Test llm-d's default scheduler at srt-slurm#45 0c209710

Carry SemiAnalysisAI/srt-slurm#45 at 0c209710 and add a DNM key that runs the llm-d frontend without an epp_config.

Signed-off-by: Cam Quilici <cjquilici@gmail.com>

Author: cquil11

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/minimaxm3/vllm/gb300-fp4/agentx/llm-d-precise-dnm.yaml

@@ -291,3 +291,15 @@ override_offload_1p1d:
       args:
         <<: *args
         kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both"},{"kv_connector":"OffloadingConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":107374182400}}]}}'
+
+# No frontend.epp_config: srtctl runs the scorers of llm-d's no-Kubernetes guide (queue
+# depth, KV-cache utilization, approximate prefix cache, no-hit LRU) over two TP4 workers.
+override_default_agg:
+  name: "minimax-m3-vllm-agg-gb300-2xtp4-llm-d-default-agentic"
+  roles:
+    agg:
+      nodes: 2
+      workers: 2
+      gpus: 4
+      env: *env
+      args: *args

inferencex-e2e/configs/nvidia-master.yaml

@@ -6362,6 +6362,33 @@ minimaxm3-fp4-gb300-vllm-llm-d-offload-agentic-disagg-dnm:
           - "CONFIG_FILE=recipes/minimaxm3/vllm/gb300-fp4/agentx/llm-d-precise-dnm.yaml:override_offload_1p1d"
         decode: { num-worker: 1, tp: 4, ep: 1, dp-attn: false }
 
+# [DNM] llm-d without an epp_config: srtctl's default scheduler (the no-Kubernetes guide's scorers) over two TP4 workers.
+minimaxm3-fp4-gb300-vllm-llm-d-default-agentic-agg-dnm:
+  image: quay.io/rh-ee-imarkov/llm-d-nokube-vllm:dspark-0814-nightly@sha256:00742fdd10e572172d49559e44c357a68c388933c74db2c4e819bf61377a6a95
+  model: nvidia/MiniMax-M3-NVFP4
+  model-prefix: minimaxm3
+  runner: cluster:gb300-nv
+  precision: fp4
+  framework: vllm
+  router: { name: llm-d-router, version: "0.10.0" }
+  multinode: true
+  disagg: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 1.0
+      search-space:
+      - spec-decoding: none
+        kv-offloading: none
+        conc-list: [8]
+        num-nodes: 2
+        worker:
+          num-worker: 2
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/minimaxm3/vllm/gb300-fp4/agentx/llm-d-precise-dnm.yaml:override_default_agg"
+
 dsv4-fp4-gb200-dynamo-vllm-agentic-mtp-agg:
   image: vllm/vllm-openai:nightly-3ee2df30337a301164c46ae444b76ee67e71c106
   model: deepseek-ai/DeepSeek-V4-Pro

inferencex-e2e/runners/srt-slurm/patches/llm-d-frontend.patch

@@ -30,10 +30,10 @@ index db321a70..716b46a0 100644
  | `num_additional_frontends`  | int  | 9             | Additional routers beyond master    |
 diff --git a/docs/llm-d.md b/docs/llm-d.md
 new file mode 100644
-index 00000000..1531e5d8
+index 00000000..e9cb0888
 --- /dev/null
 +++ b/docs/llm-d.md
-@@ -0,0 +1,247 @@
+@@ -0,0 +1,252 @@
 +# llm-d router
 +
 +`frontend.type: llm-d` routes requests with the [llm-d](https://llm-d.ai) router: the
@@ -262,6 +262,11 @@ index 00000000..1531e5d8
 +- The token producer's `vllm.url` is the first prefill worker's (else the first worker's)
 +  own vLLM API, and `modelName` defaults to the served model name.
 +
++With vLLM's `OffloadingConnector` and its `self_describing_kv_events: true`, a worker also
++reports the blocks it offloads to host memory; the producer indexes them as the `cpu` tier
++and weights tiers by `indexerConfig.kvCacheBackendConfigs` (`gpu` 1.0, `cpu` 0.8 by
++default).
++
 +The roles that publish and the producer go together: either without the other is
 +rejected, as are a producer `kvEventsConfig` that sets the socket or discovery, a
 +`token-producer` `vllm.url`, and a router not on the head node. A role with
@@ -1977,10 +1982,10 @@ index 73ebd73a..0bde38cd 100644
          return None
 diff --git a/src/srtctl/frontends/llm_d.py b/src/srtctl/frontends/llm_d.py
 new file mode 100644
-index 00000000..48c17a25
+index 00000000..a78c9349
 --- /dev/null
 +++ b/src/srtctl/frontends/llm_d.py
-@@ -0,0 +1,516 @@
+@@ -0,0 +1,521 @@
 +# SPDX-FileCopyrightText: Copyright (c) 2026 SemiAnalysis LLC. All rights reserved.
 +# SPDX-License-Identifier: Apache-2.0
 +
@@ -2002,6 +2007,11 @@ index 00000000..48c17a25
 +Envoy's admin ``/ready`` answers and the EPP reports every endpoint ready
 +(``llm_d_epp_ready_endpoints``: endpoints whose metrics it scrapes).
 +
++An external-LB DP rank (``data-parallel-external-lb``) is an endpoint of its own.
++With ``roles.<role>.kv_events`` the workers publish their KV-cache events to the
++EPP's precise prefix-cache producer on one socket srtctl owns
++(``kv_events_subscriber``, ``with_kv_events``).
++
 +Upstream, pinned: llm-d-router v0.11.0
 +(https://github.com/llm-d/llm-d-router/tree/a5cbe600ebade00cf3e9885beaf2bfacddeabce1):
 +``cmd/epp/runner/runner.go`` (``runWithFileDiscovery``),