| name | ai-inference-recipe |
| description | Creates single-point srt-slurm recipes from NVIDIA AI Inference benchmark rows. Use when the user mentions ai-inference, NVIDIA inference performance pages, benchmark rows, or asks to create a non-sweep recipe from model/GPU/framework/sequence/concurrency details. |
AI Inference Recipe
Use this skill to turn a user description of an NVIDIA AI Inference benchmark row into a local single-point srt-slurm recipe.
In this repository, single-point means the generated recipe may keep a sweep block, but every sweep key has exactly one value so it expands to one run.
Workflow
- Extract the benchmark row fields from the user request:
model, hw, framework, sequence, precision, tp, ep, conc, spec_decoding, and date when available.
- Prefer the raw Markdown source unless the user provided a local copy:
https://developer.nvidia.com/deep-learning-performance-training-inference/ai-inference.md
- Run
scripts/map_inference_row_to_recipe.py with enough --where FIELD=VALUE filters to select exactly one row.
- Emit the recipe with
--emit-pruned and --single-point.
- Read the JSON report. If multiple rows matched, ask for the missing discriminator. If cautions remain, tell the user what was inferred or left unchanged.
Command Template
python scripts/map_inference_row_to_recipe.py \
--where model=DeepSeek-R1 \
--where hw=B300 \
--where sequence=1K/1K \
--where framework=DYNAMO-TRT \
--where conc=271 \
--emit-pruned tmp/deepseek-r1-b300-dynamo-trt-1k1k-c271.yaml \
--single-point \
--json
Use --sweep-value KEY=VALUE for sweep keys that are not present in the NVIDIA row but are specified by the user, such as pp, dp, or framework-specific tuning keys.
Rules
- Do not hand-map rows to recipes when the script can do it.
- Do not modify the source recipe when producing a one-off recipe; write a new path with
--emit-pruned.
- Treat
ai-inference and ai-inference.md as equivalent user references, but prefer ai-inference.md for row selection.
- If
--single-point collapses an unspecified sweep key to the first value, surface that caution to the user.