FujitsuPolycom · FujitsuPolycom · Aug 30, 2026 · Aug 30, 2026 · Aug 30, 2026 · Aug 30, 2026
@@ -1,8 +1,35 @@
 # Serve GLM-5.3 with external DFlash7 and the exact Python-overlay runtime
 
-Status: **implemented, not qualified** for the selected source contract. The
-image builder, profile resolver, and four-rank dry-run contract pass without
-GPUs. Historical local image ID
+Status: **implemented** for reproducible image construction, profile
+resolution, and four-rank dry-run planning from the selected source contract.
+Exact local image
+`sparkring-glm53-sparkcache:dflash7-pr39-reaching-d93cb3d-arm64`, image ID
+`sha256:ed60be066d6d9eadea267bc4597a0687869f3ddb95a3e5c6f86649893a838eb8`,
+is **qualified** only for the bounded cases in the
+[exact-artifact validation record](../performance/records/glm53-flash/pr146-recurrent-publication-live-validation.md)
+from [pull request #147](https://github.com/FujitsuPolycom/sparkring/pull/147).
+It binds SparkRing `d93cb3d98305041081cf572521602625185112ae` and
+SparkCache `65b6642df1afc64366430d3aef9aca01f5c5e1c3`.
+
+The exact image completed the fixed semantic canary, a clean-restart
+8,192-token persistent restore, a 131,072-token persistent restore, tail-only
+copy-on-write publication from 131,072 to 262,144 tokens, a verified
+262,144-token persistent restore, and one C16 cohort in which 16 distinct
+request tails shared one restored 131,072-token segment. Restore timings are
+**research-only**. The fixed semantic canary is not a DFlash response-quality
+benchmark, so response quality is **unsupported**. Public OCI publication is
+also **unsupported**; the image has no published digest. No rebuild inherits
+these observations without its own exact evidence.
+
+The `ed60...` artifact uses `block_pages_v1`. Its C16 case proves that one
+exact 131,072-token prefix was restored once for requests with 16 distinct
+tails. It does not exercise `per_token_rows` different-root descriptor-segment
+coalescing. That behavior is implemented with GPU-free coverage only and is
+absent from this artifact's live qualification.
+
+### Separate historical artifact
+
+Historical local image ID
 `sha256:eef863d8bc578815a80b0e2d9f0d745102b6363415225101fd92171a2e5a55cb`
 is **qualified** only for the TP4/DCP1 startup, health, semantic generation,
 arbitrary page-boundary replay, and 131,072- and 262,144-token restore cases
@@ -20,7 +47,7 @@ identity and requires its own live checks.
 | vLLM native extensions and wheel metadata | `da4d7be6c97434f6942292ed8abbf4b32dc44355` |
 | vLLM Python source | `0b67266a0f37d6146a8403fb8482403c62f412d5`, tree `ba9484ccb33aa56e90ff2f447f15ca9b9da97639` |
 | B12X | `b1d541f9e71a35f030d45fae437630fff7507c2a`, tree `c69cdec1c59a08e8e0e549f930fa8abcfb5134ae` |
-| SparkCache shared-segment restore, tail-only copy-on-write publication, canonical CUDA configuration, and bounded page-delta reads | `65b6642df1afc64366430d3aef9aca01f5c5e1c3`, tree `41ad0a119ba109fd28900a2dcc9f9b4d8c293809`, clean source SHA-256 `a2add45a9f97446f6c2a843355161da9a5499ff7501b4750d2163591785d7345` |
+| SparkCache exact prefix restore, GPU-free row-descriptor coalescing, tail-only copy-on-write publication, canonical CUDA configuration, and bounded page-delta reads | `65b6642df1afc64366430d3aef9aca01f5c5e1c3`, tree `41ad0a119ba109fd28900a2dcc9f9b4d8c293809`, clean source SHA-256 `a2add45a9f97446f6c2a843355161da9a5499ff7501b4750d2163591785d7345` |
 | Recurrent replay-boundary producer | Patch SHA-256 `5a6561a5bbab990dcd03bfd6a485ea26c3b5a578c2fd61b76305767b16dbfba0`; produces the four postimages accepted by SparkCache lease contract SHA-256 `8adbdfa3fd4b06b213c3aab45255a0b039f1c9940a4b1fad0efd004d263227c9` |
 | 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` |
@@ -68,7 +95,7 @@ Two profiles share the same image and DFlash7 cache identity:
 | Profile | Status | Loader behavior |
 |---|---|---|
 | `glm53-flash-dflash7-python-overlay-safetensors-sparkcache-tp4-dcp1.example.json` | **implemented**, not qualified | Uses global safetensors for target and draft. This follows the qualified-compatible loader shape but still requires live qualification on the composed 0b image. |
-| `glm53-flash-dflash7-python-overlay-fastsafetensors-sparkcache-tp4-dcp1.example.json` | **implemented**, not qualified for the selected source contract | Uses global fastsafetensors with queue size one for the target and `draft_load_config={"load_format":"safetensors"}` for DFlash. The historical receipt remains evidence only for its recorded image. |
+| `glm53-flash-dflash7-python-overlay-fastsafetensors-sparkcache-tp4-dcp1.example.json` | **qualified** only for exact image `sha256:ed60be066d6d9eadea267bc4597a0687869f3ddb95a3e5c6f86649893a838eb8` and the bounded cases above; rebuilds are not qualified | Uses global fastsafetensors with queue size one for the target and `draft_load_config={"load_format":"safetensors"}` for DFlash. |
 
 The image applies an exact-input vLLM patch that passes
 `SpeculativeConfig.draft_load_config` to the DFlash model loader. The image
@@ -115,7 +142,8 @@ python scripts/sparkring_generic_launcher.py \
   plan
 ```
 
-`plan` is offline. Inspect every rank action before a lifecycle command.
