Add DeepSeek-V4 GB300 agentic SGLang recipes

[nvpohanh]

[by Codex]

Description

Port the six DeepSeek-V4-Pro GB300 SGLang agentic recipes validated in SemiAnalysisAI/InferenceX#2157:

• one aggregated TP4 recipe
• five disaggregated DEP8 Pareto recipes
• MTP and hierarchical-cache KV offload settings from the successful AgentX sweep

The recipes use agentic for the workload directory instead of a fixed input/output sequence-length label. The /infmax-workspace references are container-local paths used by the public InferenceX AgentX harness, not cluster host paths.

Validation:

• parsed all six files with PyYAML
• passed srtctl dry-run for all six files using NVIDIA/srt-slurm v1.0.25 (3435776)
• ran python scripts/update_recipe_table.py (the support matrix was already covered, so it produced no diff)

Type of change

• New recipe
• Update to an existing recipe
• Tooling / scripts
• Documentation
• Other

Checklist

• I have read the Contributing Guidelines.
• The recipe is placed under the correct recipes/<single-node|multi-node>/<model>/<GPU>/<framework>/ path.
• No internal-only paths or hostnames (e.g. cluster-specific filesystem paths or mounts) are included.
• python scripts/update_recipe_table.py has been run and the README table is up to date.
• The recipe was validated with srtctl dry-run locally or by CI.

[faradawn]

LGTM

[shljessie]

Thanks for porting the six AgentX GB300 agentic recipes from InferenceX#2157. Overall this looks good to merge.

Paths / internal-only content
No issues — paths are container-local InferenceX harness conventions (/infmax-workspace, /logs/agentic, /model/, /configs/deepgemm_cache, etc.). No cluster host paths or hostnames. The launch_gb300-*.sh reference in comments is fine.

Source attribution
PR description citation to InferenceX#2157 is sufficient. Optional: add a one-line source comment in each YAML header for traceability.

Resources block
Checked disagg-gb300-4p4d-... — omitting gpus_per_prefill / gpus_per_decode is fine; srt-slurm computes 8 and 16 from (nodes × gpus_per_node) / workers. Explicit overrides in the other recipes are optional clarity, not required.

Container image
Using container: "dynamo-sglang" (alias) is the right pattern for this repo. The exact SGLang nightly matters at runtime (MTP, hicache, MegaMoE flags, etc.) but belongs in cluster/workflow config, not the recipe YAML. For reproducibility: agentic configs in InferenceX nvidia-master.yaml currently pin nightly-dev-cu13-20260711-7de33ce8, not 425dffbd (which is the 8k1k non-agentic image). Worth confirming which image was used for the validation runs cited here.

Nits (non-blocking)
Agg recipe comment # from submission for B300 — probably copy-paste; could say GB300.
dynamo.wheel: "1.3.0.dev1" — fine if that's what AgentX validated.

LGTM with the image-version note above as a documentation/repro question, not a blocker.