Skip to content

[NVIDIA] Add Kimi K2.5 NVFP4 GB200 disaggregated vLLM benchmarks via Dynamo #1008

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
functionstackx merged 2 commits into from
Apr 7, 2026
Merged

[NVIDIA] Add Kimi K2.5 NVFP4 GB200 disaggregated vLLM benchmarks via Dynamo #1008

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
3e358a8
Add Kimi K2.5 NVFP4 GB200 disaggregated vLLM benchmarks via Dynamo
tanmayv25 Apr 6, 2026
5123f80
Fix misleading error message for unsupported dynamo-vllm combinations
tanmayv25 Apr 6, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
101 changes: 101 additions & 0 deletions .github/configs/nvidia-master.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6646,3 +6646,104 @@
tp: 8
ep: 1
dp-attn: false

kimik2.5-fp4-gb200-dynamo-vllm:
image: vllm/vllm-openai:v0.18.0-cu130
model: nvidia/Kimi-K2.5-NVFP4
model-prefix: kimik2.5
runner: gb200
precision: fp4
framework: dynamo-vllm
multinode: true
disagg: true
seq-len-configs:
- isl: 1024
osl: 1024
search-space:
- conc-list: [256, 512, 1024, 2048, 3072, 4096]
prefill:
num-worker: 1
tp: 4
ep: 4
dp-attn: true
additional-settings:
# https://github.com/NVIDIA/srt-slurm/blob/sa-submission-q2-2026/recipes/vllm/kimi-k2.5/1k1k/disagg-gb200-1p1d-dep4-dep16.yaml
- "CONFIG_FILE=recipes/vllm/kimi-k2.5/1k1k/disagg-gb200-1p1d-dep4-dep16.yaml"
decode:
num-worker: 1
tp: 16
ep: 16
dp-attn: true
- conc-list: [4, 8, 16, 32, 64, 128]
prefill:
num-worker: 1
tp: 4
ep: 4
dp-attn: true
additional-settings:
# https://github.com/NVIDIA/srt-slurm/blob/sa-submission-q2-2026/recipes/vllm/kimi-k2.5/1k1k/disagg-gb200-1p4d-dep4-tep4.yaml
- "CONFIG_FILE=recipes/vllm/kimi-k2.5/1k1k/disagg-gb200-1p4d-dep4-tep4.yaml"
decode:
num-worker: 4
tp: 4
ep: 4
dp-attn: false
- isl: 8192
osl: 1024
search-space:
- conc-list: [4, 8, 16, 32, 128]

Check warning on line 6694 in .github/configs/nvidia-master.yaml

View check run for this annotation

Claude / Claude Code Review

Missing concurrency 64 in 8k1k 1P4D sweep 

The 8k1k 1P4D search-space has `conc-list: [4, 8, 16, 32, 128]`, skipping 64, while the structurally identical 1k1k 1P4D config has `[4, 8, 16, 32, 64, 128]`. This creates a gap in the performance curve at concurrency=64 for the low-latency 8k1k configuration and appears to be a copy-paste oversight.
Comment thread
nlevin-ui marked this conversation as resolved.
prefill:
num-worker: 1
tp: 4
ep: 4
dp-attn: true
additional-settings:
# https://github.com/NVIDIA/srt-slurm/blob/sa-submission-q2-2026/recipes/vllm/kimi-k2.5/8k1k/disagg-gb200-1p4d-dep4-tep4.yaml
- "CONFIG_FILE=recipes/vllm/kimi-k2.5/8k1k/disagg-gb200-1p4d-dep4-tep4.yaml"
decode:
num-worker: 4
tp: 4
ep: 4
dp-attn: false
- conc-list: [512, 1024]
prefill:
num-worker: 3
tp: 4
ep: 4
dp-attn: true
additional-settings:
# https://github.com/NVIDIA/srt-slurm/blob/sa-submission-q2-2026/recipes/vllm/kimi-k2.5/8k1k/disagg-gb200-3p1d-dep4-dep16.yaml
- "CONFIG_FILE=recipes/vllm/kimi-k2.5/8k1k/disagg-gb200-3p1d-dep4-dep16.yaml"
decode:
num-worker: 1
tp: 16
ep: 16
dp-attn: true
- conc-list: [2048]
prefill:
num-worker: 5
tp: 4
ep: 4
dp-attn: true
additional-settings:
# https://github.com/NVIDIA/srt-slurm/blob/sa-submission-q2-2026/recipes/vllm/kimi-k2.5/8k1k/disagg-gb200-5p1d-dep4-dep8.yaml
- "CONFIG_FILE=recipes/vllm/kimi-k2.5/8k1k/disagg-gb200-5p1d-dep4-dep8.yaml"
decode:
num-worker: 1
tp: 8
ep: 8
dp-attn: true
- conc-list: [3072, 4096]
prefill:
num-worker: 6
tp: 4
ep: 4
dp-attn: true
additional-settings:
# https://github.com/NVIDIA/srt-slurm/blob/sa-submission-q2-2026/recipes/vllm/kimi-k2.5/8k1k/disagg-gb200-6p1d-dep4-dep16.yaml
- "CONFIG_FILE=recipes/vllm/kimi-k2.5/8k1k/disagg-gb200-6p1d-dep4-dep16.yaml"
decode:
num-worker: 1
tp: 16
ep: 16
dp-attn: true
13 changes: 13 additions & 0 deletions perf-changelog.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1265,3 +1265,16 @@
description:
- "Add Qwen3.5-397B-A17B-FP8 H200 SGLang MTP (EAGLE speculative decoding)"
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1001