+`plan` is offline. Inspect every rank action before a command that changes
+containers.
 
 ## Start and observe the four-rank service
 
@@ -179,8 +207,9 @@ different cache root and one-shot clear token so loader observations remain
 isolated.
 
 The pinned SparkCache source combines canonical CUDA configuration names,
-authenticated shared-segment restore, an eight-worker ordered page-delta
-reader, and tail-only copy-on-write publication. Cache identities, digest
+authenticated exact prefix restore, GPU-free row-descriptor coalescing, an
+eight-worker ordered page-delta reader, and tail-only copy-on-write
+publication. Cache identities, digest
 salts, 256-token chunk geometry, page-delta wire bytes, and the CUDA placement
 ABI are unchanged. The source change does not change the namespace, so
 compatible `page-tail-cow-v1` entries remain eligible.

@@ -14,6 +14,7 @@ qualified functional claim or a research-only observation.
 | [`glm-3.5bpw/temp1/`](glm-3.5bpw/temp1/) | 10 four-Spark TP4/DCP4 sustained-decode and Coding Peak receipts |
 | [`glm53-flash/sparkcache-dflash2-bf16-tp4-20260828/`](glm53-flash/sparkcache-dflash2-bf16-tp4-20260828/) | Sanitized post-restore semantic canary for the TP4/DCP1 SparkCache validation |
 | [`glm53-flash/sparkcache-dflash2-bf16-tp4-20260829/`](glm53-flash/sparkcache-dflash2-bf16-tp4-20260829/) | One accepted 16K prefill and C1 decode observation; capacity-limited C4 and C8 cells are excluded |
+| [`glm53-flash/adaptive-mtp-vs-dflash7-20260829/`](glm53-flash/adaptive-mtp-vs-dflash7-20260829/) | Research-only server-log diagnostic; prompt and output receipts are absent |
 | [`glm53-flash/sparkcache-dflash2-bf16-tp4-20g-20260829/`](glm53-flash/sparkcache-dflash2-bf16-tp4-20g-20260829/) | One accepted functional observation with 20 GiB of GPU KV memory per rank |
 | [`qwen38-27b/temp1/`](qwen38-27b/temp1/) | 13 two-Spark and 16 four-Spark accepted prefill, decode, and Coding Peak receipts |
 

@@ -0,0 +1,96 @@
+{
+  "schema": "sparkring-glm53-decode-log-diagnostic/v1",
+  "status": "research-only",
+  "target_model": {
+    "repository": "local-inference-lab/GLM-5.3-Flash-NVFP4",
+    "revision": "520de24eabf507659eaef7c70f14fd584527facc",
+    "cache_identity_sha256": "a35e6bf2875c1875609b8deaec404c07c6cc80259e4222fc0b51e649498bd6b9"
+  },
+  "topology": {
+    "systems": 4,
+    "tensor_parallel_size": 4,
+    "decode_context_parallel_size": 1,
+    "pipeline_parallel_size": 1
+  },
+  "adaptive_mtp": {
+    "container": "glm53-flash-public-python-overlay-mtp5-adaptive-fastsafetensors-sparkcache-tp4-r0",
+    "image_id": "sha256:d209cf986c1f14f320e53d8b425c5e3a255eef9320f25b632af50f5b5c977314",
+    "served_model": "glm-5.3-flash-nvfp4-python-overlay-0b67266-on-da4d7be-b12x-b1d541f-mtp5-adaptive-tp4",
+    "vllm_compiled_commit": "da4d7be6c97434f6942292ed8abbf4b32dc44355",
+    "vllm_python_commit": "0b67266a0f37d6146a8403fb8482403c62f412d5",
+    "b12x_commit": "b1d541f9e71a35f030d45fae437630fff7507c2a",
+    "sparkcache_commit": "20838ace3ebda570ca039cb7f1976c29da554b39",
+    "sparkring_revision": "914c94d084d6881e90660305dedaa410ef02b167",
+    "draft_identity_sha256": "2e06d909ce5bb71c0c0e3e8be74a70e3b41d92ba4c30196cfb0957fb812acef6",
+    "maximum_depth": 5,
+    "initial_depth": 3,
+    "observation_window_steps": 32,
+    "log_interval_utc": {
+      "start": "2026-08-30T03:23:25Z",
