feat(agentx): run the MiniMax-M3 MI300X LMCache point on the lmcache-server service (#3545)

* feat(agentx): run the MiniMax-M3 MI300X LMCache point on the lmcache-server service

* chore(agentx): set the perf-changelog pr-link for #3545

* fix(agentx): match the MiniMax-M3 MI300X recipe image to its config

amd-master.yaml moved minimaxm3-fp8-mi300x-vllm-agentic-mtp to
vllm/vllm-openai-rocm:v0.30.0 in #3361, but the recipe still named v0.29.0,
so the single-node adapter rejected every point.

* fix(agentx): name the lmcache-mp connector on the MiniMax-M3 MI300X LMCache point

The refreshed lmcache-server patch (SemiAnalysisAI/srt-slurm#32 at 181b2e4)
keeps a role's connector on a direct vllm serve aggregate worker, so the
variant names connector: lmcache-mp instead of hand-writing its
kv-transfer-config. The message queue timeout falls back to LMCache's
default (300 s) instead of 6000 s.

* fix(agentx): keep the measured LMCache message queue timeout on MI300X

The role connector is now the lmcache-mp preset written out plus
lmcache.mp.mq_timeout 6000, so the point keeps its measured timeout
instead of falling back to LMCache's 300 s default.

* docs(changelog): keep only the perf-affecting changes in the #3545 entry

---------

Co-authored-by: Cameron Quilici <cameron@semianalysis.com>

Author: cquil11

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/lmcache-mp-rocm.sh

@@ -1,11 +1,8 @@
 #!/usr/bin/env bash
-# Start one LMCache MP server per TP rank in the worker container before vLLM,
-# as the legacy MI300X MiniMax-M3 AgentX script did. A variant opts in with
-# LMCACHE_SHARDS and LMCACHE_L1_SHARD_GB; its kv-transfer-config lists
-# tcp://127.0.0.1:5555 through 5555 + LMCACHE_SHARDS - 1.
+# Install the ROCm LMCache wheel for the MI300X MiniMax-M3 AgentX LMCache point.
+# It runs twice: as the vLLM worker's setup_script (the connector is imported
+# there) and in the lmcache-server service's preamble (its own container).
 set -euo pipefail
-[[ -n "${LMCACHE_SHARDS:-}" ]] || exit 0
-: "${LMCACHE_L1_SHARD_GB:?}"
 version=0.5.3
 pip_install=(python3 -m pip install)
 if python3 -m pip install --help 2>/dev/null | grep -q -- --break-system-packages; then
@@ -18,29 +15,3 @@ fi
     "lmcache==${version}" \
     --find-links "https://github.com/LMCache/LMCache/releases/expanded_assets/v${version}-rocm"
 python3 -c "import cupy; import lmcache.integration.vllm.lmcache_mp_connector; import opentelemetry.exporter.prometheus"
-
-pids=()
-for ((shard = 0; shard < LMCACHE_SHARDS; shard++)); do
-    # Detached so the servers outlive this preamble and serve the vLLM step.
-    setsid lmcache server \
-        --host 127.0.0.1 --port $((5555 + shard)) \
-        --http-host 127.0.0.1 --http-port $((8080 + shard)) \
-        --l1-size-gb "$LMCACHE_L1_SHARD_GB" --l1-init-size-gb 10 \
-        --l1-read-ttl-seconds 7200 --chunk-size 256 --max-workers 2 \
-        --eviction-policy LRU --supported-transfer-mode lmcache_driven \
-        > "/logs/lmcache_server_${shard}.log" 2>&1 < /dev/null &
-    pids+=($!)
-done
-for ((shard = 0; shard < LMCACHE_SHARDS; shard++)); do
-    for ((attempt = 0; ; attempt++)); do
-        python3 -c 'import sys, urllib.request; urllib.request.urlopen(sys.argv[1], timeout=2)' \
-            "http://127.0.0.1:$((8080 + shard))/healthcheck" 2> /dev/null && break
-        if ! kill -0 "${pids[$shard]}" 2>/dev/null || (( attempt >= 600 )); then
-            echo "ERROR: LMCache server $shard did not become ready" >&2
-            tail -n 50 "/logs/lmcache_server_${shard}.log" >&2 || true
-            exit 1
-        fi
-        sleep 1
-    done
-done
-echo "LMCache: ${LMCACHE_SHARDS} servers ready, ${LMCACHE_L1_SHARD_GB} GB L1 each"

inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/mi300x-fp8-mtp/agentic.yaml

