Bind DFlash7 to concurrent page-delta reads

docs/GLM53_DFLASH7_PYTHON_OVERLAY_SPARKCACHE_TP4_QUICKSTART.md

@@ -12,7 +12,7 @@ store/restart/restore, or concurrency qualification.
 | vLLM native extensions and wheel metadata | `da4d7be6c97434f6942292ed8abbf4b32dc44355` |
 | vLLM Python source | `0b67266a0f37d6146a8403fb8482403c62f412d5`, tree `ba9484ccb33aa56e90ff2f447f15ca9b9da97639` |
 | B12X | `b1d541f9e71a35f030d45fae437630fff7507c2a`, tree `c69cdec1c59a08e8e0e549f930fa8abcfb5134ae` |
-| SparkCache reconstructed-page placement and canonical CUDA configuration | `19e2ec8b59c84ef359c2a3290f86962e3ff71d96`, tree `d8b417bb4b6d734c4403c0a73e7e42b95abd8343`, clean source SHA-256 `bc7cae86732c869ee8b2205d48ac5be6f580ee8b77a3e4ffd4c69dcd4f1bfae5` |
+| SparkCache reconstructed-page placement, canonical CUDA configuration, and bounded page-delta reads | `5ec6a9953ad5d39120298bbfc26e95a6fa4b1dc3`, tree `94c236b9dfbf5f70075eb47877fd9caaa5d8c249`, clean source SHA-256 `bc238f96e550c7ec27d4081dd1f2e741d404aaf5c8572d89ccc5e76812be4d63` |
 | DFlash draft-loader separation | Patch SHA-256 `39b567013ee7aed79f63200ed460129587933dc77fb430decdf19f78178de279`, postimage SHA-256 `98acbae2b3bb4482d83f9637c163ce7c92707ccdf6561b7e431f23337f151cf4` |
 | Unused DeepEP removal | Distribution `deep_ep==2.0.0+local`, removal receipt SHA-256 `65514f44829e7d176b0b2cacc9559ed22724e525b7041a8bcd4d2e02d1f372e3` |
 | Target | `local-inference-lab/GLM-5.3-Flash-NVFP4@520de24eabf507659eaef7c70f14fd584527facc` |
@@ -110,12 +110,14 @@ manifests. The two target-loader profiles share a namespace because loader
 choice does not change target or draft model state; each profile uses a
 different cache root and one-shot clear token while qualification is pending.
 
-[SparkCache pull request #29](https://github.com/FujitsuPolycom/sparkcache/pull/29)
-combines canonical CUDA configuration names with replacement of a partial
-terminal HMA page when an authenticated cache boundary falls inside that page.
+[SparkCache pull request #30](https://github.com/FujitsuPolycom/sparkcache/pull/30)
+combines canonical CUDA configuration names, replacement of a partial terminal
+HMA page when an authenticated cache boundary falls inside that page, and an
+eight-worker reader for authenticated page-delta chunks. The reader preserves
+manifest descriptor order after concurrent reads.
 Moving from SparkCache commit
 `5d571018de5b63a9a90e5c11e6d6e86bbff4a957` to the pinned commit
-`19e2ec8b59c84ef359c2a3290f86962e3ff71d96` does not change the namespace.
+`5ec6a9953ad5d39120298bbfc26e95a6fa4b1dc3` does not change the namespace.
 Checkpoint identities, page-delta wire schemas, record vocabulary, digest
 salts, parallel geometry, vLLM patches, the lease contract, and the CUDA
 placement ABI are unchanged. Compatible `page-tail-cow-v1` entries remain

