FujitsuPolycom · FujitsuPolycom · Aug 30, 2026
@@ -123,12 +123,26 @@ NCCL error, or traceback.
 
 ## Cache namespace impact
 
-The overlay does not change SparkCache wire fields, digest salts, 256-token
-geometry, or stored object schemas. Its embedded-MTP digest is SHA-256 over
-`glm53-embedded-mtp-runtime-v1`, the target identity, the full vLLM commit,
-maximum depth five, and `adaptive:3:32`, separated by zero bytes.
+The connector explicitly requests publication schema `tail-cow-v1`. The
+`glm53-flash-hybrid` block-page profile maps that operator value to
+`page-tail-cow-v1` in `CacheIdentity`, so it selects the tail-specific storage
+key rather than the `snapshot-v1` storage key. `snapshot-v1` entries cleanly
+miss instead of being interpreted as page-tail objects.
+
+The publication schema does not change the embedded-MTP digest. That digest is
+SHA-256 over `glm53-embedded-mtp-runtime-v1`, the target identity, the full
+vLLM commit, maximum depth five, and `adaptive:3:32`, separated by zero bytes.
+SparkCache binds publication schema independently in `CacheIdentity`, so
+changing the MTP digest would add an unrelated identity change.
 
 Including the vLLM revision gives this runtime a distinct draft-state cache
 identity from the e105 adaptive-MTP profile. Stored entries therefore
 recompute instead of crossing the KDA source boundary without byte-equivalence
 evidence.
+
+The profile retains the one-shot clear token
+`sparkring-b12x-kda-adaptive-mtp-fastsafetensors-initialization`. Clear tokens
+are durable operator-action markers, not cache-identity fields. Rotating this
+token solely for the publication-schema selection would request another
+root-wide clear even though the tail-specific storage key prevents
+cross-schema restore.
@@ -86,7 +86,7 @@ uses these sanitized inputs:
 | File | Role |
 |---|---|
 | `glm53-flash-b12x-kda-adaptive-mtp-tp4-site.example.yaml` | Four-rank TP4/DCP1 site and 20 GiB FP8 KV reservation per rank |
-| `glm53-flash-b12x-kda-mtp5-adaptive-fastsafetensors-sparkcache-tp4-dcp1.example.json` | Adaptive MTP 3→5, 32-step window, fastsafetensors queue one, native SparkCache restore |
+| `glm53-flash-b12x-kda-mtp5-adaptive-fastsafetensors-sparkcache-tp4-dcp1.example.json` | Adaptive MTP 3→5, 32-step window, fastsafetensors queue one, SparkCache C++/CUDA restore, and page-tail copy-on-write publication |
 
 Resolve both templates with
 `scripts/prepare_glm53_b12x_kda_adaptive_mtp_profile.py`. The resolver rejects

@@ -82,7 +82,7 @@
     "--async-scheduling",
     "--enable-prefix-caching",
     "--cudagraph-metrics",
-    "--kv-transfer-config", "{\"kv_connector\":\"SparkContextCacheConnector\",\"kv_connector_module_path\":\"sparkcache.spark_context_cache_connector\",\"kv_role\":\"kv_both\",\"kv_load_failure_policy\":\"recompute\",\"kv_connector_extra_config\":{\"spark_cache_root\":\"/cache/jit/sparkcache-context\",\"spark_cache_model_profile\":\"glm53-flash-hybrid\",\"spark_cache_target_checkpoint_sha256\":\"a35e6bf2875c1875609b8deaec404c07c6cc80259e4222fc0b51e649498bd6b9\",\"spark_cache_draft_checkpoint_sha256\":\"3255539158b8a4fd199b4d97d89eb5231df3b39a5370881c16099a8059b09e44\",\"spark_cache_draft_policy\":\"separate\",\"spark_cache_store\":true,\"spark_cache_restore\":true,\"spark_cache_scheduler_probe\":\"none\",\"spark_cache_streaming_snapshots\":false,\"spark_cache_native_restore\":true,\"spark_cache_max_bytes\":51539607552,\"spark_cache_low_watermark_bytes\":42949672960,\"spark_cache_ttl_seconds\":0,\"spark_cache_min_span_tokens\":4096,\"spark_cache_max_span_tokens\":524288,\"spark_cache_native_library\":\"/opt/sparkcache-src/sparkcache/native/build-cuda/libspark_cache_placement.so\",\"spark_cache_native_library_sha256\":\"REPLACE_WITH_NATIVE_LIBRARY_SHA256\",\"spark_cache_native_arena_bytes\":268435456,\"spark_cache_native_io_workers\":8,\"spark_cache_load_threads\":2,\"spark_cache_clear_once\":\"sparkring-b12x-kda-adaptive-mtp-fastsafetensors-initialization\"}}"
+    "--kv-transfer-config", "{\"kv_connector\":\"SparkContextCacheConnector\",\"kv_connector_module_path\":\"sparkcache.spark_context_cache_connector\",\"kv_role\":\"kv_both\",\"kv_load_failure_policy\":\"recompute\",\"kv_connector_extra_config\":{\"spark_cache_root\":\"/cache/jit/sparkcache-context\",\"spark_cache_model_profile\":\"glm53-flash-hybrid\",\"spark_cache_target_checkpoint_sha256\":\"a35e6bf2875c1875609b8deaec404c07c6cc80259e4222fc0b51e649498bd6b9\",\"spark_cache_draft_checkpoint_sha256\":\"3255539158b8a4fd199b4d97d89eb5231df3b39a5370881c16099a8059b09e44\",\"spark_cache_draft_policy\":\"separate\",\"spark_cache_store\":true,\"spark_cache_restore\":true,\"spark_cache_scheduler_probe\":\"none\",\"spark_cache_streaming_snapshots\":false,\"spark_cache_publication_schema\":\"tail-cow-v1\",\"spark_cache_native_restore\":true,\"spark_cache_max_bytes\":51539607552,\"spark_cache_low_watermark_bytes\":42949672960,\"spark_cache_ttl_seconds\":0,\"spark_cache_min_span_tokens\":4096,\"spark_cache_max_span_tokens\":524288,\"spark_cache_native_library\":\"/opt/sparkcache-src/sparkcache/native/build-cuda/libspark_cache_placement.so\",\"spark_cache_native_library_sha256\":\"REPLACE_WITH_NATIVE_LIBRARY_SHA256\",\"spark_cache_native_arena_bytes\":268435456,\"spark_cache_native_io_workers\":8,\"spark_cache_load_threads\":2,\"spark_cache_clear_once\":\"sparkring-b12x-kda-adaptive-mtp-fastsafetensors-initialization\"}}"
   ],
   "extra_volumes": [
     {"host": "/REPLACE/GLM53_CACHE_HOST_ROOT", "container": "/cache/jit", "mode": "rw"}

