chore: upgrade h200 gptoss to latest trtllm

[cquil11]

No description provided.

[claude]

Claude finished @cquil11's task in 51s —— View job

---

Review of PR #854

• Gather context and read changed files
• Validate master config and perf-changelog
• Validate container image accessibility
• Validate model prefix
• Review benchmark script changes
• Post review summary

LGTM — no blocking issues found.

This is a clean upgrade of the H200 GPT-OSS TensorRT-LLM config from the dev image (gpt-oss-dev) to an official release candidate (v1.3.0rc5). The sed workaround removal and enable_block_reuse default change are both reasonable cleanups enabled by the new container version. Perf changelog is properly updated.

[functionstackx]

hi @jgangani

the performance is wildly worse on the latest trtllm 1.3 rc, can u take a look? nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc5

[jgangani]

@functionstackx thanks for the perf diff. Looking at them. will get back to you

[cquil11]

@jgangani update?

[functionstackx]

@jgangani we at semianalysis gonna merge this with this tag by monday

[jgangani]

@cquil11 @functionstackx Sorry for dropping this without an update. Reason for delay is that TRTLLM rc5 container pytorch update automatically updated Triton to 3.5.1 which had regression for MoE activation kernel. We believe this has been fixed later in Triton 3.6. We are in the process of releasing a new periodic container with pytorch and triton update by next week. This week is short work week at Nvidia (we are off TH-FR). I hope we can hold off merging this until then.