runtime/glm53-flash-dflash7-python-overlay/README.md

@@ -11,10 +11,10 @@ The image combines these exact roles:
 - the 31-file vLLM Python delta at
   `0b67266a0f37d6146a8403fb8482403c62f412d5`;
 - B12X `b1d541f9e71a35f030d45fae437630fff7507c2a`;
-- SparkCache reconstructed-page placement source
-  `19e2ec8b59c84ef359c2a3290f86962e3ff71d96`, Git tree
-  `d8b417bb4b6d734c4403c0a73e7e42b95abd8343`, and deployable source SHA-256
-  `bc7cae86732c869ee8b2205d48ac5be6f580ee8b77a3e4ffd4c69dcd4f1bfae5`;
+- SparkCache reconstructed-page placement and bounded page-delta-read source
+  `5ec6a9953ad5d39120298bbfc26e95a6fa4b1dc3`, Git tree
+  `94c236b9dfbf5f70075eb47877fd9caaa5d8c249`, and deployable source SHA-256
+  `bc238f96e550c7ec27d4081dd1f2e741d404aaf5c8572d89ccc5e76812be4d63`;
 - external BF16 DFlash2 weights with SHA-256
   `b33c03475ba7322cf398828f2d8d1be376df30dc05c6b40c28c8ea8da23e410b`.
 
@@ -54,12 +54,13 @@ No PR25 compatibility profile or legacy-key translation is required by these
 profiles.
 
 The pinned SparkCache source from
-[pull request #29](https://github.com/FujitsuPolycom/sparkcache/pull/29)
-accepts canonical CUDA configuration keys and replaces a partial terminal HMA
-page when the authenticated cache boundary falls inside that page. It does not
-change cache identity, page-tail wire schemas, record geometry, vLLM patch
-bytes, the lease contract, or the CUDA placement ABI. Compatible entries
-produced by the source at commit
+[pull request #30](https://github.com/FujitsuPolycom/sparkcache/pull/30)
+accepts canonical CUDA configuration keys, replaces a partial terminal HMA
+page when the authenticated cache boundary falls inside that page, and reads
+authenticated page-delta chunks with a bounded eight-worker pool while
+preserving descriptor order. It does not change cache identity, page-tail wire
+schemas, record geometry, vLLM patch bytes, the lease contract, or the CUDA
+placement ABI. Compatible entries produced by the source at commit
 `5d571018de5b63a9a90e5c11e6d6e86bbff4a957` remain in the same namespace.
 Null-block publication failures remain unsupported by this source contract.
 Both profiles preserve B12X compute backends and the pinned PYNCCL/NCCL

runtime/glm53-flash-dflash7-python-overlay/pins.json

@@ -77,9 +77,9 @@
   },
   "sparkcache": {
     "repository": "https://github.com/FujitsuPolycom/sparkcache.git",
-    "commit": "19e2ec8b59c84ef359c2a3290f86962e3ff71d96",
-    "tree": "d8b417bb4b6d734c4403c0a73e7e42b95abd8343",
-    "source_tree_sha256": "bc7cae86732c869ee8b2205d48ac5be6f580ee8b77a3e4ffd4c69dcd4f1bfae5",
+    "commit": "5ec6a9953ad5d39120298bbfc26e95a6fa4b1dc3",
+    "tree": "94c236b9dfbf5f70075eb47877fd9caaa5d8c249",
+    "source_tree_sha256": "bc238f96e550c7ec27d4081dd1f2e741d404aaf5c8572d89ccc5e76812be4d63",
     "cuda_config_schema": "canonical-v1",
     "canonical_cuda_config_keys": [
       "spark_cache_cuda_restore",