+      "end": "2026-08-30T03:25:15Z"
+    },
+    "observations": [
+      {"timestamp": "2026-08-30T03:23:25Z", "generation_tokens_per_second": 38.7, "depth": 4, "draft_acceptance_percent": 74.2},
+      {"timestamp": "2026-08-30T03:23:35Z", "generation_tokens_per_second": 30.7, "depth": 3, "draft_acceptance_percent": 65.9},
+      {"timestamp": "2026-08-30T03:23:45Z", "generation_tokens_per_second": 34.1, "depth": 3, "draft_acceptance_percent": 62.2},
+      {"timestamp": "2026-08-30T03:23:55Z", "generation_tokens_per_second": 29.5, "depth": 4, "draft_acceptance_percent": 77.0},
+      {"timestamp": "2026-08-30T03:24:05Z", "generation_tokens_per_second": 34.6, "depth": 4, "draft_acceptance_percent": 72.4},
+      {"timestamp": "2026-08-30T03:24:15Z", "generation_tokens_per_second": 31.4, "depth": 3, "draft_acceptance_percent": 59.4},
+      {"timestamp": "2026-08-30T03:24:25Z", "generation_tokens_per_second": 36.3, "depth": 3, "draft_acceptance_percent": 67.5},
+      {"timestamp": "2026-08-30T03:24:35Z", "generation_tokens_per_second": 36.8, "depth": 4, "draft_acceptance_percent": 72.0},
+      {"timestamp": "2026-08-30T03:24:45Z", "generation_tokens_per_second": 28.0, "depth": 3, "draft_acceptance_percent": 58.2},
+      {"timestamp": "2026-08-30T03:24:55Z", "generation_tokens_per_second": 24.5, "depth": 2, "draft_acceptance_percent": 48.9},
+      {"timestamp": "2026-08-30T03:25:05Z", "generation_tokens_per_second": 31.9, "depth": 3, "draft_acceptance_percent": 75.1},
+      {"timestamp": "2026-08-30T03:25:15Z", "generation_tokens_per_second": 37.1, "depth": 4, "draft_acceptance_percent": 81.2}
+    ],
+    "sparkcache_events_in_interval": 0
+  },
+  "dflash7": {
+    "container": "glm53-flash-dflash2-bf16-sparkcache-public-tp4-r0-dflash7-stopped-20260829",
+    "rank_image_ids": [
+      "sha256:80b407904d21211d9a6b435c8ee4292ce66bfebcf802d94deb6605cdbc94d648",
+      "sha256:de4e219db4bdc2dfe4a2bcf2d5adb11655da22b5d40bf42494028c5ed128ea83",
+      "sha256:b5c29a1fc97fe7e086b62cffd832d90b292dc4f60cbf10c5593ef2f3debea365",
+      "sha256:9aeaf31b13af8a940734e31ec2963ce2714db535430ce7953a8baa5cf6dd4401"
+    ],
+    "served_model": "glm-5.3-flash-nvfp4-dflash7-bf16-tp4",
+    "vllm_commit": "da4d7be6c97434f6942292ed8abbf4b32dc44355",
+    "b12x_commit": "2fcf23a0ce269be27b2e03fece73d46e90e6aeea",
+    "sparkcache_commit": "2b86fb9d02fa3595cca5caa864b81aedce44b8bb",
+    "draft_repository": "incoai/GLM-5.3-Flash-DFlash2",
+    "draft_revision": "dc77ff1c99eeb2df044ee3d4f0094eb033fee410",
+    "draft_weights_sha256": "b33c03475ba7322cf398828f2d8d1be376df30dc05c6b40c28c8ea8da23e410b",
+    "depth": 7,
+    "log_interval_utc": {
+      "start": "2026-08-29T21:13:57Z",
+      "end": "2026-08-29T21:15:07Z"
+    },
+    "generation_tokens_per_second": [40.6, 48.2, 52.6, 94.9, 111.8, 69.8, 133.6, 136.0],
+    "sparkcache_events_in_interval": 0,
+    "post_switch_validation": {
+      "log_interval_utc": {
+        "start": "2026-08-30T04:02:49Z",
+        "end": "2026-08-30T04:04:09Z"
+      },
+      "observations": [
+        {"timestamp": "2026-08-30T04:02:49Z", "generation_tokens_per_second": 48.8, "mean_acceptance_length": 4.82, "draft_acceptance_percent": 54.6},
+        {"timestamp": "2026-08-30T04:02:59Z", "generation_tokens_per_second": 61.3, "mean_acceptance_length": 4.61, "draft_acceptance_percent": 51.6},
