inference/research
verifiedhigh · 0.90minorupdated 2026-10-04

TensorRT-LLM has no paper - the absence is verified by an empty title search, and its kernels are documented in this repo as records instead

[minor] toolchain

Fields

id
gotchas/meta-tensorrt-llm-has-no-paper-do-not-invent-one
Class
toolchain
Severity
minor
Concepts
tensorrt-llmnvidiatrtllmarxivprovenancecitation
Symptom
A reader auditing this repo for inference coverage sees that TensorRT-LLM - one of the two or three production inference engines on NVIDIA hardware - has engine, observability and multiple FLOP/kernel records but NO entry in data/papers/. The temptation is to go looking for an arXiv ID, and papers that merely benchmark AGAINST TensorRT-LLM are plentiful enough to produce a plausible-looking wrong answer.
Root cause
None. TensorRT-LLM has not published a system paper. Verified 2026-10-04: a TITLE search (search_query=ti:"TensorRT-LLM") against the arXiv API returns opensearch:totalResults 0 - an entirely empty feed. No title, no preprint, no tech report under that name.
Workaround
Cite the repository and NVIDIA's product documentation, which is what TensorRT-LLM itself does. In this repo the substance is already carried by records: engines/tensorrt-llm, obs-tensorrt-llm-observability, and the FLOP records describing its grouped-GEMM, fused-MoE, MLA and quantization kernels. A future pass finding new TensorRT-LLM kernel work should extend those records or add an engine/release-notes source, never a paper.
notes
This is the SECOND engine negative alongside meta-llama-cpp-has-no-paper-do-not-invent-one, and the pair is worth keeping together because the asymmetry is the lesson. vLLM/PagedAttention, Orca, SGLang, FlashInfer, DeepSpeed, PowerInfer, Splitwise and vAttention all DID publish. llama.cpp and TensorRT-LLM - the two engines most likely to appear in a consumer or single-vendor deployment - did not. So 'no paper' is not evidence of a coverage gap in general; it is a per-project fact that must be checked, never assumed in either direction. The empty-feed result is recorded rather than paraphrased as 'I could not find one', because 'no result' and 'search failed' are different claims and only the first is evidence. Note that this repo's FLOP records DO cite DeepGEMM's kernels and that is legitimate: a kernel can be documented from source and release notes without the project having a paper, which is the same discipline applied consistently rather than as an exception.

Notes

This is the SECOND engine negative alongside meta-llama-cpp-has-no-paper-do-not-invent-one, and the pair is worth keeping together because the asymmetry is the lesson. vLLM/PagedAttention, Orca, SGLang, FlashInfer, DeepSpeed, PowerInfer, Splitwise and vAttention all DID publish. llama.cpp and TensorRT-LLM - the two engines most likely to appear in a consumer or single-vendor deployment - did not. So 'no paper' is not evidence of a coverage gap in general; it is a per-project fact that must be checked, never assumed in either direction. The empty-feed result is recorded rather than paraphrased as 'I could not find one', because 'no result' and 'search failed' are different claims and only the first is evidence. Note that this repo's FLOP records DO cite DeepGEMM's kernels and that is legitimate: a kernel can be documented from source and release notes without the project having a paper, which is the same discipline applied consistently rather than as an exception.

Cross-references

Affects (2)