Add DSR1 FP8 H200 Dynamo TRT-LLM configurations#570
Conversation
Expand dsr1-fp8-h200-dynamo-trt section with full configuration set:
- 1k1k MTP configs (c4-c512) with CONFIG_FILE references
- 1k1k STP configs (c4-c512) with CONFIG_FILE references
- 8k1k MTP configs (c4-c512) with CONFIG_FILE references
- 8k1k STP configs (c4-c512) with CONFIG_FILE references
All configs reference recipe YAMLs in srt-slurm-trtllm repo under
recipies/trtllm/h200/{1k1k,8k1k}/{mtp,stp}/
…nsorrtllm-runtime:0.8.0
dep8 = enable_attention_dp: true = dp-attn: true tep8 = enable_attention_dp: false = dp-attn: false
Update MODEL_PATH from /models/dsr1-fp8 (old DeepSeek-R1) to /models/DeepSeek-R1-0528 (new version matching nvidia-master.yaml)
|
You must update |
|
test sweeping again to smoke test @ishandhanani additions |
|
This should not have been merged into 1 PR. Sweeps for SGLang won't pass yet. In the future please allow me to indicate when things are ready to merge into 1 mega PR. It's easier on our end if we can keep things separate |
|
sure np. nevertheless it appears sweeps (without your changes) are still failing? edit: is is bc of "recipies" LMFAO (my bad I didn't mean yall actually had to change this 😭 ) |
- Update SQUASH_FILE to use /data/containers/ with + separators - Strip nvcr.io/ prefix from path to match actual .sqsh filenames - Add CONTAINER_KEY to convert IMAGE to srt-slurm format (nvcr.io#) - Map container key to .sqsh path dynamically in srtslurm.yaml
Use the release branch for Q1 2026 submission instead of main.
LOL we had to fix it at some point. Remnant of me just moving way to fast and not having any spell check in the repo 😆 |
mb guys! |
|
@nlevin-ui left a couple more comments, mainly nits. In general, please just follow the conventions set in #585 if you can. Also see comments in #510 which may also apply to this PR. |
Link each CONFIG_FILE to its source in srt-slurm sa-submission-q1-2026 branch.
Removed outdated DSR1 FP8 H200 Dynamo TRT configuration details and re-added them in a new section.
|
ok. lgtm, thanks for all your hard work on this. |
DSR1 FP8 H200 Dynamo TRT-LLM Disagg