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

[nlevin-ui]

• New framework: dynamo-vllm (Dynamo frontend + vLLM backend)
• 6 configs: 2x 1k1k + 4x 8k1k with varying prefill/decode ratios
• Recipes sourced from NVIDIA/srt-slurm branch sa-submission-q2-2026
• Runner script updated to clone NVIDIA/srt-slurm and map vLLM container

[github-actions]

Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook

• https://cookbook.sglang.io/
• https://docs.vllm.ai/projects/recipes/en/latest/

If it is not, please create a PR first before we can merge your PR into the master branch. Let's ensure that the documentation is first class such that the entire ML community can benefit from your hard work! Thank you

• https://github.com/vllm-project/recipes
• https://github.com/sgl-project/sgl-cookbook

[claude]

Claude finished @nlevin-ui's task in 1m 19s —— View job

---

Review of PR #1008

• Gather context and read changed files
• Validate model prefixes in config keys
• Validate container image accessibility
• Validate enroot import in launch script
• Validate perf-changelog updated
• Check benchmark script style
• Post review summary

LGTM — no blocking issues found.

Validation summary:

• Container image: vllm/vllm-openai:v0.18.0-cu130 is publicly accessible on Docker Hub ✅
• Enroot import: Script already has enroot import -o $SQUASH_FILE docker://$IMAGE at line 58 ✅
• Perf-changelog: Updated with detailed entry ✅
• Model prefix: kimik2.5 is consistent with existing Kimi K2.5 configs already in nvidia-master.yaml ✅
• Config structure: 6 disaggregated configs (2× 1k1k, 4× 8k1k) with varying prefill/decode ratios look well-structured ✅
• Launch script: Clean branching for dynamo-vllm framework with correct srt-slurm repo (NVIDIA/srt-slurm) and branch (sa-submission-q2-2026) ✅
• Container mapping: "${IMAGE}": ${SQUASH_FILE} at line 172 correctly maps the vLLM Docker image name so srt-slurm recipes can resolve it ✅

[functionstackx]

thanks @nlevin-ui can u debug the failing pr validation?

[cquil11]

@claude Verify that the parameters (EP, DP, TP, etc) in nvidia-master.yaml match those in the linked recipes from CONFIG_FILES. For all added entries.

[Klaud-Cold]

Claude finished @cquil11's task in 2m 18s —— View job

---

Parameter verification: nvidia-master.yaml vs upstream recipes

• Read added entries in nvidia-master.yaml
• Identify all 6 CONFIG_FILE recipe paths
• Fetch each recipe from NVIDIA/srt-slurm branch sa-submission-q2-2026
• Compare EP, DP, TP, num-worker, and concurrency parameters
• Post verification summary

---

All 6 configs match their upstream recipes. I fetched each recipe YAML from NVIDIA/srt-slurm@sa-submission-q2-2026 and verified the parallelism parameters (num-worker, tp, ep, dp-attn) and concurrency lists against the nvidia-master.yaml entries.

The mapping used: tp ↔ gpus_per_prefill/gpus_per_decode, ep ↔ effective expert-parallel size (from data-parallel-size when using DEP, or from tensor-parallel-size when using TEP), num-worker ↔ prefill_workers/decode_workers, dp-attn ↔ whether data-parallel is active.

1. 1k1k/disagg-gb200-1p1d-dep4-dep16.yaml ✅

Parameter	nvidia-master.yaml	Recipe
Prefill num-worker / tp / ep / dp-attn	1 / 4 / 4 / true	workers=1, gpus=4, TP=1, DP=4, EP=true
Decode num-worker / tp / ep / dp-attn	1 / 16 / 16 / true	workers=1, gpus=16, TP=1, DP=16, EP=true
conc-list	[256, 512, 1024, 2048, 3072, 4096]	256x512x1024x2048x3072x4096

2. 1k1k/disagg-gb200-1p4d-dep4-tep4.yaml ✅

Parameter	nvidia-master.yaml	Recipe
Prefill num-worker / tp / ep / dp-attn	1 / 4 / 4 / true	workers=1, gpus=4, TP=1, DP=4, EP=true
Decode num-worker / tp / ep / dp-attn	4 / 4 / 4 / false	workers=4, gpus=4, TP=4, EP=true (no DP)
conc-list	[4, 8, 16, 32, 64, 128]	4x8x16x32x64x128

3. 8k1k/disagg-gb200-1p4d-dep4-tep4.yaml ✅

Parameter	nvidia-master.yaml	Recipe
Prefill num-worker / tp / ep / dp-attn	1 / 4 / 4 / true	workers=1, gpus=4, TP=1, DP=4, EP=true
Decode num-worker / tp / ep / dp-attn	4 / 4 / 4 / false	workers=4, gpus=4, TP=4, EP=true (no DP)
conc-list	[4, 8, 16, 32, 128]	4x8x16x32x128

Note: The 8k1k 1P4D concurrency list intentionally omits 64, matching the upstream recipe exactly. This confirms the recipe author's deliberate choice (as @nlevin-ui stated in the earlier review comment).

4. 8k1k/disagg-gb200-3p1d-dep4-dep16.yaml ✅

Parameter	nvidia-master.yaml	Recipe
Prefill num-worker / tp / ep / dp-attn	3 / 4 / 4 / true	workers=3, gpus=4, TP=1, DP=4, EP=true
Decode num-worker / tp / ep / dp-attn	1 / 16 / 16 / true	workers=1, gpus=16, TP=1, DP=16, EP=true
conc-list	[512, 1024]	512x1024

5. 8k1k/disagg-gb200-5p1d-dep4-dep8.yaml ✅

Parameter	nvidia-master.yaml	Recipe
Prefill num-worker / tp / ep / dp-attn	5 / 4 / 4 / true	workers=5, gpus=4, TP=1, DP=4, EP=true
Decode num-worker / tp / ep / dp-attn	1 / 8 / 8 / true	workers=1, gpus=8, TP=1, DP=8, EP=true
conc-list	[2048]	2048

6. 8k1k/disagg-gb200-6p1d-dep4-dep16.yaml ✅

Parameter	nvidia-master.yaml	Recipe
Prefill num-worker / tp / ep / dp-attn	6 / 4 / 4 / true	workers=6, gpus=4, TP=1, DP=4, EP=true
Decode num-worker / tp / ep / dp-attn	1 / 16 / 16 / true	workers=1, gpus=16, TP=1, DP=16, EP=true
conc-list	[3072, 4096]	3072x4096

Result: 6/6 configs verified — no mismatches found.

---

View job

[nlevin-ui]

thanks @nlevin-ui can u debug the failing pr validation?

@functionstackx can you be more specific? I don't see PR validation failing beyond what I had already fixed prior to your comment. I might just be missing something obvious.

[functionstackx]

thanks @nlevin-ui can u debug the failing pr validation?

@functionstackx can you be more specific? I don't see PR validation failing beyond what I had already fixed prior to your comment. I might just be missing something obvious.

nvm, i see that attempt #2 re-run has fixed it https://github.com/SemiAnalysisAI/InferenceX/actions/runs/24058086986?pr=1008 thanks guys!

[functionstackx]

lgtm passed pr validation https://github.com/SemiAnalysisAI/InferenceX/actions/runs/24058086986?pr=1008