@@ -1,11 +1,11 @@
 # MiniMax-M3 MXFP8 AgentX on MI300X with vLLM EAGLE3 (GQA draft). The KV cache
-# is GPU-resident, or backed by LMCache MP DRAM servers at the offload point.
+# is GPU-resident, or backed by an LMCache MP DRAM server at the offload point.
 base:
   schema: 2
   name: minimaxm3-fp8-mi300x-vllm-agentic
   model:
     path: hf:MiniMaxAI/MiniMax-M3-MXFP8
-    container: vllm/vllm-openai-rocm:v0.29.0
+    container: vllm/vllm-openai-rocm:v0.30.0
     precision: fp8
   resources:
     gpu_type: mi300x
@@ -24,8 +24,6 @@ base:
   health_check:
     interval_seconds: 10
     max_attempts: 360
-  # Starts the LMCache servers for the variant that sets LMCACHE_SHARDS.
-  setup_script: lmcache-mp-rocm.sh
   roles:
     agg:
       nodes: 1
@@ -72,8 +70,9 @@ base:
       AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'vllm:'
       AIPERF_APPLY_CHAT_TEMPLATE: 'true'
 
-# One variant per point. Admission is 2x CONC. The DRAM point splits its
-# 1298 GB budget into one LMCache L1 shard per TP rank (1298 / 8 = 162 GB).
+# One variant per point. Admission is 2x CONC. The DRAM point runs one
+# lmcache-server service on the node with the 8 x 162 GB L1 the per-rank
+# servers used to split between them.
 override_tp8_c2:
   roles:
     agg:
@@ -125,14 +124,40 @@ override_tp8_c10:
       KV_OFFLOADING: 'none'
 
 override_tp8_c16_lmcache:
+  # Installs LMCache in the worker container; the service installs its own.
+  setup_script: lmcache-mp-rocm.sh
+  services:
+    - name: lmcache
+      type: lmcache-server
+      preamble: bash /configs/lmcache-mp-rocm.sh
+      # The default probe with more time: the install counts against it.
+      readiness:
+        http:
+          port: 8751
+          path: /healthcheck
+        timeout_seconds: 900
+      args:
+        - --l1-size-gb
+        - '1296'
+        - --l1-init-size-gb
+        - '10'
+        - --l1-read-ttl-seconds
+        - '7200'
+        - --chunk-size
+        - '256'
+        - --max-workers
+        - '2'
+        - --eviction-policy
+        - LRU
+        - --supported-transfer-mode
+        - lmcache_driven
   roles:
     agg:
       args:
         max-num-seqs: 32
-        kv-transfer-config: '{"kv_connector":"LMCacheMPConnector","kv_connector_module_path":"lmcache.integration.vllm.lmcache_mp_connector","kv_role":"kv_both","kv_connector_extra_config":{"lmcache.mp.server_urls":"tcp://127.0.0.1:5555,tcp://127.0.0.1:5556,tcp://127.0.0.1:5557,tcp://127.0.0.1:5558,tcp://127.0.0.1:5559,tcp://127.0.0.1:5560,tcp://127.0.0.1:5561,tcp://127.0.0.1:5562","lmcache.mp.mq_timeout":6000.0}}'
-      env:
-        LMCACHE_SHARDS: '8'
-        LMCACHE_L1_SHARD_GB: '162'
+        # The lmcache-mp preset (the node's lmcache-server on port 8750) plus the
+        # measured message queue timeout.
+        connector: '{"kv_connector":"LMCacheMPConnector","kv_connector_module_path":"lmcache.integration.vllm.lmcache_mp_connector","kv_role":"kv_both","kv_connector_extra_config":{"lmcache.mp.host":"tcp://localhost","lmcache.mp.port":8750,"lmcache.mp.mq_timeout":6000.0}}'
   benchmark:
     env:
       CONC: '16'

inferencex-e2e/perf-changelog.yaml

@@ -9053,3 +9053,12 @@
   description:
     - "Update the GB300 DeepSeek-V4.1-Flash SGLang AgentX curve to lmsysorg/sglang:dev-cu13-nightly-0924 (digest pinned): pure TP4 (EP1) at C1/C2 with Engram in HBM, and TP4/EP4 at C4+ with per-rank host Engram, 16K prefill chunks, no prefill-decode interval, 4096 SWA prefix tails and a 128 decode graph batch; all TP4 points use static ragged verify; TP2 is unchanged."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3421
+
+- config-keys:
+    - minimaxm3-fp8-mi300x-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "All points now run vllm/vllm-openai-rocm:v0.30.0 (previously v0.29.0)."
+    - "LMCache DRAM point (TP8, concurrency 16): one LMCache server per node with 1296 GB L1 shared by all 8 TP ranks, replacing 8 per-rank servers of 162 GB L1 each."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3545