[Doc] Broken links in READMEs referencing files removed by legacy TensorRT backend cleanup

[jlonge4]

Summary

I scanned all 123 README files in the repo — 503 links total (236 repo-relative, 224 unique external URLs) — and found broken links falling into two groups:

• Mechanical link errors (typos, renames, wrong relative depth) — fixed in [None][doc] Fix broken Dynamo TensorRT LLM deployment link in DeepSeek-V3 README #18378 and [None][doc] Fix broken links across example and test READMEs #18379.
• Stale references to intentionally deleted files — listed below. These can't be fixed with a link swap, because the surrounding prose describes workflows and files that no longer exist. Each needs a maintainer decision about the intended replacement, so I'm filing rather than guessing.

Every entry below was verified: repo-relative paths were resolved against the working tree, and external URLs were probed for their HTTP status.

---

1. Legacy TensorRT example removal (#15763)

examples/models/core/llama/ (with convert_checkpoint.py and its README) and examples/run.py were removed, but two model READMEs still walk users through them as the primary workflow.

examples/models/core/exaone/README.md

Line	Broken target
397	../llama/convert_checkpoint.py
537	../../../run.py
561	../llama/README.md

examples/models/contrib/hyperclovax/README.md

Line	Broken target
93	../../core/llama/convert_checkpoint.py
233	../../../run.py
257	../../../../tensorrt_llm/models/llama/model.py
258	../../core/llama/README.md

These READMEs are documenting the removed TensorRT checkpoint-conversion flow end to end, so they likely need rewriting against the PyTorch backend (or an explicit deprecation note).

2. Legacy backend test/example removal (#15918, #16369, #16610)

tests/integration/defs/accuracy/README.md — test_llm_api.py was removed in #15918; examples/summarize.py in #16610.

Line	Broken target
33	../../../../examples/summarize.py
35	./test_llm_api.py
141	./test_llm_api.py
270	./test_llm_api.py

Lines 35/141/270 all read "test_llm_api.py or test_llm_api_pytorch.py" — presumably the first half should just be dropped. Line 33 describes CLI-flow tasks "calling a standalone script like summarize.py", which no longer has a referent.

cpp/tests/README.md — test_e2e.py was removed in #16369, but line 9 still documents it alongside test_multi_gpu.py, and the pytest ... --collect-only example block below it references the same file.

Line	Broken target
9	../../tests/integration/defs/cpp/test_e2e.py

3. QA perf test lists consolidated

tests/integration/defs/perf/README_release_test.md — sections 3.1/3.2 list four test cycles; three no longer exist.

Line	Broken target	Status
115	../../test_lists/qa/llm_perf_cluster.yml	Merged into core (#9714)
117	../../test_lists/qa/llm_perf_nim.yml	Removed (#10657)
122	../../test_lists/qa/llm_perf_sanity.yml	Merged into core (#14952)

(The fourth, llm_perf_full.yml → llm_perf_core.yml, was a straight rename and is fixed in #18379.) Since three of the four cycles collapsed into llm_perf_core.yml, that section probably wants rewriting rather than relinking.

4. Referenced file never existed

jenkins/scripts/perf/README.md:327 points at tests/scripts/perf-sanity/README.md for a "Step-by-Step: Adding or Re-enabling Disaggregated Perf Sanity Tests" guide. Two problems: the relative depth is wrong (../../ resolves to jenkins/tests/...), and no README exists at tests/scripts/perf-sanity/ at any depth — that directory contains only aggregated/, cache_transceiver_precheck/, disaggregated/, and visual_gen/. Either the guide was never committed or it lives somewhere else.

5. External URLs returning 404

examples/auto_deploy/README.md:116 — https://github.com/NVIDIA/Model-Optimizer/blob/main/examples/llm_autodeploy/README.md (404). The llm_autodeploy example directory no longer exists in Model-Optimizer. For "generating quantized model checkpoint" the closest current example looks like examples/hf_ptq, but that's a semantic call for someone who knows the intended flow.

NVIDIA/llm-compiler (404) — referenced as the standalone "Paragraf" package. This may simply be unpublished, in which case it's intentional and can be ignored; flagging it in case the repo was meant to be public by now.

File	Line(s)
examples/auto_deploy/README.md	5
examples/auto_deploy/paragraf/README.md	22, 25, 32, 34

---

Checked and deliberately excluded

These came up in the scan but are not bugs:

• http://localhost:16686/ — Jaeger UI, intentionally local.
• artifactory.nvidia.com (×2) — NVIDIA-internal, unreachable from public networks.
• huggingface.co/naver/hyperclovax — returns 401, a gated model rather than a dead link.
• en.wikipedia.org/wiki/Power_(statistics) — false positive from my own link parser truncating at the closing paren.
• people.eecs.berkeley.edu/~hendrycks/data.tar — the entire host was unreachable from my network, including its root, so I could not verify this either way. Worth a second look by someone who can reach it.

Notes

Happy to send PRs for any of these if maintainers can confirm the intended replacements — I held off here specifically because each one changes documented workflow content rather than just a URL.

[MannXo]

I'd like to give this a shot if it's unassigned.