@@ -28,6 +28,7 @@
 MTP_CACHE_IDENTITY_SHA256 = (
     "3255539158b8a4fd199b4d97d89eb5231df3b39a5370881c16099a8059b09e44"
 )
+PUBLICATION_SCHEMA = "tail-cow-v1"
 
 
 class ResolveError(ValueError):
@@ -104,6 +105,14 @@ def resolve(
         raise ResolveError("adaptive MTP observation window must be 32")
     if profile.get("environment", {}).get("VLLM_FASTSAFETENSORS_QUEUE_SIZE") != "1":
         raise ResolveError("fastsafetensors queue size must be one")
+    transfer = json.loads(_argument(profile, "--kv-transfer-config"))
+    connector_extra = transfer.get("kv_connector_extra_config")
+    if not isinstance(connector_extra, dict):
+        raise ResolveError("profile must contain SparkCache connector extra config")
+    if connector_extra.get("spark_cache_publication_schema") != PUBLICATION_SCHEMA:
+        raise ResolveError(
+            "profile must select SparkCache tail-cow-v1 publication"
+        )
     attestation = " ".join(str(value) for value in profile.get("attestation_hook", []))
     if SPARKCACHE_SOURCE_SHA256 not in attestation:
         raise ResolveError("profile does not attest the integrated SparkCache source")

@@ -12,6 +12,7 @@
 from prepare_glm53_b12x_kda_adaptive_mtp_profile import (
     LEASE_CONTRACT_SHA256,
     MTP_CACHE_IDENTITY_SHA256,
+    PUBLICATION_SCHEMA,
     SPARKCACHE_COMMIT,
     SPARKCACHE_SOURCE_SHA256,
     VLLM_COMMIT,
@@ -77,6 +78,7 @@ def test_profile_pins_adaptive_mtp_fastsafetensors_and_sparkcache() -> None:
     assert extra["spark_cache_native_arena_bytes"] == 256 * 1024**2
     assert extra["spark_cache_native_io_workers"] == 8
     assert extra["spark_cache_load_threads"] == 2
+    assert extra["spark_cache_publication_schema"] == PUBLICATION_SCHEMA
     assert extra["spark_cache_clear_once"] == (
         "sparkring-b12x-kda-adaptive-mtp-fastsafetensors-initialization"
     )
@@ -155,6 +157,26 @@ def test_resolver_rejects_runtime_or_loader_identity_drift() -> None:
         resolve(changed, copy.deepcopy(site), **arguments)
 
 
+def test_resolver_rejects_publication_schema_drift() -> None:
+    profile = json.loads(PROFILE.read_text(encoding="utf-8"))
+    site = yaml.safe_load(SITE.read_text(encoding="utf-8"))
+    transfer_index = profile["extra_vllm_args"].index("--kv-transfer-config") + 1
+    transfer = json.loads(profile["extra_vllm_args"][transfer_index])
+    transfer["kv_connector_extra_config"].pop("spark_cache_publication_schema")
+    profile["extra_vllm_args"][transfer_index] = json.dumps(transfer)
+
+    with pytest.raises(ResolveError, match="tail-cow-v1 publication"):
+        resolve(
+            profile,
+            site,
+            image="image",
+            image_id="sha256:" + "a" * 64,
+            parent_image="parent",
+            parent_image_id="sha256:" + "b" * 64,
+            native_library_sha256="c" * 64,
+        )
+
+
 def test_quickstart_names_the_executable_builder_and_profile_contracts() -> None:
     guide = QUICKSTART.read_text(encoding="utf-8")
     assert "runtime/glm53-flash-b12x-kda-adaptive-mtp/build-image.sh" in guide