Riley dixon/hipfile shim layer

.buildkite/k3_harness/setup-blend-env.sh

@@ -53,26 +53,17 @@ TEST_VENV_BIN="/workspace/.venv/bin"
 # When flashinfer and flashinfer-cubin resolve to different patch versions, skip strict check.
 export FLASHINFER_DISABLE_VERSION_CHECK=1
 
-# vLLM: pinned to a known-good cu129 nightly wheel as a temporary workaround.
-# Background: vLLM's PyPI stable is now built against CUDA 13 (libcudart.so.13)
-# which this CUDA-12 container can't load. The cu129 nightly channel auto-rolls
-# on every upstream commit and resolver-based installs have proven fragile
-# during the cu12/cu13 split (unsafe-best-match picks PyPI stable; first-index
-# with a version cap backtracked all the way to vllm 0.2.5). Pinning a wheel
-# URL bypasses resolution for vLLM; transitive deps still resolve from PyPI +
-# PyTorch cu129.
-# TODO: bump the URL when this pinned build goes stale, or revert to dynamic
-# selection once the cu12/cu13 transition settles.
-VLLM_WHEEL_URL="https://wheels.vllm.ai/8189a15914ca48461acf106f126c58ef7e41c9ee/vllm-0.20.2rc1.dev112%2Bg8189a1591.cu129-cp38-abi3-manylinux_2_34_x86_64.whl"
-echo "--- :python: Installing vLLM (pinned wheel: ${VLLM_WHEEL_URL})"
+# vLLM: force cu12.9 variant; PyPI default is cu13 (libcudart.so.13), unloadable in this cu12 container.
+echo "--- :python: Installing vLLM (nightly cu129)"
 
 # --index-strategy unsafe-best-match is needed so uv considers PyPI for vLLM's
 # transitive deps (setuptools>=77 specifically — the PyTorch cu129 index only
 # carries setuptools<=70.2.0, and uv's default first-index would refuse to
-# look at PyPI for it). Safe here because vLLM itself is pinned by URL, so
-# resolver strategy can't influence which vLLM gets installed.
-"${UV_BIN}" pip install -p "${TEST_VENV_BIN}/python" -U "${VLLM_WHEEL_URL}" \
-    --extra-index-url "https://download.pytorch.org/whl/cu129" \
+# look at PyPI for it). VLLM_PRECOMPILED_WHEEL_VARIANT pins the vLLM CUDA
+# variant, so the resolver strategy can't pull a mismatched build.
+VLLM_PRECOMPILED_WHEEL_VARIANT=cu129 "${UV_BIN}" pip install -p "${TEST_VENV_BIN}/python" -U vllm --pre \
+    --extra-index-url https://wheels.vllm.ai/nightly/cu129 \
+    --extra-index-url https://download.pytorch.org/whl/cu129 \
     --index-strategy unsafe-best-match
 
 
@@ -103,6 +94,8 @@ echo "--- :python: Installing LMCache from source"
 # Skip setuptools_scm git describe; the repo carries non-PEP-440 tags
 # (nightly, nightly-cu13) that crash the newer vcs_versioning backend.
 export SETUPTOOLS_SCM_PRETEND_VERSION_FOR_LMCACHE="${SETUPTOOLS_SCM_PRETEND_VERSION_FOR_LMCACHE:-0.0.0+ci}"
+# Select cu12 nixl wheel
+export LMCACHE_CUDA_MAJOR=12
 "${UV_BIN}" pip install -p "${DEFAULT_VENV_BIN}/python" -e . --no-build-isolation
 "${UV_BIN}" pip install -p "${TEST_VENV_BIN}/python" -e . --no-build-isolation
 # Work around openai_harmony vocab download/load issues for GPT-OSS (vLLM recipes troubleshooting).

.buildkite/k3_tests/multiprocess/scripts/run-long-doc-qa-l2.sh

@@ -388,14 +388,14 @@ def get_counter(name):
     return 0.0
 
 # L1 metrics
-l1_write_keys = get_counter('lmcache_mp_l1_write_keys_total')
+l1_write_keys = get_counter('lmcache_mp_l1_write_chunks_total')
 
