CD perf-instability-cookbook
Cookbook of how TensorRT-LLM performance *instabilities* happen and how they were mitigated, distilled from perf-instability commits on main. Consult when a metric is unstable across runs, iterations, or ranks (first-iter JIT stall spikes, cross-run throughput variance, non-deterministic hangs or logs, NCCL negotiation flakiness, perf-test read-after-write races) to match the symptom against known instability patterns and their precedent mitigations. Distinct from `perf-regression-cookbook` (known-good → bad transitions) and `perf-optimization-casebook` (offensive optimizations): here perf is *variable* rather than uniformly slower. Each pattern precisely tracks all related commit hashes and NVBug IDs. Also consult when reviewing a change that adds JIT compilation, opportunistic collectives, or GC-managed collective resources.
Cookbook of how TensorRT-LLM performance instabilities happen and how they were mitigated, distilled from perf-instability commits on main.
As a process D 43/100 · Unfinished process — References files that are not bundled: references/<module>/<slug>.md
How to improve
- The SKILL.md body is over 5,000 tokens: move reference detail into references/ and load it when needed.
- The text references files that are not there: add them or drop the references.
- 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 · 1
✓ No critical or high findings
Medium and low: 1
-
low Secrets in code
secret-high-entropy-tokenreferences/communication/nccl-symmetric-off-on-gb10-spark.md:39High-entropy token-like string (may be an id, hash or a credential)Nemo…FP4 at `--tp_size 2` work on 1.2.0rc6
Files scanned: 30. Evidence is masked. Grey chips explain why severity was lowered.
Against the Agent Skills spec
- warning
body-longSKILL.md body ≈ 5372 tokens (recommended < 5000); move details to references/ - warning
missing-refreference to a missing file: references/<module>/<slug>.md - note
edit-residuethe text marks something as outdated (lines 194): check that old rules are not kept next to new ones — the full check reads the text for contradictions
Process rating: all ten parameters 43/100
- 0Tools and files. 1 referenced file(s) missing: 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. 23 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 5372 tokens
- 100Steps. 48 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 (6 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)
- +3Description length 838: 120–800 characters recommended
- +3Output format is not stated: the model decides each time
- +4No input/output examples
- +2Single-language instructions
- +4Structure: 13 headings
- +3Step-by-step instructions: 48 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: 66.