runtime/glm53-flash-dflash7-python-overlay/test_dflash7_python_overlay.py

@@ -38,13 +38,13 @@ def test_pins_bind_the_exact_dflash7_composition() -> None:
         "b1d541f9e71a35f030d45fae437630fff7507c2a"
     )
     assert pins["sparkcache"]["commit"] == (
-        "19e2ec8b59c84ef359c2a3290f86962e3ff71d96"
+        "5ec6a9953ad5d39120298bbfc26e95a6fa4b1dc3"
     )
     assert pins["sparkcache"]["tree"] == (
-        "d8b417bb4b6d734c4403c0a73e7e42b95abd8343"
+        "94c236b9dfbf5f70075eb47877fd9caaa5d8c249"
     )
     assert pins["sparkcache"]["source_tree_sha256"] == (
-        "bc7cae86732c869ee8b2205d48ac5be6f580ee8b77a3e4ffd4c69dcd4f1bfae5"
+        "bc238f96e550c7ec27d4081dd1f2e741d404aaf5c8572d89ccc5e76812be4d63"
     )
     assert pins["sparkcache"]["cuda_config_schema"] == "canonical-v1"
     assert pins["sparkcache"]["canonical_cuda_config_keys"] == [
@@ -105,6 +105,15 @@ def test_verifier_requires_the_dflash7_deployment_label() -> None:
         "glm53-flash-dflash7-python-overlay"
     )
     assert labels["org.sparkcache.cuda-config-schema"] == "canonical-v1"
+    assert labels["org.sparkcache.source-revision"] == (
+        "5ec6a9953ad5d39120298bbfc26e95a6fa4b1dc3"
+    )
+    assert labels["org.sparkcache.source-tree"] == (
+        "94c236b9dfbf5f70075eb47877fd9caaa5d8c249"
+    )
+    assert labels["org.sparkcache.source-sha256"] == (
+        "bc238f96e550c7ec27d4081dd1f2e741d404aaf5c8572d89ccc5e76812be4d63"
+    )
     assert labels["org.jovian.vllm.commit"] != labels[
         "org.sparkring.vllm.python.commit"
     ]

scripts/config/glm53-flash-dflash7-python-overlay-fastsafetensors-sparkcache-tp4-dcp1.example.json

@@ -150,9 +150,9 @@
     "max_num_seqs": "32",
     "sparkcache_publication_schema": "tail-cow-v1",
     "sparkcache_effective_publication_schema": "page-tail-cow-v1",
-    "sparkcache_source_sha256": "bc7cae86732c869ee8b2205d48ac5be6f580ee8b77a3e4ffd4c69dcd4f1bfae5",
-    "sparkcache_source_revision": "19e2ec8b59c84ef359c2a3290f86962e3ff71d96",
-    "sparkcache_source_tree": "d8b417bb4b6d734c4403c0a73e7e42b95abd8343",
+    "sparkcache_source_sha256": "bc238f96e550c7ec27d4081dd1f2e741d404aaf5c8572d89ccc5e76812be4d63",
+    "sparkcache_source_revision": "5ec6a9953ad5d39120298bbfc26e95a6fa4b1dc3",
+    "sparkcache_source_tree": "94c236b9dfbf5f70075eb47877fd9caaa5d8c249",
     "vllm_native_revision": "da4d7be6c97434f6942292ed8abbf4b32dc44355",
     "vllm_python_revision": "0b67266a0f37d6146a8403fb8482403c62f412d5",
     "vllm_python_tree": "ba9484ccb33aa56e90ff2f447f15ca9b9da97639",
@@ -182,9 +182,9 @@
     "org.sparkcache.deployment-profile": "glm53-flash-dflash7-python-overlay",
     "org.sparkcache.cuda-config-schema": "canonical-v1",
     "org.sparkcache.cuda-placement-library-sha256": "REPLACE_WITH_CUDA_PLACEMENT_LIBRARY_SHA256",
-    "org.sparkcache.source-revision": "19e2ec8b59c84ef359c2a3290f86962e3ff71d96",
-    "org.sparkcache.source-tree": "d8b417bb4b6d734c4403c0a73e7e42b95abd8343",
-    "org.sparkcache.source-sha256": "bc7cae86732c869ee8b2205d48ac5be6f580ee8b77a3e4ffd4c69dcd4f1bfae5",
+    "org.sparkcache.source-revision": "5ec6a9953ad5d39120298bbfc26e95a6fa4b1dc3",
+    "org.sparkcache.source-tree": "94c236b9dfbf5f70075eb47877fd9caaa5d8c249",
+    "org.sparkcache.source-sha256": "bc238f96e550c7ec27d4081dd1f2e741d404aaf5c8572d89ccc5e76812be4d63",
     "org.sparkcache.vllm-contract-sha256": "6defde9551cbb586fd09bb2d3020495531b6573397875a767eaae1dbad126024",
     "org.sparkring.runtime.removed-deep-ep-distribution": "deep_ep==2.0.0+local",
     "org.sparkring.runtime.deep-ep-removal-receipt-sha256": "65514f44829e7d176b0b2cacc9559ed22724e525b7041a8bcd4d2e02d1f372e3",
@@ -196,6 +196,6 @@
   "attestation_hook": [
     "/bin/sh",
     "-ec",
-    "test -f /models/target/config.json && test -f /models/target/model.safetensors.index.json && test \"$(sha256sum /models/target/config.json | cut -d ' ' -f1)\" = 676382abd1e90a6c85f0c8f33d45441ecd45fd514fd7b63ce5610e732d8e4996 && test \"$(sha256sum /models/target/model.safetensors.index.json | cut -d ' ' -f1)\" = 0d1d9e6b226e76520e182de10d4e7194cc885c5cb1bf885bb90de1916ce312cb && test -f /dflash-draft/config.json && test -f /dflash-draft/model.safetensors && test \"$(sha256sum /dflash-draft/config.json | cut -d ' ' -f1)\" = c4aeac0101196a6e26705b34c45230bcd0c7c68ee2d2d1efdb242087f3712573 && test \"$(sha256sum /dflash-draft/model.safetensors | cut -d ' ' -f1)\" = b33c03475ba7322cf398828f2d8d1be376df30dc05c6b40c28c8ea8da23e410b && test \"$(sha256sum /opt/sparkring/nccl/libnccl.so.2 | cut -d ' ' -f1)\" = 5f1c3f10d5ace66d4ba584415bbfe42b6ac1a0a9116a3b81dcbe50516ad924b3 && grep -q '\"tail-cow-v1\"' /opt/sparkcache-src/sparkcache/spark_context_cache_config.py && test \"$(sha256sum /opt/sparkring/runtime/python-overlay/vllm-python-overlay.json | cut -d ' ' -f1)\" = e5e528288b173399611a4930fecc4182b7208bc1564881d52ca5d2c5c4ae0f6a && test \"$(sha256sum /opt/sparkring/runtime/python-overlay/source-receipt.json | cut -d ' ' -f1)\" = REPLACE_WITH_SOURCE_RECEIPT_SHA256 && test \"$(python3 -c 'import json; print(json.load(open(\"/opt/sparkring/runtime/python-overlay/retained-native.json\", encoding=\"utf-8\"))[\"native_elf_manifest_sha256\"])')\" = REPLACE_WITH_VLLM_NATIVE_ELF_MANIFEST_SHA256 && test \"$(python3 -c 'import json; print(json.load(open(\"/opt/sparkring/runtime/python-overlay/retained-native.json\", encoding=\"utf-8\"))[\"native_dispatch_manifest_sha256\"])')\" = REPLACE_WITH_VLLM_NATIVE_DISPATCH_MANIFEST_SHA256 && test \"$(cat /opt/sparkring/runtime/python-overlay/sparkcache-source-tree.sha256)\" = bc7cae86732c869ee8b2205d48ac5be6f580ee8b77a3e4ffd4c69dcd4f1bfae5 && test \"$(sha256sum /opt/sparkcache-src/sparkcache/runtime_patches/vllm-kv-block-lease-contract-glm53-b12x-kda-adaptive-mtp.json | cut -d ' ' -f1)\" = 6defde9551cbb586fd09bb2d3020495531b6573397875a767eaae1dbad126024 && test \"$(sha256sum /opt/sparkcache-src/sparkcache/native/build-cuda/libspark_cache_placement.so | cut -d ' ' -f1)\" = REPLACE_WITH_CUDA_PLACEMENT_LIBRARY_SHA256 && test \"$(sha256sum /opt/sparkring/runtime/python-overlay/deep-ep-removal-receipt.json | cut -d ' ' -f1)\" = 65514f44829e7d176b0b2cacc9559ed22724e525b7041a8bcd4d2e02d1f372e3 && python3 -c 'import importlib.util; assert importlib.util.find_spec(\"deep_ep\") is None' && python3 /opt/sparkcache-src/sparkcache/runtime_patches/verify_lease_contract.py --vllm-root /usr/local/lib/python3.12/dist-packages --contract /opt/sparkcache-src/sparkcache/runtime_patches/vllm-kv-block-lease-contract-glm53-b12x-kda-adaptive-mtp.json"
+    "test -f /models/target/config.json && test -f /models/target/model.safetensors.index.json && test \"$(sha256sum /models/target/config.json | cut -d ' ' -f1)\" = 676382abd1e90a6c85f0c8f33d45441ecd45fd514fd7b63ce5610e732d8e4996 && test \"$(sha256sum /models/target/model.safetensors.index.json | cut -d ' ' -f1)\" = 0d1d9e6b226e76520e182de10d4e7194cc885c5cb1bf885bb90de1916ce312cb && test -f /dflash-draft/config.json && test -f /dflash-draft/model.safetensors && test \"$(sha256sum /dflash-draft/config.json | cut -d ' ' -f1)\" = c4aeac0101196a6e26705b34c45230bcd0c7c68ee2d2d1efdb242087f3712573 && test \"$(sha256sum /dflash-draft/model.safetensors | cut -d ' ' -f1)\" = b33c03475ba7322cf398828f2d8d1be376df30dc05c6b40c28c8ea8da23e410b && test \"$(sha256sum /opt/sparkring/nccl/libnccl.so.2 | cut -d ' ' -f1)\" = 5f1c3f10d5ace66d4ba584415bbfe42b6ac1a0a9116a3b81dcbe50516ad924b3 && grep -q '\"tail-cow-v1\"' /opt/sparkcache-src/sparkcache/spark_context_cache_config.py && test \"$(sha256sum /opt/sparkring/runtime/python-overlay/vllm-python-overlay.json | cut -d ' ' -f1)\" = e5e528288b173399611a4930fecc4182b7208bc1564881d52ca5d2c5c4ae0f6a && test \"$(sha256sum /opt/sparkring/runtime/python-overlay/source-receipt.json | cut -d ' ' -f1)\" = REPLACE_WITH_SOURCE_RECEIPT_SHA256 && test \"$(python3 -c 'import json; print(json.load(open(\"/opt/sparkring/runtime/python-overlay/retained-native.json\", encoding=\"utf-8\"))[\"native_elf_manifest_sha256\"])')\" = REPLACE_WITH_VLLM_NATIVE_ELF_MANIFEST_SHA256 && test \"$(python3 -c 'import json; print(json.load(open(\"/opt/sparkring/runtime/python-overlay/retained-native.json\", encoding=\"utf-8\"))[\"native_dispatch_manifest_sha256\"])')\" = REPLACE_WITH_VLLM_NATIVE_DISPATCH_MANIFEST_SHA256 && test \"$(cat /opt/sparkring/runtime/python-overlay/sparkcache-source-tree.sha256)\" = bc238f96e550c7ec27d4081dd1f2e741d404aaf5c8572d89ccc5e76812be4d63 && test \"$(sha256sum /opt/sparkcache-src/sparkcache/runtime_patches/vllm-kv-block-lease-contract-glm53-b12x-kda-adaptive-mtp.json | cut -d ' ' -f1)\" = 6defde9551cbb586fd09bb2d3020495531b6573397875a767eaae1dbad126024 && test \"$(sha256sum /opt/sparkcache-src/sparkcache/native/build-cuda/libspark_cache_placement.so | cut -d ' ' -f1)\" = REPLACE_WITH_CUDA_PLACEMENT_LIBRARY_SHA256 && test \"$(sha256sum /opt/sparkring/runtime/python-overlay/deep-ep-removal-receipt.json | cut -d ' ' -f1)\" = 65514f44829e7d176b0b2cacc9559ed22724e525b7041a8bcd4d2e02d1f372e3 && python3 -c 'import importlib.util; assert importlib.util.find_spec(\"deep_ep\") is None' && python3 /opt/sparkcache-src/sparkcache/runtime_patches/verify_lease_contract.py --vllm-root /usr/local/lib/python3.12/dist-packages --contract /opt/sparkcache-src/sparkcache/runtime_patches/vllm-kv-block-lease-contract-glm53-b12x-kda-adaptive-mtp.json"
   ]
 }