- config-keys:
- kimik2.5-fp4-gb200-dynamo-vllm
description:
- "Add Kimi K2.5 NVFP4 GB200 disaggregated multinode vLLM benchmark via Dynamo frontend"
- "Image: vllm/vllm-openai:v0.18.0-cu130"
- "Model: nvidia/Kimi-K2.5-NVFP4 with NixlConnector KV transfer, FLASHINFER_MLA attention"
- "1k1k configs: high-throughput DEP (1P1D dep4/dep16), low-latency TEP (1P4D dep4/tep4)"
- "8k1k configs: low-latency TEP (1P4D), mid-curve DEP (3P1D dep16), high-throughput (5P1D dep8, 6P1D dep16)"
- "Recipes sourced from NVIDIA/srt-slurm branch sa-submission-q2-2026"
- "New framework: dynamo-vllm (Dynamo frontend + vLLM backend)"
- "Runner script updated to clone NVIDIA/srt-slurm and map vLLM container image"
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1008
21 changes: 18 additions & 3 deletions runners/launch_gb200-nv.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,6 +34,14 @@
echo "Unsupported model prefix: $MODEL_PREFIX. Supported prefixes are: gptoss or dsr1"
exit 1
fi
elif [[ $FRAMEWORK == "dynamo-vllm" ]]; then
if [[ $MODEL_PREFIX == "kimik2.5" && $PRECISION == "fp4" ]]; then
export MODEL_PATH="/mnt/lustre01/models/kimi-k2.5-nvfp4"
export SRT_SLURM_MODEL_PREFIX="kimi-k2.5-nvfp4"
else
echo "Unsupported model prefix: $MODEL_PREFIX. Supported prefixes for dynamo-vllm: kimik2.5"
exit 1

Check warning on line 43 in runners/launch_gb200-nv.sh

View check run for this annotation

Claude / Claude Code Review

Misleading error message omits precision constraint for dynamo-vllm 

The error message on line 42 says "Supported prefixes for dynamo-vllm: kimik2.5" but the guard condition checks both MODEL_PREFIX=="kimik2.5" AND PRECISION=="fp4", so a user passing MODEL_PREFIX=kimik2.5 with PRECISION=fp8 will see a self-contradictory message that lists kimik2.5 as both unsupported and supported, hiding the real problem (unsupported precision). The message should say "Supported combinations for dynamo-vllm: kimik2.5/fp4" to clearly communicate the constraint.
Comment thread
nlevin-ui marked this conversation as resolved.
Outdated
fi
else
export MODEL_PATH=$MODEL
fi
Expand Down Expand Up @@ -112,9 +120,15 @@
rm -rf "$SRT_REPO_DIR"
fi

git clone https://github.com/ishandhanani/srt-slurm.git "$SRT_REPO_DIR"
cd "$SRT_REPO_DIR"
git checkout sa-submission-q1-2026
if [[ $FRAMEWORK == "dynamo-vllm" ]]; then
git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
cd "$SRT_REPO_DIR"
git checkout sa-submission-q2-2026
else
git clone https://github.com/ishandhanani/srt-slurm.git "$SRT_REPO_DIR"
cd "$SRT_REPO_DIR"
git checkout sa-submission-q1-2026
fi

echo "Installing srtctl..."
curl -LsSf https://astral.sh/uv/install.sh | sh
Expand Down Expand Up @@ -155,6 +169,7 @@
containers:
dynamo-trtllm: ${SQUASH_FILE}
dynamo-sglang: ${SQUASH_FILE}
"${IMAGE}": ${SQUASH_FILE}
nginx-sqsh: ${NGINX_SQUASH_FILE}
EOF

Expand Down
Loading