 # L2 metrics
-store_keys = get_counter('lmcache_mp_l2_store_keys_total')
-store_succeeded = get_counter('lmcache_mp_l2_store_succeeded_keys_total')
-prefetch_lookups = get_counter('lmcache_mp_l2_prefetch_lookups_total')
-prefetch_hits = get_counter('lmcache_mp_l2_prefetch_hit_keys_total')
-prefetch_loaded = get_counter('lmcache_mp_l2_prefetch_loaded_keys_total')
+store_keys = get_counter('lmcache_mp_l2_store_submitted_objects_chunks_total')
+store_succeeded = get_counter('lmcache_mp_l2_store_completed_objects_chunks_total')
+prefetch_lookups = get_counter('lmcache_mp_l2_prefetch_lookup_requests_total')
+prefetch_hits = get_counter('lmcache_mp_l2_prefetch_hit_chunks_total')
+prefetch_loaded = get_counter('lmcache_mp_l2_prefetch_load_completed_chunks_total')
 
 print('=' * 60)
 print('Data Flow Metrics')
@@ -518,8 +518,8 @@ def has_label(base_name: str, label: str) -> bool:
 # (kind, metric_name, optional_label_to_assert_present_or_None)
 checks = [
     # ── Newer counters (with label dimensions) ─────────────────────
-    ("counter", "lmcache_mp_l2_store_completed_total", "l2_name"),
-    ("counter", "lmcache_mp_l2_load_completed_total", "l2_name"),
+    ("counter", "lmcache_mp_l2_store_completed_requests_total", "l2_name"),
+    ("counter", "lmcache_mp_l2_load_completed_requests_total", "l2_name"),
     ("counter", "lmcache_mp_lookup_requested_tokens_total", "model_name"),
     ("counter", "lmcache_mp_lookup_hit_tokens_total", "model_name"),
     ("counter", "lmcache_mp_num_chunks_loaded_total", "worker_id"),
@@ -528,10 +528,10 @@ checks = [
     # ``unit="GB/s"`` actually reports as
     # ``<name>_GB_per_second_count`` / ``..._sum`` / ``..._bucket``.
     # Match by base name and let the helper tolerate the unit suffix.
-    ("hist", "lmcache_mp_l0_l1_store_throughput_gbs", None),
-    ("hist", "lmcache_mp_l0_l1_load_throughput_gbs", None),
-    ("hist", "lmcache_mp_l2_store_throughput_gbs", "l2_name"),
-    ("hist", "lmcache_mp_l2_load_throughput_gbs", "l2_name"),
+    ("hist", "lmcache_mp_l0_l1_store_throughput", None),
+    ("hist", "lmcache_mp_l0_l1_load_throughput", None),
+    ("hist", "lmcache_mp_l2_store_throughput", "l2_name"),
+    ("hist", "lmcache_mp_l2_load_throughput", "l2_name"),
 ]
 
 failed = False

.buildkite/k3_tests/multiprocess/scripts/run-restart-recovery.sh

@@ -5,7 +5,7 @@
 #
 # Flow:
 #   1. Run a `lmcache bench engine --workload random-prefill` round.
-#   2. Snapshot lmcache_mp_l1_write_keys_total via /metrics.
+#   2. Snapshot lmcache_mp_l1_write_chunks_total via /metrics.
 #   3. Kill the LMCache server, relaunch on the same port.
 #   4. Wait for the new server to be ready and for the worker to
 #      re-register (poll /status until gpu_context_meta is non-empty).
@@ -97,7 +97,7 @@ run_bench_round() {
 }
 
 scrape_l1_write_keys() {
-    # Print the latest lmcache_mp_l1_write_keys_total value (single sample,
+    # Print the latest lmcache_mp_l1_write_chunks_total value (single sample,
     # unlabeled) from the LMCache HTTP server's /metrics endpoint.
     python3 - <<EOF
 import sys, urllib.request
@@ -111,7 +111,7 @@ total = None
 for line in body.splitlines():
     if line.startswith("#"):
         continue
-    if not line.startswith("lmcache_mp_l1_write_keys_total"):
+    if not line.startswith("lmcache_mp_l1_write_chunks_total"):
         continue
     parts = line.rsplit(" ", 1)
     if len(parts) != 2:
@@ -122,7 +122,7 @@ for line in body.splitlines():
         continue
     total = (total or 0.0) + val
 if total is None:
-    print(f"ERROR: lmcache_mp_l1_write_keys_total not found at {url}", file=sys.stderr)
+    print(f"ERROR: lmcache_mp_l1_write_chunks_total not found at {url}", file=sys.stderr)
     sys.exit(1)
 print(int(total))
 EOF
@@ -211,7 +211,7 @@ ROUND1_WRITES=$(scrape_l1_write_keys) || {
     echo "FAIL: could not scrape /metrics after round 1"
     exit 1
 }
-echo "Round 1 lmcache_mp_l1_write_keys_total = $ROUND1_WRITES"
+echo "Round 1 lmcache_mp_l1_write_chunks_total = $ROUND1_WRITES"
 
 if [ "$ROUND1_WRITES" -le 0 ]; then
     echo "FAIL: round 1 produced no L1 writes; benchmark setup is broken"
@@ -257,7 +257,7 @@ ROUND2_WRITES=$(scrape_l1_write_keys) || {
     echo "FAIL: could not scrape /metrics after round 2"
     exit 1
 }
-echo "Round 2 lmcache_mp_l1_write_keys_total = $ROUND2_WRITES"
+echo "Round 2 lmcache_mp_l1_write_chunks_total = $ROUND2_WRITES"
 
