docs(DeepSeek-V4): add h200|big verified recipes + tune H200 Pro parameters

[yushengsu-thu]

Summary

• Mark h200|big|low-latency, h200|big|balanced, h200|big|max-throughput as verified in the interactive command generator.
• Tune H200 Pro (big) parameters based on testing:
  • SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: 256 → 128 (balanced & max-throughput)
  • --cuda-graph-max-bs: 32 → 8, --max-running-requests: 64 → 32 (low-latency)
  • --mem-fraction-static: 0.82 → 0.88 (low-latency / balanced / max-throughput)
  • Balanced recipe: add dedicated --cuda-graph-max-bs 8 and --max-running-requests 32 for H200 Pro

Test plan

• Run mint dev locally and verify the interactive command generator produces correct commands for all H200 Pro combinations
• Verify no regressions for existing verified combos (H200 small, Blackwell, etc.)

Made with Cursor

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[Copilot]

Pull request overview

Updates the DeepSeek-V4 deployment interactive command generator snippet to treat H200 Pro (big) recipes as verified and to tune several H200 Pro-specific launch parameters based on recent testing.

Changes:

• Mark h200|big|low-latency, h200|big|balanced, and h200|big|max-throughput as verified (runnable) recipes.
• Adjust H200 Pro (big) DeepEP dispatch-token cap for balanced/max-throughput and reduce low-latency cuda-graph/max-running-requests.
• Increase H200 Pro (big) --mem-fraction-static to 0.88 in multiple recipes and add explicit balanced H200 Pro cuda-graph/max-running-requests flags.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The balanced recipe comment above says H200 big: cg=128 max-run=128 (same) but for hardware === "h200" && isBig the code now emits --cuda-graph-max-bs 8 and --max-running-requests 32. Update the comment to reflect the new H200 Pro tuned values to avoid confusing users (and to preserve the "mirror of allinone" intent documented earlier in the file).

[Copilot]

--mem-fraction-static was changed to 0.88 for all isBig low-latency runs. The block comment above says Blackwell big should use mem-frac 0.82, and the PR description suggests this tuning is specific to H200 Pro. Consider gating 0.88 to hardware === "h200" && isBig and keeping the prior value for other big variants to avoid altering existing verified B200/GB300 commands.

[Copilot]

The surrounding allinone summary comment for H200 big low-latency still mentions cg=32 max-run=64 (and mem-frac 0.82), but the actual flags now emit --cuda-graph-max-bs 8 and --max-running-requests 32 (and mem-frac 0.88). Please update the comment to match the new tuned values so the snippet remains self-consistent.