docs: correct TRT-LLM FPM availability and KV-transfer success metric side

[dagil-nvidia]

Overview

Three docs were inconsistent with the 1.2.0 TRT-LLM implementation:

1. docs/backends/trtllm/trtllm-observability.md:159 said trtllm_kv_transfer_success_total is recorded on the prefill side. The 1.2.0 implementation records it on the decode worker, in lock-step with the sibling KV-transfer latency/bytes/speed histogram counts. Verified against components/src/dynamo/trtllm/request_handlers/handler_base.py and components/src/dynamo/trtllm/metrics.py.

2. docs/components/planner/README.md:127 said FPM is "currently only available for vllm". TRT-LLM has supported FPM for non-attention-DP workers since FpmDirectPublisher landed (PRs feat(trtllm): publish ForwardPassMetrics via FpmDirectPublisher (non-attention-DP) #8356 and perf(trtllm): publisher off request-loop + batched KV walk #8892); SGLang has the Dynamo-side pipeline wired but the upstream SGLang FPM module is not in the 1.2.0 runtime image. Replaced with a per-backend support matrix.

3. docs/kubernetes/model-deployment-guide.md:316 carried the same vLLM-only claim in the "Recommended Backend" table. Updated to match the planner README matrix.

Changes

• docs/backends/trtllm/trtllm-observability.md — correct counter side
• docs/components/planner/README.md — per-backend FPM support matrix
• docs/kubernetes/model-deployment-guide.md — update FPM row

No code or example changes. Pure documentation correctness fix.

Test plan

• Documentation renders cleanly (markdown only, no link/anchor changes)
• Cross-link from planner/README.md to sglang-observability.md#forward-pass-metrics-fpm resolves
• No bug-tracker URLs leak into user-facing docs

Fixes NVBug 6204453 / DYN-3091.

---

Summary by CodeRabbit

• Documentation
  • Clarified TensorRT-LLM Prometheus metric documentation regarding KV cache transfer tracking.
  • Expanded ForwardPassMetrics backend support details for vLLM, TensorRT-LLM, and SGLang in planner configuration.
  • Updated backend selection guidance for load-based scaling scenarios with improved clarity on supported configurations.

[github-actions]

🌿 Fern Docs Preview: https://nvidia-preview-37e83a35-7b68-496b-93eb-b3ed8f9eae4b.docs.buildwithfern.com/dynamo/dev

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: c489e625-9d3e-45a2-96d6-fc150d7713d4

📥 Commits

Reviewing files that changed from the base of the PR and between 99db412 and e5299dd.

📒 Files selected for processing (3)

• docs/backends/trtllm/trtllm-observability.md
• docs/components/planner/README.md
• docs/kubernetes/model-deployment-guide.md

---

Walkthrough

Updated documentation across three files to clarify backend support for observability features: refined TensorRT-LLM's KV transfer metric description to specify decode-worker recording behavior, and expanded guidance on ForwardPassMetrics availability across vLLM, TensorRT-LLM, and SGLang backends with specific gating conditions and runtime prerequisites.

Changes

Backend Capability and Observability Documentation Clarifications

Layer / File(s)	Summary
TensorRT-LLM KV transfer observability metric docs/backends/trtllm/trtllm-observability.md	trtllm_kv_transfer_success_total metric description refined to specify recording occurs on decode worker when non-zero KV-transfer timing is observed in RequestPerfMetrics.timing_metrics, and clarifies lock-step tracking with sibling histogram _count values.
ForwardPassMetrics backend support guidance docs/components/planner/README.md, docs/kubernetes/model-deployment-guide.md	Expanded documentation to clarify ForwardPassMetrics availability across backends: vLLM (any configuration), TensorRT-LLM (non-attention-DP only via attention_dp_size gating), and SGLang (wired in Dynamo but upstream module not present in 1.2.0 runtime image).

🎯 1 (Trivial) | ⏱️ ~3 minutes

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately summarizes the main changes: correcting TRT-LLM FPM availability documentation and fixing the KV-transfer success metric side specification.
Description check	✅ Passed	The description provides a comprehensive overview of changes, implementation verification sources, and testing methodology that aligns with the template structure.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

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

---

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.}