recipes(qwen3.5): refresh fp8 mtp-off wideep configs

[zhengd-nv]

Summary

Refresh of the Qwen3.5 FP8 mtp-off wideep recipes introduced by #128 driven by decode-side performance
sweeps on GB200.

Changes per recipe

• setup_script: switched from rebuild-deepep.sh to setup-router-and-deepep.sh. Even though sglang-router is not actually used by these recipes, this setup script additionally seeds the decode SGLANG_DG_CACHE_DIR (/tmp/deepgemm-cache, node-local) from the shared /configs cache; the underlying deepep-rebuild step is the same.
• identity: declares the sglang repo, container image, and framework version actually exercised (lmsysorg/sglang:0.5.10.post1), so jobs that apply these recipes can verify they're running the intended version.
• name: prefixed with wideep- so the runtime job name matches the file name (e.g. qwen3.5-wideep-1p1d-dep8-dpep-ccsweep).
• Drop deprecated decode option prefill-round-robin-balance: true.
• DEP16+: lower decode mem-fraction-static from 0.80 to 0.75. DeepEP buffers scale linearly with ep_size; at 0.80 CUDA graph capture OOMs inside cuda_graph_runner.capture_one_batch_size for DEP16/DEP32.
• Add SGLANG_HEALTH_STARTING_OK / SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION to both prefill and decode env; align decode SGLANG_DG_CACHE_DIR and SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK.
• Extend 1P ccsweep ranges so both the prefill-unsaturated and prefill-saturated regimes are sampled (DEP16 → cc=4096, DEP32 → cc=8192).

New recipe

• wideep-3p1d-dep16-dpep-cc4096.yaml: mirrors wideep-3p1d-dep32-dpep-cc4096.yaml but for DEP16. At cc=4096 (per-DP=256), 1P and 2P prefill remain saturated; only 3P unlocks the decode at full per-DP capacity, and this point produces the highest realized out/gpu of all tested configurations.

Caveat

6p1d-dep32-cc8192 (and any other multi-prefill recipe that fans out past ~5 workers) tends to hit zmq.error.ZMQError: Address already in use during sglang engine init when run on plain main. #134 (port jitter / odd-port allocation) resolves this; on plain main, multiple resubmissions may be required for the 6P+ configurations to start cleanly.

Test plan

• srtctl dry-run -f <recipe> for all 8 recipes (node counts + identity fields verified)
• srtctl apply -f wideep-1p1d-dep8-dpep-ccsweep.yaml: completes the full cc sweep (8 → 2048); output throughput matches the prior locally-generated config within <1 %
• srtctl apply -f wideep-6p1d-dep32-dpep-cc8192.yaml (on a branch carrying Sglang port jitter #134): initializes cleanly (no ZMQError) and completes the cc=8192 point

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (main@3183b4c). Learn more about missing BASE report.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #149   +/-   ##
=======================================
  Coverage        ?   65.07%           
=======================================
  Files           ?       67           
  Lines           ?     8214           
  Branches        ?        0           
=======================================
  Hits            ?     5345           
  Misses          ?     2869           
  Partials        ?        0           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.