Remove moe_dense_tp_size for TP configs

[kyleliang-nv]

Summary by CodeRabbit

• Chores
  • Updated model serving configurations across multiple hardware and precision variants to streamline settings by removing explicit configuration overrides, allowing system defaults to apply.

[coderabbitai]

📝 Walkthrough

Walkthrough

This pull request systematically removes moe-dense-tp-size: 1 configuration entries from the prefill and decode sections of sglang_config across 20 recipe YAML files spanning GB200 and GB300 models with various floating-point precisions and sequence length configurations.

Changes

Cohort / File(s)	Summary
GB200 FP4 Configuration Cleanup recipes/gb200-fp4/1k1k/1p2d-mtp.yaml, recipes/gb200-fp4/1k1k/low-latency.yaml, recipes/gb200-fp4/1k8k/low-latency.yaml, recipes/gb200-fp4/8k1k/low-latency.yaml	Removed moe-dense-tp-size: 1 from sglang_config prefill and decode sections.
GB200 FP8 Configuration Cleanup recipes/gb200-fp8/1k1k/low-latency-mtp.yaml, recipes/gb200-fp8/1k1k/low-latency.yaml, recipes/gb200-fp8/1k8k/low-latency-mtp.yaml, recipes/gb200-fp8/1k8k/low_latency.yaml, recipes/gb200-fp8/8k1k/low-latency-mtp.yaml, recipes/gb200-fp8/8k1k/low-latency.yaml	Removed moe-dense-tp-size: 1 from sglang_config prefill and decode sections across all FP8 variants.
GB300 FP4 Configuration Cleanup recipes/gb300-fp4/128k8k/5-low-latency.yaml, recipes/gb300-fp4/1k1k/low_latency.yaml, recipes/gb300-fp4/1k8k/low-latency.yaml, recipes/gb300-fp4/8k1k/low_latency.yaml	Removed moe-dense-tp-size: 1 from sglang_config prefill and decode sections; 128k8k file also includes trailing newline fix.
GB300 FP8 Configuration Cleanup recipes/gb300-fp8/1k1k/mtp/low-latency.yaml, recipes/gb300-fp8/1k1k/stp/low-latency.yaml, recipes/gb300-fp8/1k8k/mtp/low-latency.yaml, recipes/gb300-fp8/1k8k/stp/low-latency.yaml, recipes/gb300-fp8/8k1k/mtp/low-latency.yaml, recipes/gb300-fp8/8k1k/stp/low-latency.yaml	Removed moe-dense-tp-size: 1 from sglang_config prefill and decode sections across all FP8 variants with multi-tensor-parallel (mtp) and single-tensor-parallel (stp) configurations.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

Possibly related PRs

• files for 8k/1k-fp4 #54 — Modifies sglang_config in recipes/gb200-fp4/8k1k/low-latency.yaml where this PR removes moe-dense-tp-size entries.
• Update GB300-FP4 configs #131 — Related to MOE dense tensor-parallelism configuration changes in GB300-FP4 recipe files.
• Add all GB200/GB300 FP8 MTP recipes #134 — Modifies moe-dense-tp-size settings in sglang_config parallelism blocks for prefill/decode.

Suggested reviewers

• trevor-m
• gracehonv

Poem

🐰 Through configs wide, the cleanup hops,
Removing moe-dense from prefill tops!
Twenty files trimmed with careful care,
Defaults take charge through the parser's air.
Cleaner recipes, ready to soar! ✨

🚥 Pre-merge checks | ✅ 3

✅ Passed checks (3 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title directly summarizes the primary change across all modified files: removing the moe-dense-tp-size configuration parameter from multiple YAML recipe files.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment
• Commit unit tests in branch kylliang/update_low_lat_config

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}