feat(frontends): add the llm-d router

.github/CODEOWNERS

@@ -1,5 +1,5 @@
 # Default owners for all files
-* @alec-flowers @nlevin-ui @csahithi
+* @alec-flowers @ishandhanani @nlevin-ui @csahithi
 
 # Curated examples owners
 /examples/ @kyleliang-nv @hjjq @qiching @kedarpotdar-nv

docs/cluster-config.md

@@ -44,8 +44,6 @@ reporting:
 
 **output_dir**: When set, job logs are written to `output_dir/{job_id}/logs` instead of `srtctl_root/outputs/{job_id}/logs`. Useful for CI/CD and ephemeral environments.
 
-**record_launch_plan**: Set `true` to record the realized launch plan (`output.record_launch_plan`) for every recipe on the cluster. Because the launch plan lives under the normal log directory, existing S3 postprocessing and external collectors that archive the complete log directory include it without a separate upload path.
-
 **containers**: A map from alias to image path or registry URI. One resolver replaces image aliases in `model.container`, `roles.<role>.container`, frontend/benchmark images, exporter images and `services[].container`. Literal paths and registry URIs pass through untouched. Free-form maps (`environment`, `roles.<role>.env`, `roles.<role>.args`, `services[].env`, `container_mounts`) and the `identity` block are never rewritten.
 
 **default_bash_preamble**: A shell snippet (e.g. `"ulimit -n 1048576 -s unlimited -u 1048576"`) prepended to every container srun launched by srtctl: workers, frontends, telemetry, benchmark, postprocess. Runs before per-call `bash_preamble` and the main command, so cluster-wide ulimits apply to everything downstream. Silently dropped for distroless containers (e.g. `prom/node-exporter`) that bypass the bash wrapper; a WARNING log is emitted in that case.

docs/config-reference.md

@@ -55,23 +55,15 @@ Fields: [ModelConfig](schema-reference.md#modelconfig).
 
 ## output
 
-Output paths and optional reproducibility artifacts.
+Output configuration with formattable paths.
 
 ```yaml
 output:
   log_dir: "./outputs/{job_id}/logs"
-  record_launch_plan: false    # Save exact realized srun scripts and a manifest
 ```
 
 The `log_dir` supports FormattablePath templating. See [FormattablePath Template System](runtime-env.md#formattablepath-template-system).
 
-When `record_launch_plan` is enabled, srtctl creates `logs/launch-plan/manifest.json` and one executable shell
-script for every realized `srun` invocation. Recording happens after Slurm assigns nodes, ports, heterogeneous
-groups, mounts, and container paths, so the scripts describe what was actually launched rather than a pre-submit
-estimate. Secret-like environment values are never persisted; the scripts name the environment variables that
-must be supplied for replay. The resolved recipe, outer `sbatch_script.sh`, lockfile, and resource snapshot remain
-alongside this directory and are referenced by the manifest.
-
 ## health_check
 
 Health check configuration for worker readiness, and the worker log watch that fails a run early.

docs/engines.md

@@ -144,7 +144,7 @@ Set `engine.dp_launch_mode: per_gpu` only when temporarily preserving the legacy
 
 When `TP x PP` fits on one node, srtslurm derives `--data-parallel-size-local` and `--data-parallel-start-rank`, then enables `--data-parallel-hybrid-lb` so every node-local process registers with the Dynamo frontend. When `TP x PP` is larger than the node-local GPU allocation, srtslurm instead derives the multi-node rendezvous arguments and makes every process except the global leader headless. For example, both DP4 x TP4 and DP2 x TP8 are selected automatically on four-GPU nodes.
 
-Do not set `data-parallel-size-local`, `data-parallel-start-rank`, `data-parallel-hybrid-lb`, or `headless` manually; srtslurm owns those values. The allocation must be regular: `DP x TP x PP` must match the endpoint GPU count, and a TP/PP replica must divide evenly within or across nodes.
+Do not set `data-parallel-size-local`, `data-parallel-start-rank`, `data-parallel-hybrid-lb`, or `headless` manually; srtslurm owns those values. Behind a router frontend, `data-parallel-external-lb: true` instead launches every DP rank as its own `vllm serve --data-parallel-rank <r>` on its own HTTP port, so the router addresses each rank ([llm-d](llm-d.md#routing-to-dp-ranks)). The allocation must be regular: `DP x TP x PP` must match the endpoint GPU count, and a TP/PP replica must divide evenly within or across nodes.
 
 ### TRT-LLM metrics publication
 

docs/frontends.md

@@ -1,14 +1,14 @@
 # Frontends and Dynamo
 
-The `frontend:` block (which router fronts the workers) and the `dynamo:` block (how Dynamo is installed and wired). Router-specific pages: [SGLang Router](sglang-router.md), [vLLM Router](vllm-router.md), [Shepherd Model Gateway](smg.md).
+The `frontend:` block (which router fronts the workers) and the `dynamo:` block (how Dynamo is installed and wired). Router-specific pages: [SGLang Router](sglang-router.md), [vLLM Router](vllm-router.md), [Shepherd Model Gateway](smg.md), [llm-d Router](llm-d.md).
 
 ## frontend
 
 Frontend/router configuration.
 
 ```yaml
 frontend:
-  # Frontend type: "dynamo" (default), "sglang-router", "vllm-router", "smg", direct "sglang", "vllm", "trtllm_serve",
+  # Frontend type: "dynamo" (default), "sglang-router", "vllm-router", "smg", "llm-d", direct "sglang", "vllm", "trtllm_serve",
   # or "none" for a services-only job (no router, no OpenAI endpoint, no worker-count health gate; only
   # valid without engine roles, see services[].nodes)
   type: dynamo