[Klaud Cold][agentic experiment][Variant I] Kimi-K3 B200 agg TP8xPP2 agentic — DSpark spec decoding @ golden AL / Kimi-K3 B200 聚合式 TP8xPP2 智能体实验——DSpark 投机解码（黄金 AL） - #2376

Closed

functionstackx wants to merge 27 commits into

mainfrom

klaud/kimik3-b200-agentic-direct-vllm-dspark

Commits on Jul 27, 2026

feat: add Kimi-K3 MXFP4 B200 aggregated TP8xPP2 Dynamo-vLLM agentic recipe

functionstackx
and
claude
committed
b4fd077
docs: link PR #2355 in changelog entry and MODELS rows

functionstackx
and
claude
committed
4dbbdc8
fix: pin agentic srt-slurm to NVIDIA v1.0.36, dynamo 1.2.1, conc-8 smoke test

functionstackx
and
claude
committed
e8d42a7
fix: drop OpenAI-frontend tool-choice flags from dynamo-vllm worker args

functionstackx
and
claude
committed
c1e2a56
fix: try dynamo wheel 1.2.0.dev20260426 for Kimi-K3 frontend tokenizer

functionstackx
and
claude
committed
ef35fd1
fix: pin dynamo to day-zero Kimi-K3 commit, restore kimi_k3 parser flags

functionstackx
and
claude
committed
be6c56e
fix: use dynamo namespaced --dyn-* kimi_k3 parser args on the worker

functionstackx
and
claude
committed
c6917e6
fix: disable aiperf conv-aware routing (session_control 400-rejected)

functionstackx
and
claude
committed
a26853a
fix: patch kimi-k3 image mamba_hybrid index_fill_ dtype via setup_script

functionstackx
and
claude
committed
f61eafb
feat: agentic experiment D — direct vllm serve via srt-slurm PR #278

functionstackx
and
claude
committed
8fd319c
docs: point changelog and MODELS rows at experiment PR #2359

functionstackx
and
claude
committed
c0ace4d
fix: drop gpu-memory-utilization to 0.90 (flashinfer MoE workspace OOM)

functionstackx
and
claude
committed
862024d
fix: expandable_segments allocator, drop NCCL_CUMEM_ENABLE (prefill OOM)

functionstackx
and
claude
committed
4370988
Revert "fix: expandable_segments allocator, drop NCCL_CUMEM_ENABLE (prefill OOM)"

functionstackx
and
claude
committed
0c5fe11
feat: add VLLM_PREFIX_CACHE_RETENTION_INTERVAL=32768

functionstackx
and
claude
committed
e675dd2
feat: widen agentic conc list to 1/8/16/32

functionstackx
and
claude
committed
4b0c3a4
fix: remove VLLM_PREFIX_CACHE_RETENTION_INTERVAL (K3 scheduler block)

functionstackx
and
claude
committed
479b74b

Commits on Jul 28, 2026

feat: agentic experiment I — variant D + DSpark spec decoding, golden AL

functionstackx
and
claude
committed
245e36e
Merge remote-tracking branch 'origin/main' into klaud/kimik3-b200-agentic-direct-vllm-dspark
functionstackx
committed
ef94613
docs: point changelog and MODELS rows at experiment PR #2376

functionstackx
and
claude
committed
6dc9b4a
feat: label variant I spec-decoding as dspark

functionstackx
and
claude
committed
b3bde33
refactor: name the variant I recipe file by its spec method (dspark)

functionstackx
and
claude
committed
08217d9
Merge remote-tracking branch 'origin/main' into klaud/kimik3-b200-agentic-direct-vllm-dspark
functionstackx
committed
58dcbe1
fix: DSpark synthetic rejection sampling at the golden AL

functionstackx
and
claude
committed
7176883
fix: verify DSpark draft against a pp=1 parallel-config view

functionstackx
and
claude
committed
41dd5a9
feat: transport DSpark aux hidden states across PP stages

functionstackx
and
claude
committed
18d66d5
fix: lift V2 runner's dspark-under-PP guard (aux transport in place)

functionstackx
and
claude
committed
83fc622

Navigation Menu

Uh oh!

[Klaud Cold][agentic experiment][Variant I] Kimi-K3 B200 agg TP8xPP2 agentic — DSpark spec decoding @ golden AL / Kimi-K3 B200 聚合式 TP8xPP2 智能体实验——DSpark 投机解码（黄金 AL） - #2376

Commits on Jul 27, 2026

feat: add Kimi-K3 MXFP4 B200 aggregated TP8xPP2 Dynamo-vLLM agentic recipe

docs: link PR #2355 in changelog entry and MODELS rows

fix: pin agentic srt-slurm to NVIDIA v1.0.36, dynamo 1.2.1, conc-8 smoke test

fix: drop OpenAI-frontend tool-choice flags from dynamo-vllm worker args

fix: try dynamo wheel 1.2.0.dev20260426 for Kimi-K3 frontend tokenizer

fix: pin dynamo to day-zero Kimi-K3 commit, restore kimi_k3 parser flags

fix: use dynamo namespaced --dyn-* kimi_k3 parser args on the worker

fix: disable aiperf conv-aware routing (session_control 400-rejected)

fix: patch kimi-k3 image mamba_hybrid index_fill_ dtype via setup_script

feat: agentic experiment D — direct vllm serve via srt-slurm PR #278

docs: point changelog and MODELS rows at experiment PR #2359

fix: drop gpu-memory-utilization to 0.90 (flashinfer MoE workspace OOM)

fix: expandable_segments allocator, drop NCCL_CUMEM_ENABLE (prefill OOM)

Revert "fix: expandable_segments allocator, drop NCCL_CUMEM_ENABLE (prefill OOM)"

feat: add VLLM_PREFIX_CACHE_RETENTION_INTERVAL=32768

feat: widen agentic conc list to 1/8/16/32

fix: remove VLLM_PREFIX_CACHE_RETENTION_INTERVAL (K3 scheduler block)

Commits on Jul 28, 2026

feat: agentic experiment I — variant D + DSpark spec decoding, golden AL

Merge remote-tracking branch 'origin/main' into klaud/kimik3-b200-agentic-direct-vllm-dspark

docs: point changelog and MODELS rows at experiment PR #2376

feat: label variant I spec-decoding as dspark

refactor: name the variant I recipe file by its spec method (dspark)

Merge remote-tracking branch 'origin/main' into klaud/kimik3-b200-agentic-direct-vllm-dspark

fix: DSpark synthetic rejection sampling at the golden AL

fix: verify DSpark draft against a pp=1 parallel-config view

feat: transport DSpark aux hidden states across PP stages

fix: lift V2 runner's dspark-under-PP guard (aux transport in place)