SKILLEMALL.ai
GitHub ↗ RU Sign in

CD perf-regression-cookbook

Cookbook of how TensorRT-LLM performance regressions happen and how they were fixed, distilled from NVBug-fix commits on main. Consult when a perf regression is observed or suspected (throughput/TTFT/ITL drop, perf CI bar failure, memory-footprint growth) to match the symptom against known regression patterns and their precedent fixes. Each pattern precisely tracks all related commit hashes and NVBug IDs. Also consult when reviewing a risky change to check which regression classes it could reintroduce.

NVIDIA/TensorRT-LLM Agent Skills author: NVIDIA 80 files body ≈ 5 841 tokens Open the sourcegithub.com↗ analyzed 5 d ago

Cookbook of how TensorRT-LLM performance regressions happen and how they were fixed, distilled from NVBug-fix commits on main.

As a process D 43/100 · Unfinished process — References files that are not bundled: references/spec-decode/index.md, references/tokenization/index.md, references/<module>/<slug>.md

ProcedureCustomer supporttype and topics are labelled automatically from the skill text
Runs in: Claude / Claude Code ChatGPT / Codex Gemini CLI Cursor GitHub Copilot / VS Code Qwen Code Kimi Code CLI Z.ai (GLM) DeepSeek (Deep Code) Mistral Vibe LangChain (deepagents) OpenClaw Goose OpenCode
JSON
Technical rating
C
84/100
safety, quality, tests
Safety 60%
98
Quality 40%
63
Run on models
none yet
Process rating
D
43/100
Unfinished process
References files that are not bundled: references/spec-decode/index.md, references/tokenization/index.md, references/<module>/<slug>.md
Tools and files w 18
0
Inputs and preconditions w 11
0
When it triggers w 12
20
the three weakest of ten parameters · all ten

How to improve

  1. The SKILL.md body is over 5,000 tokens: move reference detail into references/ and load it when needed.
  2. The text references files that are not there: add them or drop the references.
For the model run — optional
  • Your own cases (evals/evals.json, 4–6 real requests with expected answers): the full check would then run those instead of a model-drafted suite.
  • A spec.yaml with trigger phrases and assertions — a behaviour contract for CI; `skilltest init` writes a template.

Guard findings · 2

✓ No critical or high findings

Medium and low: 2
  • low Secrets in code secret-high-entropy-token references/gemm-and-quantization/fp8-pertensor-quant-slower-than-te.md:31
    High-entropy token-like string (may be an id, hash or a credential) (quoted — discussed, not commanded)
    though TE's `Floa…zer` kernel is faster on H100 —
    quoted
  • low Secrets in code secret-high-entropy-token references/model-definition/vlm-vision-tower-loaded-for-text-only-bench.md:52
    High-entropy token-like string (may be an id, hash or a credential) (quoted — discussed, not commanded)
    `Qwen…ion` / `Qwen…ion`
    quoted

Files scanned: 80. Evidence is masked. Grey chips explain why severity was lowered.

Against the Agent Skills spec

  • warning body-long SKILL.md body ≈ 5841 tokens (recommended < 5000); move details to references/
  • warning missing-ref reference to a missing file: references/spec-decode/index.md
  • warning missing-ref reference to a missing file: references/tokenization/index.md
  • warning missing-ref reference to a missing file: references/<module>/<slug>.md

Process rating: all ten parameters 43/100

Will not run. References files that are not bundled: references/spec-decode/index.md, references/tokenization/index.md, references/<module>/<slug>.md
  • 0Tools and files. 3 referenced file(s) missing: references/spec-decode/index.md, references/tokenization/index.md, references/<module>/<slug>.md
  • 0Inputs and preconditions. Does not say what the process needs to start
  • 20When it triggers. No condition that starts the skill
  • 30Running it twice. 31 mutating operations with no state check
  • 40Result and completion. Does not say what the result is
  • 50Failures and branches. 0 branches, has a failure section
  • 70Execution cost. Instruction body is 5841 tokens
  • 100Steps. 36 steps
  • 100Consistency. Name and required fields are in place
  • 100Progress reporting. Reports progress
  • medium Safety rules and hard prohibitions inside a skill: they belong in the system prompt, here they protect nothing
  • low The response is described with custom markup (7 tags): a typed call is more reliable

Everything here is measured from the skill text rather than judged by a model, so the numbers are checkable. A parameter weighs more when it is a more common reason for the process to stall.

Quality signals

  • +5Description has no quoted example phrases that should trigger the skill
  • +4Description does not say when NOT to use the skill (false activations)
  • +3Output format is not stated: the model decides each time
  • +4No input/output examples
  • +2Single-language instructions
  • +3Description length 507: enough signal without eating the budget
  • +4Structure: 12 headings
  • +3Step-by-step instructions: 36 items
  • +4Reference files are cited in the instructions (1 of 1)
  • +1License stated

Quality base 70; lint remarks subtract, signals add up to 100. Result: 63.