[NV] DeepSeek-V4-Pro trtllm disagg receipes for STP and MTP

[richardhuo-nv]

Summary

Adds disaggregated TRT-LLM inference benchmarks for DeepSeek-V4-Pro in MXFP4 precision
on GB300 via the Dynamo frontend, covering both Standard Token Processing (STP) and
Multi-Token Prediction (MTP) configurations.

Changes

.github/configs/nvidia-master.yaml — 819 lines added

• New config key dsv4-fp4-gb300-dynamo-trt (STP): 27 scenarios across two sequence length regimes
  • ISL 1024 / OSL 1024: 14 concurrency points (4 → 16384)
  • ISL 8192 / OSL 1024: 13 concurrency points (4 → 4301)
• New config key dsv4-fp4-gb300-dynamo-trt-mtp (MTP): 27 scenarios across two sequence length regimes
  • ISL 1024 / OSL 1024: 14 concurrency points (8 → 8192)
  • ISL 8192 / OSL 1024: 13 concurrency points (8 → 4301)
• Container: nvcr.io#nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1
• Recipes sourced from NVIDIA/srt-slurm branch sa-submission-q2-2026
• Disaggregated prefill/decode topology with varying num-worker, tp, ep, and
  dp-attn per scenario

runners/launch_gb300-nv.sh — 7 lines added

• New branch for framework=dynamo-trt + model_prefix=dsv4
• Overrides SRT_SLURM_MODEL_PREFIX to deepseek-ai/DeepSeek-V4-Pro to match the
  recipe's model.path key (HuggingFace ID format rather than short prefix)
• Clones NVIDIA/srt-slurm and checks out sa-submission-q2-2026

perf-changelog.yaml — 12 lines added

• Documents new config keys, concurrency coverage, container, and recipe source

---

Note

Low Risk
Changes are benchmark YAML, changelog, and CI runner branching only; no application runtime or auth logic.

Overview
Adds GB300 disaggregated TRT-LLM (Dynamo) benchmark coverage for DeepSeek-V4-Pro MXFP4 via two new keys in nvidia-master.yaml: dsv4-fp4-gb300-dynamo-trt (STP) and dsv4-fp4-gb300-dynamo-trt-mtp (MTP with spec-decoding: mtp). Each key defines fixed-seq-len scenarios for ISL 1024/OSL 1024 and ISL 8192/OSL 1024, with concurrency sweeps and per-point prefill/decode topology (num-worker, tp, ep, dp-attn) wired to CONFIG_FILE recipes on NVIDIA/srt-slurm branch sa-submission-q2-2026, using container nvcr.io#nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1.

launch_gb300-nv.sh gains a dynamo-trt + dsv4 path that clones that branch and overrides SRT_SLURM_MODEL_PREFIX to deepseek-ai/DeepSeek-V4-Pro so srt-slurm model aliases match the recipe’s HuggingFace-style model.path (instead of the short dsv4 prefix used elsewhere). perf-changelog.yaml documents the new config keys and coverage.

^{Reviewed by Cursor Bugbot for commit 4ebf43a. Bugbot is set up for automated code reviews on this repo. Configure here.}