 # ── Step 4: Compare metrics ──────────────────────────────────
 echo ""

.buildkite/k3_tests/unit/run.sh

@@ -35,7 +35,6 @@ pytest --maxfail=1 --cov=lmcache \
     --cov-report term --cov-report=html:coverage-test \
     --cov-report=xml:coverage-test.xml --html=durations/test.html \
     --ignore=tests/disagg --ignore=tests/v1/test_pos_kernels.py \
-    --ignore=tests/v1/test_nixl_storage.py \
     --ignore=tests/skipped \
     --ignore=tests/v1/storage_backend/test_eic.py
 

.buildkite/operator/integration/pipeline.yaml

@@ -0,0 +1,28 @@
+steps:
+  - label: "Detect operator changes"
+    key: "operator_detect_changes"
+    command: |
+      set -eu
+
+      BASE_REF="origin/dev"
+
+      if ! git fetch --quiet origin dev 2>/dev/null; then
+        echo "WARN: could not fetch $$BASE_REF; running tests defensively."
+        buildkite-agent pipeline upload .buildkite/operator/integration/tests.yaml
+        exit 0
+      fi
+
+      if ! CHANGED=$(git diff --name-only "$$BASE_REF...HEAD" -- operator/ .buildkite/operator/ 2>/dev/null); then
+        echo "WARN: git diff failed; running tests defensively."
+        buildkite-agent pipeline upload .buildkite/operator/integration/tests.yaml
+        exit 0
+      fi
+
+      if [ -z "$$CHANGED" ]; then
+        echo "No operator or operator-pipeline changes vs $$BASE_REF; skipping integration tests."
+        exit 0
+      fi
+
+      echo "Operator changes detected:"
+      echo "$$CHANGED"
+      buildkite-agent pipeline upload .buildkite/operator/integration/tests.yaml

.buildkite/operator/integration/tests.yaml

@@ -0,0 +1,30 @@
+env:
+  GPU_KIND_CLUSTER: "operator-test-e2e-gpu-${BUILDKITE_BUILD_NUMBER}"
+
+steps:
+  - label: "Operator Integration: test-e2e-gpu-kind"
+    key: "operator_test_e2e_gpu_kind"
+    timeout_in_minutes: 90
+    command: |
+      echo $$PWD # for debugging
+
+      cd operator
+      make test-e2e-gpu-kind
+
+  - label: "Cleanup Kind cluster"
+    key: "operator_cleanup_kind"
+    depends_on: "operator_test_e2e_gpu_kind"
+    allow_dependency_failure: true
+    timeout_in_minutes: 10
+    command: |
+      cd operator
+      make cleanup-test-e2e-gpu-kind
+
+  - label: "Cleanup workspace"
+    depends_on: "operator_cleanup_kind"
+    allow_dependency_failure: true
+    command: |
+      export TARGET="$$PWD"
+      echo "Deleting current workspace $$TARGET"
+      cd /
+      rm -rf "$$TARGET" || true

.buildkite/pipeline.yml

@@ -25,6 +25,7 @@ steps:
         export CXX=hipcc
         export BUILD_WITH_HIP=1
         uv pip install torch torchvision --index-url https://download.pytorch.org/whl/rocm7.0
+        uv pip install -r requirements/rocm_core.txt
       fi
 
       uv pip install -r requirements/common.txt
@@ -37,8 +38,8 @@ steps:
         --cov-report term --cov-report=html:coverage-test \
         --cov-report=xml:coverage-test.xml --html=durations/test.html \
         --ignore=tests/disagg --ignore=tests/v1/test_pos_kernels.py \
-        --ignore=tests/v1/test_nixl_storage.py \
         --ignore=tests/v1/test_nixl_batched_contains.py \
+        --ignore=tests/v1/test_device_id_race.py \
         --ignore=tests/skipped \
         --ignore=tests/v1/storage_backend/test_eic.py
 

.github/CODEOWNERS

@@ -9,21 +9,24 @@
 /lmcache/v1/memory_management.py       @sammshen @deng451e @chunxiaozheng @DongDongJu @ApostaC
 
