You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I would like to help with this. I will draft a TRT-LLM inference recipe under inference_benchmarks/ (model export, engine build, and how it aligns with the torch/NIM paths in cosmos3-super-generator.md). Note upfront: I do not have A100/TRT-LLM hardware here, so I can document and script the recipe but cannot reproduce the benchmark numbers — maintainers will need to validate the exact latencies. Hope that is still useful; happy to adjust scope.
Update: while preparing this, I found that PR #384 (merged 2026-10-21, two days after this issue was filed) removed the entire inference_benchmarks/ directory — including cosmos3-super-generator.md — from this public repo, syncing its tree with nvidia-cosmos/cosmos-private. So there is no longer a public file for the TRT-LLM rows (PBR #308000) to be documented against.
Given that, the TRT-LLM recipe this issue asks for can realistically only come from the NVIDIA team with access to the private repo / internal benchmark runs. I do not have A100/TRT-LLM hardware, so I cannot independently reproduce the 0.2s/14s/66s numbers. Closing my involvement here unless maintainers want a community-side TRT-LLM export/serve guide added somewhere under docs/ — happy to do that if useful.
Activity
[-]TRT results in the benchmark[/-][+]TRT-LLM results in the benchmark[/+]akushonkamen commentedon Oct 22, 2026
I would like to help with this. I will draft a TRT-LLM inference recipe under
inference_benchmarks/(model export, engine build, and how it aligns with the torch/NIM paths in cosmos3-super-generator.md). Note upfront: I do not have A100/TRT-LLM hardware here, so I can document and script the recipe but cannot reproduce the benchmark numbers — maintainers will need to validate the exact latencies. Hope that is still useful; happy to adjust scope.akushonkamen commentedon Oct 22, 2026
Update: while preparing this, I found that PR #384 (merged 2026-10-21, two days after this issue was filed) removed the entire
inference_benchmarks/directory — includingcosmos3-super-generator.md— from this public repo, syncing its tree withnvidia-cosmos/cosmos-private. So there is no longer a public file for the TRT-LLM rows (PBR #308000) to be documented against.Given that, the TRT-LLM recipe this issue asks for can realistically only come from the NVIDIA team with access to the private repo / internal benchmark runs. I do not have A100/TRT-LLM hardware, so I cannot independently reproduce the 0.2s/14s/66s numbers. Closing my involvement here unless maintainers want a community-side TRT-LLM export/serve guide added somewhere under
docs/— happy to do that if useful.