Publish ettin-150m-memory-reranker-ft-v1 (qualified ONNX export and complete attribution)

MODIFICATIONS.md

@@ -14,20 +14,23 @@ What changed:
   with rank-16 LoRA+. The adapter was merged into the model weights.
   Relevance grades were frontier-model judgments under a frozen protocol,
   not human annotations.
-- **Export.** The merged model was exported to ONNX as one graph taking
+- **Export.** The merged model was exported to ONNX graphs taking
   `input_ids` and `attention_mask` and emitting one score per query-passage
-  pair, in FP16 for the CUDA realization, with the parameters stored as
-  external data.
-- **Serving contract.** `serving.json` records the runtime and sequence
-  geometry the graph was qualified under: 1,153 total pair tokens, batch 16,
-  CUDA first with CPU as the floor.
+  pair. CUDA uses FP16. DirectML widens the same stored weights and floating
+  calculations to FP32, and uses compatible inferred reshapes. Both graphs
+  share one external parameter file; the CUDA graph and weights are unchanged.
+- **Serving contracts.** `serving.json` and `serving.directml.json` use the
+  same descriptor format for each graph: 1,153 total pair tokens and batch
+  ceilings of 16 on CUDA and 8 on DirectML. CPU handles permitted control
+  operations, not a fallback execution of the whole reranker.
 
 Files modified or generated by Daecore:
 
 - `model.onnx`: the generated ONNX graph of the fine-tuned reranker;
-- `model.onnx.data`: the fine-tuned parameters used by that graph;
+- `model.directml.onnx`: the FP32 calculation graph for DirectML;
+- `model.onnx.data`: the fine-tuned parameters shared by both graphs;
 - `config.json`: the export configuration of the modified reranker;
-- `serving.json`: the Daecore runtime and sequence-geometry contract.
+- `serving.json` and `serving.directml.json`: the runtime and sequence contracts.
 
 Unchanged: `tokenizer.json` and `tokenizer_config.json` preserve the upstream
 tokenizer contract.

README.md

@@ -40,8 +40,8 @@ The limitations section spells out each.
 | Input | one query and one passage, 1,153 tokens total |
 | Pool | a fixed 50-passage candidate set; the model reorders it and never adds to it |
 | Output | one relevance score per pair, used for ordering |
-| Serving precision | FP16 through ONNX Runtime with the CUDA overlay |
-| Deployment state | protected gate passed; package qualified; not yet published or activated |
+| Serving precision | FP16 on CUDA; FP32 calculations on DirectML, sharing the same stored weights |
+| Deployment state | CUDA release plus an approved DirectML graph; installation is a separate update |
 
 ## Intended use
 
@@ -126,7 +126,50 @@ sits at rank 47. On that panel the export reranked a pool in 0.584 s median
 and 0.912 s at the 95th percentile, and the Torch FP16 model in 0.532 s and
 0.777 s.
 
-## Limitations, ranked
+### DirectML comparison
+
+The DirectML package adds `model.directml.onnx` alongside the unchanged CUDA
+`model.onnx`. Both read the same `model.onnx.data`; no second parameter file
+or training framework is needed at runtime. DirectML uses full-precision
+calculations because the half-precision graph had larger score differences.
+CUDA keeps its faster half-precision graph. Each provider plans batches from
+available memory and pair length, with measured ceilings of 16 rows for CUDA
+and 8 for DirectML.
+
+On a current 970-query replay (48,500 fixed-pool pairs), DirectML completed
+without a failure: 2.231 seconds median, 3.247 seconds at the 95th percentile,
+and 4.286 seconds maximum on an RTX 3060 Ti. No call exceeded ten seconds.
+This panel retains its per-query evidence; 2,498 pairs are unjudged, and their
+grades remain unknown. It is separate from the lost historical gate evidence.
+
+Against the published CUDA graph, nDCG@10 changed by +0.000017. Eight first
+choices and 36 top-ten sets changed. Both versions returned known-useful
+evidence for 816 of 970 queries; DirectML returned 5,528 known-useful passages
+versus 5,524. The unchanged calibration was applied for comparison only; its
+DirectML registration is not implied by these results.
+
+Two strict cross-precision checks fail and were explicitly accepted:
+Hit@1 and Hit@5 each lose one query, and one passage moves five positions
+against a limit of four. The affected useful passages remain in the returned
+sets; the five-position move is rank 33 to 28, outside the returned three.
+On the 482 fully judged pools, no Hit metric decreases.
+
+Running the same FP32 graph on CUDA for all 165 queries with changed
+deliveries or relevant ranking differences isolates the provider error:
+8,250 pairs have a 95th-percentile score difference of 0.00000954 and a
+maximum of 0.000149. One pair exceeds the diagnostic 0.0001 bound; one full
+ordering changes and no first choice changes. The two Hit losses and the
+five-position move also occur on CUDA FP32, identifying precision rather
+than DirectML as their cause.
+
+On the supported sequential build policy, concurrent embedding and recall
+kept reranking below 3.2 seconds in three rounds. Eight queued recalls across
+four threads completed within 9.8 seconds. Forcing all three GPU models to
+remain resident on this 8 GB card exceeded the memory budget and failed the
+ten-second limit; those forced-overload timings are not a passing result.
+Worker-kill recovery completed for reranking, embedding, and classification.
+
+## Limitations, ranked
 