 # Multiprocess
-/lmcache/v1/multiprocess/              @ApostaC @OasisGit
+/lmcache/v1/multiprocess/              @ApostaC @OasisGit @hlin99
 /lmcache/v1/multiprocess/protocols/observability.py  @royyhuang @ApostaC
 /lmcache/v1/multiprocess/http_server.py              @royyhuang @ApostaC @OasisGit
-/lmcache/v1/mp_observability             @royyhuang @ApostaC @OasisGit @sammshen
+/lmcache/v1/mp_observability/            @yoo-kumaneko @royyhuang @ApostaC @OasisGit @sammshen
 
 # Distributed / L2
 /lmcache/v1/distributed/               @ApostaC @maobaolong @chunxiaozheng
-/lmcache/v1/distributed/eviction.py                              @YaoJiayi
-/lmcache/v1/distributed/eviction_policy/                         @YaoJiayi
-/lmcache/v1/distributed/storage_controllers/eviction_controller.py @YaoJiayi
+/lmcache/v1/distributed/eviction.py                              @YaoJiayi @hlin99
+/lmcache/v1/distributed/eviction_policy/                         @YaoJiayi @hlin99
+/lmcache/v1/distributed/storage_controllers/eviction_controller.py @YaoJiayi @hlin99
 /lmcache/v1/distributed/l2_adapters/mooncake_store_l2_adapter.py  @maobaolong
 /lmcache/v1/distributed/l2_adapters/resp_l2_adapter.py           @sammshen
 
 # GPU connector
-/lmcache/v1/gpu_connector/             @sammshen @ApostaC
+/lmcache/v1/gpu_connector/             @sammshen @ApostaC @hlin99
+
+# Platform
+/lmcache/v1/platform/                  @maobaolong @hlin99
 
 # Lookup client
 /lmcache/v1/lookup_client/             @maobaolong @YaoJiayi @sammshen
@@ -41,13 +44,13 @@
 /lmcache/v1/storage_backend/connector/fs_adapter.py            @maobaolong @chunxiaozheng
 /lmcache/v1/storage_backend/connector/audit_connector.py       @maobaolong @chunxiaozheng
 /lmcache/v1/storage_backend/connector/audit_adapter.py         @maobaolong @chunxiaozheng
-/lmcache/v1/storage_backend/connector/redis_connector.py      @sammshen
-/lmcache/v1/storage_backend/connector/redis_adapter.py        @sammshen
-/lmcache/v1/storage_backend/connector/lm_connector.py         @sammshen
-/lmcache/v1/storage_backend/connector/lm_adapter.py           @sammshen
+/lmcache/v1/storage_backend/connector/redis_connector.py      @sammshen @hlin99
+/lmcache/v1/storage_backend/connector/redis_adapter.py        @sammshen @hlin99
+/lmcache/v1/storage_backend/connector/lm_connector.py         @sammshen @hlin99
+/lmcache/v1/storage_backend/connector/lm_adapter.py           @sammshen @hlin99
 
 # Integration
-/lmcache/integration/vllm/             @sammshen @maobaolong @YaoJiayi @deng451e
+/lmcache/integration/vllm/             @sammshen @maobaolong @YaoJiayi @deng451e @hlin99
 /lmcache/integration/sglang/           @OasisGit @DongDongJu @sammshen
 /lmcache/integration/tensorrt_llm/     @sammshen
 
@@ -56,12 +59,15 @@
 
 # C extensions
 /csrc/                                 @ApostaC @YaoJiayi @sammshen
-/csrc/storage_backends/                @sammshen
+/csrc/storage_backends/                @sammshen @hlin99
 
 # Native connector L2 adapters
-/lmcache/v1/distributed/l2_adapters/native_connector_l2_adapter.py  @sammshen
-/lmcache/v1/distributed/l2_adapters/native_plugin_l2_adapter.py     @sammshen
-/lmcache/v1/distributed/l2_adapters/fs_native_l2_adapter.py         @sammshen
+/lmcache/v1/distributed/l2_adapters/native_connector_l2_adapter.py  @sammshen @hlin99
+/lmcache/v1/distributed/l2_adapters/native_plugin_l2_adapter.py     @sammshen @hlin99
+/lmcache/v1/distributed/l2_adapters/fs_native_l2_adapter.py         @sammshen @hlin99
+
+# Offline analysis tools
+/lmcache/tools/                        @yoo-kumaneko
 
 # Tests
 /tests/                                @hickeyma @sammshen @ApostaC @deng451e