[WIP][AMD][MI355x] dsv4-fp4-mi355x-vllm-agentic, install latest lmcache

benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm.sh

@@ -276,8 +276,7 @@ EOF
 
         git clone https://github.com/LMCache/LMCache.git
         cd LMCache
-        # https://github.com/LMCache/LMCache/pull/3853
-        git checkout 9229067cec0b3a63bb8a39368d101db7ac0bc3c1
+        git checkout v0.5.1
         pip install -r requirements/build.txt
         pip install grpcio==1.78.0
         CXX=hipcc BUILD_WITH_HIP=1 pip install -e .   --no-build-isolation
@@ -317,6 +316,9 @@ EOF
         LMCACHE_TX_MODE="lmcache_driven"
 
         echo "Starting LMCache MP server..."
+        # TEMP DEBUG: remove before merging
+        df -h /dev/shm
+        free -h
         LMCACHE_CMD=(
             lmcache server
             --host "$LMCACHE_HOST"
@@ -330,6 +332,7 @@ EOF
             --max-workers "$LMCACHE_MAX_WORKERS"
             --eviction-policy LRU
             --supported-transfer-mode "$LMCACHE_TX_MODE"
+            --no-separate-object-group
         )
         printf '%q ' "${LMCACHE_CMD[@]}" > "$RESULT_DIR/lmcache_command.txt"
         printf '\n' >> "$RESULT_DIR/lmcache_command.txt"

configs/amd-master.yaml

@@ -1422,7 +1422,7 @@ qwen3.5-fp8-mi355x-sglang-agentic-hicache:
 # DP-attention on/off and EP=8.
 
 dsv4-fp4-mi355x-vllm-agentic:
-  image: vllm/vllm-openai-rocm:nightly-09663abde0f50944a8d5ea30120666024b503faa
+  image: vllm/vllm-openai-rocm:v0.25.1
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:mi355x-amds
@@ -1435,7 +1435,7 @@ dsv4-fp4-mi355x-vllm-agentic:
       search-space:
       - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, conc-list: [1, 4, 8, 16, 32, 40, 48] }
       - { tp: 8, ep: 1, dp-attn: true, kv-offloading: none, conc-list: [64, 72], router: { name: vllm-router, version: "0.1.14" } }
-      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "9229067cec0b3a63bb8a39368d101db7ac0bc3c1" }, conc-list: [32, 40, 48] }
+      - { tp: 4, ep: 1, dp-attn: false, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "9229067cec0b3a63bb8a39368d101db7ac0bc3c1" }, conc-list: [32, 40, 48] }
 
 # DSv4-Pro FP4 on MI355X via SGLang. Uses a rocm720 mi35x image built off the
 # amd/deepseek_v4 branch in sgl-project/sglang; the SHA is encoded in the

perf-changelog.yaml

@@ -5060,3 +5060,13 @@
     - "Re-pin VLLM_ROUTER_IMAGE to vllm/vllm-router:nightly-20260716-1fbcde7 (previous nightly-20260629-e667ebb was garbage-collected from Docker Hub)"
     - "Exclude known-bad nodes mia1-p01-g09,g14 from the disagg node pool"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2301
+
+- config-keys:
+    - dsv4-fp4-mi355x-vllm-agentic
+  description:
+    - "Bump LMCache source build from pinned commit 9229067 to the v0.5.1 release tag (CXX=hipcc BUILD_WITH_HIP=1)"
+    - "Add --no-separate-object-group to the LMCache MP server launch command"
+    - "Bump image from vllm/vllm-openai-rocm:nightly-09663abde0f50944a8d5ea30120666024b503faa (rotated off Docker Hub retention, 404s) to the stable vllm/vllm-openai-rocm:v0.25.1 release tag"
+    - "Drop the lmcache DRAM arm from TP8 to TP4 (conc-list [32, 40, 48] unchanged): halves the LMCache L1 pool partition (TOTAL_CPU_DRAM_GB / (8/TP)) so the MP server comes up inside the healthcheck window"
+    - "generate_sweep_configs.py: agentic entries now stamp run-eval: true at construction time (was false) — SWE-bench eval is ready, so agentic points default to eval-on in sweeps that skip mark_eval_entries (--no-evals); eval-marked invocations still overwrite per the existing policy"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2216

utils/matrix_logic/generate_sweep_configs.py

@@ -783,6 +783,7 @@ def generate_full_sweep(args, all_config_data, runner_data):
                                 ),
                                 Fields.DISAGG.value: disagg,
                                 Fields.SCENARIO_TYPE.value: "agentic-coding",
+                                Fields.RUN_EVAL.value: True,
                             }
                             if kv_offload_backend is not None:
                                 entry[Fields.KV_OFFLOAD_BACKEND.value] = kv_offload_backend
@@ -816,6 +817,7 @@ def generate_full_sweep(args, all_config_data, runner_data):
                                     + (f"_spec-{spec_decoding}" if spec_decoding != "none" else "")
                                 ),
                                 Fields.SCENARIO_TYPE.value: "agentic-coding",
+                                Fields.RUN_EVAL.value: True,
                             }
                             if kv_offload_backend is not None:
                                 entry[Fields.KV_OFFLOAD_BACKEND.value] = kv_offload_backend
@@ -1081,6 +1083,7 @@ def generate_test_config_sweep(args, all_config_data, runner_data=None):
                                 ),
                                 Fields.DISAGG.value: disagg,
                                 Fields.SCENARIO_TYPE.value: "agentic-coding",
+                                Fields.RUN_EVAL.value: True,
                             }
                             if kv_offload_backend is not None:
                                 entry[Fields.KV_OFFLOAD_BACKEND.value] = kv_offload_backend
@@ -1113,6 +1116,7 @@ def generate_test_config_sweep(args, all_config_data, runner_data=None):
                                     + (f"_spec-{spec_decoding}" if spec_decoding != "none" else "")
                                 ),
                                 Fields.SCENARIO_TYPE.value: "agentic-coding",
+                                Fields.RUN_EVAL.value: True,
                             }
                             if kv_offload_backend is not None:
                                 entry[Fields.KV_OFFLOAD_BACKEND.value] = kv_offload_backend

utils/matrix_logic/validation.py

@@ -334,6 +334,8 @@ class MultiNodeAgenticMatrixEntry(BaseModel):
     exp_name: str = Field(alias=Fields.EXP_NAME.value)
     disagg: bool
     scenario_type: str = Field(alias=Fields.SCENARIO_TYPE.value)
+    run_eval: bool = Field(alias=Fields.RUN_EVAL.value)
+    eval_only: bool = Field(alias=Fields.EVAL_ONLY.value, default=False)
 
     @model_validator(mode='after')
     def validate_worker_hardware_pair(self):