 1. **Model-judged relevance.** No human grades; session agreement 0.858 and
    kappa 0.682 on a 419-query audit. Scores measure agreement with that
@@ -142,14 +185,16 @@ and 0.912 s at the 95th percentile, and the Torch FP16 model in 0.532 s and
 5. **First-stage ceiling.** 112 of 553 development queries and 95 of 990 gate
    queries have no useful passage in any top-50 pool; ranks 51 to 200 were
    never judged.
-6. **No CPU or DirectML realization is qualified.** Hosts without the CUDA
-   provider serve the fused first-stage order with a typed notice.
+6. **Provider limits.** CPU reranking is not qualified. DirectML has been
+   measured on an NVIDIA RTX 3060 Ti, not AMD or Intel hardware, and its
+   cross-precision policy exceptions above were accepted with disclosure.
+   Hosts without a qualified reranker serve fusion order with a typed notice.
 7. **The reproducible parity read fails the frozen policy.** The recorded
    development-panel pass cannot be re-run, and the 2026-09-04 recheck on the
    training partition fails two checks at exact FP16 ties. The operator
    accepted the recorded qualification with this disclosure rather than
-   exporting an FP32 graph or amending the policy; an FP32 export and re-run
-   would settle it.
+   exporting an FP32 graph or amending the policy. The current FP32 provider
+   comparison is separate evidence and does not erase that historical result.
 
 ## Reproducibility
 
@@ -161,7 +206,10 @@ and 0.912 s at the 95th percentile, and the Torch FP16 model in 0.532 s and
 - Recipe id `final-listnet-adjacent-loraplus16-full-coverage-w16-lr5e5`;
   registry manifest `../model_registry/ettin-150m-memory-reranker-ft-v1.json`;
   experiment log `../history/README.md`.
-- Serving graph SHA-256 `ce05a6a810007372c6de0f4c56dc94f652e1d28f2eeddf1878b64a170b4c48c2`.
+- CUDA graph SHA-256 `7df4a3b3cd89d216c016abd7d169f0fd215e9f79c06a2ffed57bac1abf556073`.
+- DirectML graph SHA-256 `05abdc03628b599712a993381919aef1e22fcc5ad6384608abc2e069c402e369`.
+- Shared weights SHA-256 `78ddab681c2f506d62bbcd475fec5c10c62d867ed60d97a0170828efaff5bdde`.
+- CUDA serving-contract SHA-256 `ce05a6a810007372c6de0f4c56dc94f652e1d28f2eeddf1878b64a170b4c48c2`.
 - Build stack: Transformers 5.2.0, Sentence Transformers 5.5.1, PyTorch
   2.10.0; export opset 18.
 

