dsv4-b300-sglang-mtp: restore TP4 and DP-attn search-space entries#1180
dsv4-b300-sglang-mtp: restore TP4 and DP-attn search-space entries#1180
Conversation
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook If it is not, please create a PR first before we can merge your PR into the master branch. Let's ensure that the documentation is first class such that the entire ML community can benefit from your hard work! Thank you PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. If re-running failed jobs is attempted, PR authors are responsible for ensuring it passes. See GitHub's docs on re-running failed jobs: https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow As a rule of thumb, generally, PR authors should request a review & get a PR approval from the respective companies' CODEOWNERS before requesting a review from core maintainers. If additional help is needed, PR authors can reach out to core maintainers over Slack. |
There was a problem hiding this comment.
LGTM, mechanical YAML config restoration.
Extended reasoning...
Overview
This PR adds 4 lines to .github/configs/nvidia-master.yaml, restoring two search-space entries (TP4 ep=1 conc 16-128, and TP4 ep=4 dp-attn conc 64-512) under dsv4-fp4-b300-sglang-mtp for both the 1k1k and 8k1k seq-len-configs. The entries follow the exact same schema as the existing TP8 entry and other entries throughout the file.
Security risks
None. This is a benchmark/CI search-space configuration file with no security-sensitive content (no auth, secrets, permissions, or external execution paths affected).
Level of scrutiny
Low. This is a pure config restoration — the PR description explicitly states these entries were previously removed in commit 287ef26 and are being put back. No logic changes, no new code paths.
Other factors
- Single-file YAML diff with 4 added lines, no removals.
- Bug hunting system found no issues.
- Entries follow established schema patterns visible elsewhere in the file.
|
/sweep test-config --config-files .github/configs/nvidia-master.yaml --config-keys dsv4-fp4-b300-sglang-mtp |
|
@yhyang201 Kicking off a sweep. Run: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/24962086988 |
|
|
|
/sweep test-config --config-files .github/configs/nvidia-master.yaml --config-keys dsv4-fp4-b300-sglang-mtp |
|
@yhyang201 Kicking off a sweep. Run: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/24962803567 |
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
Hi @yhyang201 please hold off on triggering sweeps til we get CI unblocked |
|
/sweep test-config --config-files .github/configs/nvidia-master.yaml --config-keys dsv4-fp4-b300-sglang-mtp |
|
@Qiaolin-Yu Kicking off a sweep. Run: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/24977935266 |
| export SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE=1 | ||
| export SGLANG_OPT_FIX_HASH_MEGA_MOE=1 | ||
| # DP-attn path: flashinfer_mxfp4 + DP-attn (covers conc 16-256). | ||
| export SGLANG_OPT_SWA_EVICT_DROP_PAGE_MARGIN=1 |
There was a problem hiding this comment.
This is fine for now since we are in early stages of v4 support, but eventually we'd like these to just be default settings based on scenario in engine
| --disable-flashinfer-autotune | ||
| --deepep-config "$DEEPEP_CONFIG" | ||
| --cuda-graph-max-bs 256 | ||
| --schedule-conservativeness 2 |
There was a problem hiding this comment.
Can you explain why --schedule-conservativeness 2?
There was a problem hiding this comment.
Still can be default, you can remove it
Summary
Test plan
🤖 Generated with Claude Code