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

benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm.sh

@@ -235,7 +235,7 @@ EOF
 
     wait_for_lmcache_ready() {
         { set +x; } 2>/dev/null
-        local attempts="${LMCACHE_READY_ATTEMPTS:-120}"
+        local attempts="${LMCACHE_READY_ATTEMPTS:-1800}"
         local tail_pid=""
 
         while [ ! -f "$LMCACHE_LOG" ]; do
@@ -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 "$LMCACHE_HOST"
@@ -330,6 +332,14 @@ EOF
             --max-workers "$LMCACHE_MAX_WORKERS"
             --eviction-policy LRU
             --supported-transfer-mode "$LMCACHE_TX_MODE"
+            --no-separate-object-group
+            --no-l1-use-lazy
+            # lmcache_driven transfer mode uses CPU SHM by default for the L1
+            # pool (MixedMemoryAllocator's default shm-name); that pool is
+            # sized to the full --l1-size-gb and blows past the container's
+            # /dev/shm cap, SIGBUSing. --shm-name "" forces the pickle
+            # (non-SHM) path instead.
+            --shm-name ""
         )
         printf '%q ' "${LMCACHE_CMD[@]}" > "$RESULT_DIR/lmcache_command.txt"
         printf '\n' >> "$RESULT_DIR/lmcache_command.txt"

configs/amd-master.yaml

@@ -2199,7 +2199,7 @@ 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

perf-changelog.yaml

@@ -4927,3 +4927,13 @@
     - "Expand mi355x-disagg runner pool from 3 to 9 nodes."
     - "Concurrency sweep: 1K1K and 8K1K both [ 8, 16, 32, 64, 128 ]."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2094
+
+- 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 and --no-l1-use-lazy to the LMCache MP server launch command"
+    - "Bump default LMCACHE_READY_ATTEMPTS from 120 to 1800 (eager L1 pool allocation takes longer to become healthy)"
+    - "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"
+    - "Add --shm-name '' to the LMCache MP server launch command: lmcache_driven transfer mode's L1 pool defaults to CPU-SHM backing (sized to the full --l1-size-gb with --no-l1-use-lazy), exceeding the CI container's /dev/shm cap and SIGBUSing on startup. --shm-name '' forces the pickle (non-SHM) path"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2216

utils/matrix_logic/generate_sweep_configs.py

@@ -733,6 +733,7 @@ def generate_full_sweep(args, all_config_data, runner_data):
                                 ),
                                 Fields.DISAGG.value: disagg,
                                 Fields.SCENARIO_TYPE.value: "agentic-coding",
+                                Fields.RUN_EVAL.value: False,
                             }
                             if kv_offload_backend is not None:
                                 entry[Fields.KV_OFFLOAD_BACKEND.value] = kv_offload_backend
@@ -766,6 +767,7 @@ 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: False,
                             }
                             if kv_offload_backend is not None:
                                 entry[Fields.KV_OFFLOAD_BACKEND.value] = kv_offload_backend
@@ -1033,6 +1035,7 @@ 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: False,
                             }
                             if kv_offload_backend is not None:
                                 entry[Fields.KV_OFFLOAD_BACKEND.value] = kv_offload_backend
@@ -1065,6 +1068,7 @@ 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: False,
                             }
                             if kv_offload_backend is not None:
                                 entry[Fields.KV_OFFLOAD_BACKEND.value] = kv_offload_backend

utils/matrix_logic/validation.py

@@ -291,6 +291,8 @@ class SingleNodeAgenticMatrixEntry(BaseModel):
     duration: int = Field(alias=Fields.DURATION.value)
     exp_name: str = Field(alias=Fields.EXP_NAME.value)
     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_kv_offload_fields(self):
@@ -329,6 +331,8 @@ 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):
@@ -889,8 +893,10 @@ class ChangelogMatrixEntry(BaseModel):
                       ] = Field(default_factory=dict)
     multi_node: dict[str, list[Union[MultiNodeMatrixEntry, MultiNodeAgenticMatrixEntry]]
                      ] = Field(default_factory=dict)
-    evals: list[SingleNodeMatrixEntry] = Field(default_factory=list)
-    multinode_evals: list[MultiNodeMatrixEntry] = Field(default_factory=list)
+    evals: list[Union[SingleNodeMatrixEntry, SingleNodeAgenticMatrixEntry]
+                ] = Field(default_factory=list)
+    multinode_evals: list[Union[MultiNodeMatrixEntry, MultiNodeAgenticMatrixEntry]
+                          ] = Field(default_factory=list)
     changelog_metadata: ChangelogMetadata