+        {"timestamp": "2026-08-30T04:03:09Z", "generation_tokens_per_second": 74.5, "mean_acceptance_length": 5.69, "draft_acceptance_percent": 67.0},
+        {"timestamp": "2026-08-30T04:03:19Z", "generation_tokens_per_second": 60.9, "mean_acceptance_length": 4.54, "draft_acceptance_percent": 50.5},
+        {"timestamp": "2026-08-30T04:03:29Z", "generation_tokens_per_second": 61.3, "mean_acceptance_length": 4.51, "draft_acceptance_percent": 50.1},
+        {"timestamp": "2026-08-30T04:03:39Z", "generation_tokens_per_second": 56.3, "mean_acceptance_length": 4.33, "draft_acceptance_percent": 47.6},
+        {"timestamp": "2026-08-30T04:03:49Z", "generation_tokens_per_second": 64.4, "mean_acceptance_length": 5.07, "draft_acceptance_percent": 58.1},
+        {"timestamp": "2026-08-30T04:03:59Z", "generation_tokens_per_second": 72.1, "mean_acceptance_length": 5.30, "draft_acceptance_percent": 61.4},
+        {"timestamp": "2026-08-30T04:04:09Z", "generation_tokens_per_second": 72.2, "mean_acceptance_length": 5.47, "draft_acceptance_percent": 63.9}
+      ],
+      "sparkcache_events_in_interval": 0
+    }
+  },
+  "user_reported_matched_coding_peak": {
+    "dflash7_tokens_per_second": 70.0,
+    "adaptive_mtp_tokens_per_second": 33.5,
+    "receipt_present": false
+  },
+  "prompt_receipt_present": false,
+  "output_receipt_present": false
+}
@@ -0,0 +1,135 @@
+{
+  "schema": "sparkring-glm53-dflash7-pr25-live-validation/v1",
+  "status": {
+    "artifact_construction": "implemented",
+    "startup_health": "qualified",
+    "semantic_smoke": "qualified",
+    "flat_restore": "qualified",
+    "page_delta_restore_correctness": "qualified",
+    "page_delta_restore_performance": "research-only",
+    "persistent_128k_restore": "qualified",
+    "shared_prefix_waves": "qualified",
+    "dflash_quality": "unsupported"
+  },
+  "artifact": {
+    "image": "sparkring-glm53-sparkcache:dflash7-vllm-python-0b67266-native-da4d7be-b12x-b1d541f-arm64",
+    "image_id": "sha256:9faa36a9f37aee16d97ab9214ef3153b4d200121126e6b2dee5ebb63109fea18",
+    "oci_revision_label": "e2d92fdc7d0306d664d6fd9f296dc2adcaf0fe05",
+    "sparkring_revision": "e2d92fdc7d0306d664d6fd9f296dc2adcaf0fe05",
+    "vllm_compiled_revision": "da4d7be6c97434f6942292ed8abbf4b32dc44355",
+    "vllm_python_revision": "0b67266a0f37d6146a8403fb8482403c62f412d5",
+    "b12x_revision": "b1d541f9e71a35f030d45fae437630fff7507c2a",
+    "sparkcache_revision": "5d571018de5b63a9a90e5c11e6d6e86bbff4a957",
+    "sparkcache_tree": "e864ed9ad64f771188fdb59aa9738e348134d636",
+    "sparkcache_source_sha256": "f7c0565521fddeff7085e4cc08043cb8d1e2bde33abc67f83b8608a162d05b88",
+    "cuda_placement_library_sha256": "a2e495162bf3d58b01613cd82ac15c8e15031dd7d6de7299700d2c58d905ada8",
+    "dflash_loader_patch_sha256": "39b567013ee7aed79f63200ed460129587933dc77fb430decdf19f78178de279",
+    "dflash_loader_postimage_sha256": "98acbae2b3bb4482d83f9637c163ce7c92707ccdf6561b7e431f23337f151cf4"
+  },
+  "conditions": {
+    "hardware": "four NVIDIA DGX Spark systems",
+    "topology": {
+      "tensor_parallel_size": 4,
+      "decode_context_parallel_size": 1,
+      "pipeline_parallel_size": 1
+    },
+    "target_loader": "fastsafetensors",
+    "draft_loader": "safetensors",
+    "draft_method": "dflash",
+    "draft_tokens": 7,
+    "draft_tensor_parallel_size": 4,
+    "kv_cache_dtype": "fp8",
+    "vllm_block_size": 256,
+    "max_num_seqs": 32,
+    "sparkcache_publication_schema": "tail-cow-v1",
+    "sparkcache_effective_publication_schema": "page-tail-cow-v1",
+    "sparkcache_config_surface": "PR25 legacy-key compatibility",
+    "sparkcache_accepted_legacy_keys": [
+      "spark_cache_native_restore",
+      "spark_cache_native_library",
+      "spark_cache_native_library_sha256",
+      "spark_cache_native_arena_bytes",
+      "spark_cache_native_io_workers"
+    ]
+  },
+  "measurements": {
+    "startup": {
+      "target_fastsafetensors_seconds": 69.52,
+      "draft_safetensors_seconds": 4.18,
+      "total_model_load_seconds": 79.27,
+      "warm_ready_seconds_approx": 190.0,
+      "healthy_ranks": 4,
+      "restart_count_by_rank": [0, 0, 0, 0],
+      "oom_killed_by_rank": [false, false, false, false]
+    },
+    "semantic_smoke": {
+      "raw_completion_tokens": [2, 2],
+      "scope": "continued raw completion only; no quality score"
+    },
+    "flat_restore": {
+      "restored_tokens": 11520,
+      "rank": 0,
+      "restore_milliseconds": 29.258
+    },
+    "page_delta_restore": {
+      "restored_tokens": 17152,
+      "restore_milliseconds": 703.826,
+      "restore_read_milliseconds": 579.259,
+      "chunk_count": 67,
+      "type_error_observed": false
+    },
+    "prime_128k": {
+      "request_seconds": 55.522,
+      "completion_token": 13,
+      "snapshot_milliseconds": 2440.7,
+      "commit_milliseconds": 1687.7
+    },
+    "unrelated_sentinel": {
+      "completion_token": 271
+    },
+    "persistent_restore_128k": {
+      "restored_tokens": 131072,
+      "bytes_per_rank": 813068464,
+      "rank_restore_milliseconds": {
+        "minimum": 123.69,
+        "maximum": 153.253
+      },
+      "logical_tokens_per_second_at_slowest_rank": 855265.476
+    },
+    "shared_prefix_waves": [
+      {
+        "concurrency": 2,
+        "wall_seconds": 0.808,
+        "http_200": 2,
+        "completion_token": 13,
+        "restore_events_observed": 1,
+        "retention_observed": true
+      },
+      {
+        "concurrency": 8,
+        "wall_seconds": 1.506,
+        "http_200": 8,
+        "completion_token": 13,
+        "restore_events_observed": 1,
+        "retention_observed": true
+      },
+      {
+        "concurrency": 16,
+        "wall_seconds": 1.781,
+        "http_200": 16,
+        "completion_token": 13,
+        "restore_events_observed": 1,
+        "retention_observed": true
+      }
+    ]
+  },
+  "limitations": [
+    "The 7168-to-12032 base-geometry incompatibility is addressed only in draft SparkCache PR28 and remains present in this artifact.",
+    "Null-block publication rejection at 6912 tokens remains under investigation.",
+    "Page-delta read and reconstruction latency is too high for a performance claim.",
+    "No DFlash quality benchmark was run on this artifact.",
+    "One observed wave per concurrency does not establish latency variability, throughput, or soak behavior.",
+    "The collection timestamp, complete command transcript, and independent clock audit are not part of this bounded receipt.",
+    "This receipt qualifies only image sha256:9faa36a9f37aee16d97ab9214ef3153b4d200121126e6b2dee5ebb63109fea18; it does not qualify image sha256:eef863d8bc578815a80b0e2d9f0d745102b6363415225101fd92171a2e5a55cb or image sha256:ed60be066d6d9eadea267bc4597a0687869f3ddb95a3e5c6f86649893a838eb8."
+  ]
+}