampling DSpark

another for if they only support greedy

functionstackx added a commit that referenced this pull request Jul 28, 2026
Replace draft_sample_method=probabilistic + rejection_sample_method=block
with rejection_sample_method=synthetic. The golden AL (3.84 at k=7) was
MEASURED under probabilistic+block (#2366/#2368); the benchmark injects it
via synthetic rejection sampling per the golden-AL convention
(golden_al_distribution/README.md), the same pattern as the DSv4 MTP
recipes (rejection_sample_method synthetic + synthetic_acceptance_length).

中文:将 draft_sample_method=probabilistic + rejection_sample_method=block
替换为 rejection_sample_method=synthetic。黄金 AL(k=7 时 3.84)是在
probabilistic+block 下测得(#2366/#2368);基准测试按黄金 AL 约定通过合成
拒绝采样注入该值(golden_al_distribution/README.md),与 DSv4 MTP 配方的
模式一致。

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

Development

Successfully merging this pull request may close these issues.

2 participants