docs: Clean up incomplete recipes and clarify Kubernetes-only focus

[BenHamm]

🎯 Objective

Clean up the recipes/ directory to focus exclusively on production-ready Kubernetes deployments for release 0.6.1. Remove incomplete configurations and clarify that benchmark recipes are tools for users, not published performance results.

---

📊 Summary of Changes

34 files changed: +210 insertions, -1,542 deletions

Deleted Content

• 5 incomplete model directories (no K8s deploy.yaml manifests):

  • deepseek-r1-distill-llama-8b/
  • gemma3/
  • llama4/
  • qwen2-vl-7b-instruct/
  • qwen3/ (kept qwen3-32b-fp8/ which is complete)

• run.sh script (228 lines) - non-K8s automation tool

• 12 standalone engine config YAMLs from deepseek-r1/trtllm/:

  • agg/simple/, agg/mtp/, agg/wide_ep/ (all subdirs)
  • disagg/simple/, disagg/mtp/
  • disagg/wide_ep/*.yaml (standalone configs only)
  • Kept: disagg/wide_ep/gb200/ (has complete K8s manifests)

Added/Modified Content

• Comprehensive new recipes/README.md with:

  • Complete recipe table showing Deployment and Benchmark Recipe availability
  • Prerequisites section linking to correct K8s installation guides
  • Step-by-step deployment instructions
  • Troubleshooting section
  • Removed extraneous links (docs.nvidia.com, license section)

• gpt-oss-120b/trtllm/disagg/README.md - Documents incomplete recipe status

---

📦 What Remains

4 Models | 10 Complete Deployments | 7 with Benchmark Recipes

Model	Framework	Recipes	Status
llama-3-70b	vLLM	agg, disagg-single-node, disagg-multi-node	✅ All with deploy.yaml + perf.yaml
qwen3-32b-fp8	TensorRT-LLM	agg, disagg	✅ All with deploy.yaml + perf.yaml
gpt-oss-120b	TensorRT-LLM	agg, disagg*	✅ agg complete; disagg documented as incomplete
deepseek-r1	SGLang + TensorRT-LLM	sglang/disagg-8gpu, sglang/disagg-16gpu, trtllm/disagg/wide_ep/gb200	✅ 2 functional, 1 complete

---

🔑 Key Changes Explained

1. Removed Incomplete Model Directories

Issue: 5 model directories contained only standalone engine configs (.yaml files) without Kubernetes deploy.yaml manifests. These were added in PR #3772 but don't follow the documented recipe structure.

Resolution: Remove them from 0.6.1. Can be re-added properly in 0.7.0 with complete K8s manifests.

2. Removed run.sh Script

Issue: Automated deployment script for non-K8s environments. Adds complexity and isn't aligned with K8s-only focus.

Resolution: Remove for 0.6.1. Recipes directory should contain declarative K8s manifests only.

3. Cleaned Up Standalone Engine Configs

Issue: deepseek-r1/trtllm/ contained many standalone engine config YAMLs (agg.yaml, decode.yaml, prefill.yaml) that aren't wrapped in K8s manifests. These configs are meant to be embedded as ConfigMaps within deploy.yaml files.

Resolution: Remove standalone configs. The one complete recipe (disagg/wide_ep/gb200/deploy.yaml) already embeds its configs as ConfigMaps.

4. Clarified "Benchmark Recipe" vs "Benchmarked"

Issue: Previous README implied we have published benchmark results. We actually provide benchmark tools (perf.yaml with AIPerf).

Resolution:

• Changed table columns from "Benchmarked | Status" to "Deployment | Benchmark Recipe"
• Legend now clarifies: "Benchmark Recipe: ✅ = Includes perf.yaml for running AIPerf benchmarks"
• Removed claims of having performance numbers

---

📁 Directory Structure (After)

recipes/
├── README.md (comprehensive guide)
├── CONTRIBUTING.md
├── hf_hub_secret/
├── llama-3-70b/
│   ├── model-cache/
│   └── vllm/ (3 complete recipes)
├── qwen3-32b-fp8/
│   ├── model-cache/
│   └── trtllm/ (2 complete recipes)
├── gpt-oss-120b/
│   ├── model-cache/
│   └── trtllm/
│       ├── agg/ (complete)
│       └── disagg/ (documented as incomplete)
└── deepseek-r1/
    ├── model-cache/
    ├── sglang/ (2 functional recipes)
    └── trtllm/disagg/wide_ep/gb200/ (1 complete recipe)

---

🤔 Review Focus Areas

1. Scope of deletions - Are we okay removing these 5 models from 0.6.1?
2. README accuracy - Does the new table and documentation clearly convey what's available?
3. Incomplete gpt-oss disagg - Should we keep it documented, or delete it entirely?
4. DeepSeek-R1 structure - Only GB200 recipe remains. Should we flatten the directory structure?

[biswapanda]

LGTM. went over the shorter README and verified links are fine.

minor nit: recipes: .. PR title needs update. Valid prefix are ~ fix: , docs: