Bind DFlash7 to canonical SparkCache CUDA keys

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 | `5d571018de5b63a9a90e5c11e6d6e86bbff4a957`, tree `e864ed9ad64f771188fdb59aa9738e348134d636` |
+| SparkCache reconstructed-page placement and canonical CUDA configuration | `b95aa8ab0068dc66a6892a5c311d7e9dd4a9c55a`, tree `723fc604d73911a9e907798fd7932e4fc9c95df5`, clean source SHA-256 `48e008ba0cbd12f1ffae1c28388ea83310f41c6219c955e13d63ab171290d8de` |
 | DFlash draft-loader separation | Patch SHA-256 `39b567013ee7aed79f63200ed460129587933dc77fb430decdf19f78178de279`, postimage SHA-256 `98acbae2b3bb4482d83f9637c163ce7c92707ccdf6561b7e431f23337f151cf4` |
 | Target | `local-inference-lab/GLM-5.3-Flash-NVFP4@520de24eabf507659eaef7c70f14fd584527facc` |
 | External draft | `incoai/GLM-5.3-Flash-DFlash2@dc77ff1c99eeb2df044ee3d4f0094eb033fee410`, BF16 weights SHA-256 `b33c03475ba7322cf398828f2d8d1be376df30dc05c6b40c28c8ea8da23e410b` |
@@ -54,6 +54,9 @@ postimage SHA-256
 `98acbae2b3bb4482d83f9637c163ce7c92707ccdf6561b7e431f23337f151cf4`.
 Both profiles remain unqualified until live four-rank gates pass.
 
+SparkCache PR #26 accepts the canonical CUDA keys used by both profiles. No
+local PR25 compatibility profile or legacy-key rewrite is part of this path.
+
 ## Resolve the profile and inspect the plan
 
 Copy `scripts/config/glm53-flash-tp4-site.example.yaml`
@@ -93,3 +96,9 @@ MTP profiles. `tail-cow-v1` also separates these entries from snapshot-v1
 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.
+
+Moving from SparkCache PR #25 to PR #26 does not change the namespace. The
+checkpoint identities, publication schema, record vocabulary, digest salts,
+parallel geometry, vLLM patches, lease contract, and CUDA placement ABI are
+unchanged. Compatible PR25 `page-tail-cow-v1` entries remain eligible; only
+the configuration vocabulary and SparkCache source receipt change.

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

@@ -12,8 +12,8 @@ The image combines these exact roles:
   `0b67266a0f37d6146a8403fb8482403c62f412d5`;
 - B12X `b1d541f9e71a35f030d45fae437630fff7507c2a`;
 - SparkCache reconstructed-page placement source
-  `5d571018de5b63a9a90e5c11e6d6e86bbff4a957`, Git tree
-  `e864ed9ad64f771188fdb59aa9738e348134d636`;
+  `b95aa8ab0068dc66a6892a5c311d7e9dd4a9c55a`, Git tree
+  `723fc604d73911a9e907798fd7932e4fc9c95df5`;
 - external BF16 DFlash2 weights with SHA-256
   `b33c03475ba7322cf398828f2d8d1be376df30dc05c6b40c28c8ea8da23e410b`.
 
@@ -43,3 +43,11 @@ profile uses global fastsafetensors for the target and an exact
 `draft_load_config` selecting safetensors for DFlash. The image applies and
 verifies the draft-loader patch before installing SparkCache patches. See
 `docs/GLM53_DFLASH7_PYTHON_OVERLAY_SPARKCACHE_TP4_QUICKSTART.md`.
+
+The SparkCache source accepts the canonical CUDA configuration keys directly.
+No PR25 compatibility profile or legacy-key translation is required by these
+profiles.
+
+The PR26 source update does not change cache identity, page-tail publication,
+record geometry, vLLM patch bytes, the lease contract, or the CUDA placement
+ABI. Existing compatible PR25 cache entries remain in the same namespace.

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

@@ -77,9 +77,17 @@
   },
   "sparkcache": {
     "repository": "https://github.com/FujitsuPolycom/sparkcache.git",
-    "commit": "5d571018de5b63a9a90e5c11e6d6e86bbff4a957",
-    "tree": "e864ed9ad64f771188fdb59aa9738e348134d636",
-    "source_tree_sha256": "f7c0565521fddeff7085e4cc08043cb8d1e2bde33abc67f83b8608a162d05b88",
+    "commit": "b95aa8ab0068dc66a6892a5c311d7e9dd4a9c55a",
+    "tree": "723fc604d73911a9e907798fd7932e4fc9c95df5",
+    "source_tree_sha256": "48e008ba0cbd12f1ffae1c28388ea83310f41c6219c955e13d63ab171290d8de",
+    "cuda_config_schema": "canonical-v1",
+    "canonical_cuda_config_keys": [
+      "spark_cache_cuda_restore",
+      "spark_cache_cuda_placement_library",
+      "spark_cache_cuda_placement_library_sha256",
+      "spark_cache_cuda_placement_arena_bytes",
+      "spark_cache_cuda_restore_io_workers"
+    ],
     "contract": {
       "path": "sparkcache/runtime_patches/vllm-kv-block-lease-contract-glm53-b12x-kda-adaptive-mtp.json",
       "sha256": "6defde9551cbb586fd09bb2d3020495531b6573397875a767eaae1dbad126024",

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

@@ -47,6 +47,15 @@ def _longpath_run(argv, *, cwd=None):
 
 
 _original_run = common.run
+_original_copy_file = common.copy_file
+
+
+def _routed_copy_file(source: Path, destination: Path) -> None:
+    """Place the DFlash7 pins in the shared prepared-context topology."""
+
+    if source == COMMON / "pins.json":
+        source = PINS
+    _original_copy_file(source, destination)
 
 
 def _render_containerfile() -> str:
@@ -79,6 +88,15 @@ def _render_containerfile() -> str:
         if old not in text:
             raise PrepareError(f"shared Containerfile omits required text: {old}")
         text = text.replace(old, new)
+    deployment = (
+        'org.sparkcache.deployment-profile="glm53-flash-dflash7-python-overlay"'
+    )
+    if deployment not in text:
+        raise PrepareError("rendered Containerfile omits the DFlash7 deployment label")
+    text = text.replace(
+        deployment,
+        'org.sparkcache.cuda-config-schema="canonical-v1" \\\n      ' + deployment,
+    )
     return text
 
 
@@ -96,6 +114,17 @@ def _render_verify_image() -> str:
         if old not in text:
             raise PrepareError(f"shared image verifier omits required text: {old}")
         text = text.replace(old, new)
+    deployment = (
+        '"org.sparkcache.deployment-profile": '
+        '"glm53-flash-dflash7-python-overlay",'
+    )
+    if deployment not in text:
+        raise PrepareError("rendered verifier omits the DFlash7 deployment label")
+    text = text.replace(
+        deployment,
+        '"org.sparkcache.cuda-config-schema": "canonical-v1",\n'
+        f"        {deployment}",
+    )
     return text
 
 
@@ -132,16 +161,19 @@ def _replace_runtime_files(context: Path) -> None:
 def prepare(output: Path, *, repository_root: Path = ROOT) -> dict[str, Any]:
     original_pins = common.PINS
     original_run = common.run
+    original_copy_file = common.copy_file
     try:
         common.PINS = PINS
         common.run = _longpath_run
+        common.copy_file = _routed_copy_file
         common.prepare(output, repository_root=repository_root)
         _replace_runtime_files(output)
         common.verify_context(output)
         return json.loads((output / "receipt.json").read_text(encoding="utf-8"))
     finally:
         common.PINS = original_pins
         common.run = original_run
+        common.copy_file = original_copy_file
 
 
 def verify_context(context: Path) -> dict[str, Any]:

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

@@ -33,8 +33,22 @@ def test_pins_bind_the_exact_dflash7_composition() -> None:
         "b1d541f9e71a35f030d45fae437630fff7507c2a"
     )
     assert pins["sparkcache"]["commit"] == (
-        "5d571018de5b63a9a90e5c11e6d6e86bbff4a957"
+        "b95aa8ab0068dc66a6892a5c311d7e9dd4a9c55a"
     )
+    assert pins["sparkcache"]["tree"] == (
+        "723fc604d73911a9e907798fd7932e4fc9c95df5"
+    )
+    assert pins["sparkcache"]["source_tree_sha256"] == (
+        "48e008ba0cbd12f1ffae1c28388ea83310f41c6219c955e13d63ab171290d8de"
+    )
+    assert pins["sparkcache"]["cuda_config_schema"] == "canonical-v1"
+    assert pins["sparkcache"]["canonical_cuda_config_keys"] == [
+        "spark_cache_cuda_restore",
+        "spark_cache_cuda_placement_library",
+        "spark_cache_cuda_placement_library_sha256",
+        "spark_cache_cuda_placement_arena_bytes",
+        "spark_cache_cuda_restore_io_workers",
+    ]
     workload = pins["workload"]
     assert workload["role"] == "external-dflash7"
     assert workload["draft"]["weights_sha256"] == (
@@ -59,6 +73,7 @@ def test_rendered_image_metadata_names_dflash7_not_adaptive_mtp() -> None:
     recipe = prepare._render_containerfile()
     assert "SparkRing GLM-5.3 DFlash7 Python overlay" in recipe
     assert "glm53-flash-dflash7-python-overlay" in recipe
+    assert 'org.sparkcache.cuda-config-schema="canonical-v1"' in recipe
     assert "010-dflash-draft-load-config.patch" in recipe
     assert "98acbae2b3bb4482d83f9637c163ce7c92707ccdf6561b7e431f23337f151cf4" in recipe
     label_section = recipe[recipe.index("LABEL org.opencontainers.image.title=") :]
@@ -74,6 +89,7 @@ def test_verifier_requires_the_dflash7_deployment_label() -> None:
     assert labels["org.sparkcache.deployment-profile"] == (
         "glm53-flash-dflash7-python-overlay"
     )
+    assert labels["org.sparkcache.cuda-config-schema"] == "canonical-v1"
     assert labels["org.jovian.vllm.commit"] != labels[
         "org.sparkring.vllm.python.commit"
     ]

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

@@ -33,6 +33,7 @@ def expected_output_labels(pins: dict[str, Any]) -> dict[str, str]:
     labels["org.sparkcache.deployment-profile"] = (
         "glm53-flash-dflash7-python-overlay"
     )
+    labels["org.sparkcache.cuda-config-schema"] = "canonical-v1"
     return labels
 
 

scripts/config/README.md

@@ -107,6 +107,8 @@ SparkCache CUDA restore with page-tail copy-on-write publication.
 Resolve either profile with
 `scripts/prepare_glm53_dflash7_python_overlay_profile.py`. Follow the
 [DFlash7 Python-overlay quickstart](../../docs/GLM53_DFLASH7_PYTHON_OVERLAY_SPARKCACHE_TP4_QUICKSTART.md).
+Both templates use only SparkCache PR #26 canonical CUDA keys; they do not
+require a PR25 compatibility profile.
 
 ## DeepSeek-V4-Flash-0731
 

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

@@ -148,9 +148,9 @@
     "max_num_seqs": "32",
     "sparkcache_publication_schema": "tail-cow-v1",
     "sparkcache_effective_publication_schema": "page-tail-cow-v1",
-    "sparkcache_source_sha256": "f7c0565521fddeff7085e4cc08043cb8d1e2bde33abc67f83b8608a162d05b88",
-    "sparkcache_source_revision": "5d571018de5b63a9a90e5c11e6d6e86bbff4a957",
-    "sparkcache_source_tree": "e864ed9ad64f771188fdb59aa9738e348134d636",
+    "sparkcache_source_sha256": "48e008ba0cbd12f1ffae1c28388ea83310f41c6219c955e13d63ab171290d8de",
+    "sparkcache_source_revision": "b95aa8ab0068dc66a6892a5c311d7e9dd4a9c55a",
+    "sparkcache_source_tree": "723fc604d73911a9e907798fd7932e4fc9c95df5",
     "vllm_native_revision": "da4d7be6c97434f6942292ed8abbf4b32dc44355",
     "vllm_python_revision": "0b67266a0f37d6146a8403fb8482403c62f412d5",
     "vllm_python_tree": "ba9484ccb33aa56e90ff2f447f15ca9b9da97639",
@@ -175,10 +175,11 @@
     "org.opencontainers.image.base.name": "ghcr.io/fujitsupolycom/sparkring-glm53-runtime@sha256:864adfe68f458223e186a19844ac80c7adc7365e5db1f25e109b85fc19850dcd",
     "org.sparkring.base.image-id": "sha256:7e8c0ebcb2001efb4cdab0ec9d20d53972e62db3688230044e22e61ffb1d35d5",
     "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": "5d571018de5b63a9a90e5c11e6d6e86bbff4a957",
-    "org.sparkcache.source-tree": "e864ed9ad64f771188fdb59aa9738e348134d636",
-    "org.sparkcache.source-sha256": "f7c0565521fddeff7085e4cc08043cb8d1e2bde33abc67f83b8608a162d05b88",
+    "org.sparkcache.source-revision": "b95aa8ab0068dc66a6892a5c311d7e9dd4a9c55a",
+    "org.sparkcache.source-tree": "723fc604d73911a9e907798fd7932e4fc9c95df5",
+    "org.sparkcache.source-sha256": "48e008ba0cbd12f1ffae1c28388ea83310f41c6219c955e13d63ab171290d8de",
     "org.sparkcache.vllm-contract-sha256": "6defde9551cbb586fd09bb2d3020495531b6573397875a767eaae1dbad126024",
     "org.sparkring.source-receipt-sha256": "REPLACE_WITH_SOURCE_RECEIPT_SHA256",
     "org.sparkring.nccl.commit": "73cf112295c33aee2b895f329f592f2a9b4b0f97",
@@ -188,6 +189,6 @@
   "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)\" = f7c0565521fddeff7085e4cc08043cb8d1e2bde33abc67f83b8608a162d05b88 && 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 && 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)\" = 48e008ba0cbd12f1ffae1c28388ea83310f41c6219c955e13d63ab171290d8de && 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 && 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"
   ]
 }