model.directml.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:05abdc03628b599712a993381919aef1e22fcc5ad6384608abc2e069c402e369
+size 2242802

publication-manifest.json

@@ -1,6 +1,6 @@
 {
   "schema": "daecore.retrieval-publication-manifest",
-  "staged_at": "2026-09-07T00:27:07+00:00",
+  "tool_sha256": "2369a55f4fa3f007f800834880b6e4a65624c91589b469f48133ebc058c3d2ef",
   "model_id": "ettin-150m-memory-reranker-ft-v1",
   "public_repo": "Daecore/ettin-150m-memory-reranker-ft-v1",
   "role": "cross-encoder-reranker",
@@ -9,15 +9,21 @@
     "revision": "025501c4e0f9bbeb4c5b198318e0089ff061cc14",
     "tree_sha256": "e0edfb38cc495f23507de50233d0965f4e72ad0872e70e6f4e8fc889d394ea01"
   },
-  "export_receipt_sha256": "0b9698df82a890a1688ab23dca2e32182fbccd31f6b460deea0222678d8ce99c",
+  "export_receipt_sha256": "4faf07d568cc5da6400e9aa2645d73b2a2562cb720742c64106b2a70f43b12ba",
   "serving_sha256": "ce05a6a810007372c6de0f4c56dc94f652e1d28f2eeddf1878b64a170b4c48c2",
   "source_model_tree_sha256": "bd8028b5dbc6ab2cd2ab9a6019346de7b25536f50022617194bcb8503d4467ad",
+  "staged_at": "2026-09-11T14:57:19+00:00",
   "files": {
     "config.json": {
       "sha256": "a897acffc99f97238bb5bc0e1bccd6b7fe8fdfebe1450d4db3a9ab026cce6795",
       "size": 2020,
       "binding": "export receipt (file hash)"
     },
+    "model.directml.onnx": {
+      "sha256": "05abdc03628b599712a993381919aef1e22fcc5ad6384608abc2e069c402e369",
+      "size": 2242802,
+      "binding": "export receipt (file hash)"
+    },
     "model.onnx": {
       "sha256": "7df4a3b3cd89d216c016abd7d169f0fd215e9f79c06a2ffed57bac1abf556073",
       "size": 2231565,
@@ -28,6 +34,11 @@
       "size": 299270144,
       "binding": "export receipt (file hash)"
     },
+    "serving.directml.json": {
+      "sha256": "7a43e53663efc6e82db9b29a13fa3775a35d73bd6731e1163601f8f36748925f",
+      "size": 711,
+      "binding": "export receipt (file hash)"
+    },
     "serving.json": {
       "sha256": "381c2fba69bd65cd1e90eac10b4048f1bcaac35f77bac358709194e05432f742",
       "size": 704,
@@ -44,8 +55,8 @@
       "binding": "export receipt (file hash)"
     },
     "README.md": {
-      "sha256": "7b2135ffa768c9d919fd8d71ab47fcb7971ba7bfc9e7f33b00183b533e43bb9b",
-      "size": 8755,
+      "sha256": "19b97e33bddbcb587446bbbb164cd091fc2cb534911cf0aafb0efc8013bda68c",
+      "size": 11806,
       "binding": "packaging record"
     },
     "LICENSE": {
@@ -59,8 +70,8 @@
       "binding": "packaging record"
     },
     "MODIFICATIONS.md": {
-      "sha256": "128a860497b8e31ea659025420a0ca7861b2481bf3d58fda1ad9fe4bcea7710c",
-      "size": 1639,
+      "sha256": "1f8c677387fb72ef31b7a107e866868a13821b6282debebc6f8150eac9866b18",
+      "size": 1958,
       "binding": "packaging record"
     }
   }

serving.directml.json

@@ -0,0 +1,25 @@
+{
+  "external_data": "model.onnx.data",
+  "geometry": {
+    "dynamic_batch": true,
+    "dynamic_sequence": true,
+    "max_batch": 8,
+    "max_sequence": 1153
+  },
+  "graph": "model.directml.onnx",
+  "inputs": [
+    "input_ids",
+    "attention_mask"
+  ],
+  "output": "scores",
+  "precision": "float32",
+  "provider_chain": [
+    "DmlExecutionProvider",
+    "CPUExecutionProvider"
+  ],
+  "role": "cross-encoder-reranker",
+  "schema": "daecore.retrieval-onnx-serving.v1",
+  "sha256": "b83ffa3a44f19d145ecb87e36a56da290b25def03db267e33c1c57a7a46dda76",
+  "source_model_tree_sha256": "bd8028b5dbc6ab2cd2ab9a6019346de7b25536f50022617194bcb8503d4467ad",
+  "torch_required_at_runtime": false
+}