Read authenticated page-delta chunks concurrently

deploy/deepseek_v4/tp4_profile.json

@@ -4,7 +4,7 @@
   "cache_model_profile": "deepseek-v4-fp8-hma",
   "published_runtime_base": "ghcr.io/fujitsupolycom/gb10-vllm-serving@sha256:6fc26fdad81a18f0fff67ce0a05f6d90165625ea2e1cac8a6f39bfb462017028",
   "sparkcache": {
-    "source_sha256": "bc7cae86732c869ee8b2205d48ac5be6f580ee8b77a3e4ffd4c69dcd4f1bfae5"
+    "source_sha256": "bc238f96e550c7ec27d4081dd1f2e741d404aaf5c8572d89ccc5e76812be4d63"
   },
   "model": {
     "repository": "deepseek-ai/DeepSeek-V4-Flash-0731",

deploy/glm52_35bpw/profile.json

@@ -5,7 +5,7 @@
   "published_runtime_base": "ghcr.io/fujitsupolycom/gb10-vllm-serving@sha256:6fc26fdad81a18f0fff67ce0a05f6d90165625ea2e1cac8a6f39bfb462017028",
   "base_image_requirement": "exact GLM-5.2 3.5-bpw R7 image recorded by the source container inspection",
   "sparkcache": {
-    "source_sha256": "bc7cae86732c869ee8b2205d48ac5be6f580ee8b77a3e4ffd4c69dcd4f1bfae5"
+    "source_sha256": "bc238f96e550c7ec27d4081dd1f2e741d404aaf5c8572d89ccc5e76812be4d63"
   },
   "model": {
     "repository": "brandonmusic/GLM-5.2-EXL3-TR3v4-3.5bpw-MTP78",

sparkcache/persistent_context_cache/cache_manifest.py

@@ -2263,8 +2263,7 @@ def _read_context_chunks(
         descriptors: Sequence[Mapping[str, Any]],
         required: frozenset[StateRecord],
     ) -> tuple[ContextChunk, ...]:
-        result = []
-        for descriptor in descriptors:
+        def _read_one(descriptor: Mapping[str, Any]) -> ContextChunk:
             encoded = (
                 self.root / "chunks" / f"{descriptor['sha256']}.spcc"
             ).read_bytes()
@@ -2280,8 +2279,12 @@ def _read_context_chunks(
                 or chunk.logical_end != descriptor["logical_end"]
             ):
                 raise CacheFormatError("chunk range disagrees with descriptor")
-            result.append(chunk)
-        return tuple(result)
+            return chunk
+
+        if not descriptors:
+            return ()
+        with ThreadPoolExecutor(max_workers=min(8, len(descriptors))) as pool:
+            return tuple(pool.map(_read_one, descriptors))
 
     def publish_prefix_aliases(
         self,

sparkcache/persistent_context_cache/test_cache_manifest.py

@@ -98,6 +98,41 @@ def _clear_once_in_subprocess(
 
 
 class ManifestStoreTests(unittest.TestCase):
+    def test_page_delta_chunk_reads_overlap_and_preserve_descriptor_order(
+        self,
+    ) -> None:
+        identity = _identity()
+        chunks = (_chunk(0, 256), _chunk(256, 512))
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            store = ManifestStore(root)
+            digest = hashlib.sha256(b"parallel-page-delta-reads").hexdigest()
+            store.commit(
+                identity=identity,
+                context_digest=digest,
+                chunks=chunks,
+                span_tokens=512,
+            )
+            manifest_path = (
+                root / "manifests" / identity.storage_key / f"{digest}.json"
+            )
+            descriptors = json.loads(manifest_path.read_bytes())["chunks"]
+            overlap = threading.Barrier(len(descriptors), timeout=2.0)
+            original_read_bytes = Path.read_bytes
+
+            def read_with_overlap(path: Path) -> bytes:
+                if path.suffix == ".spcc":
+                    overlap.wait()
+                return original_read_bytes(path)
+
+            with mock.patch.object(Path, "read_bytes", read_with_overlap):
+                restored = store._read_context_chunks(
+                    descriptors,
+                    identity.required_records,
+                )
+
+            self.assertEqual(restored, chunks)
+
     def test_page_extension_materializes_full_snapshot_after_base_root_removal(
         self,
     ) -> None: