From 1cc1caa38f026b87817c00282e959573fe9686da Mon Sep 17 00:00:00 2001 From: "github-actions[bot]" Date: Tue, 14 Jul 2026 18:11:50 +0800 Subject: [PATCH 1/4] recipe(mpnet): add dell-research-harvard/lt-wikidata-comp-en (feature-extraction, sentence-similarity) Add new fp16 and w8a16 recipes for the LinkTransformer mpnet embedder, mirroring the sentence-transformers/multi-qa-mpnet-base-dot-v1 template. Recipe-only contribution: mpnet is already vendor-supported with no winml code changes; CPU build and perf were verified. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com> --- examples/recipes/README.md | 4 +- .../feature-extraction_fp16_config.json | 64 +++++++++++++++ .../feature-extraction_w8a16_config.json | 82 +++++++++++++++++++ .../sentence-similarity_fp16_config.json | 64 +++++++++++++++ .../sentence-similarity_w8a16_config.json | 82 +++++++++++++++++++ 5 files changed, 295 insertions(+), 1 deletion(-) create mode 100644 examples/recipes/dell-research-harvard_lt-wikidata-comp-en/feature-extraction_fp16_config.json create mode 100644 examples/recipes/dell-research-harvard_lt-wikidata-comp-en/feature-extraction_w8a16_config.json create mode 100644 examples/recipes/dell-research-harvard_lt-wikidata-comp-en/sentence-similarity_fp16_config.json create mode 100644 examples/recipes/dell-research-harvard_lt-wikidata-comp-en/sentence-similarity_w8a16_config.json diff --git a/examples/recipes/README.md b/examples/recipes/README.md index 1077e4f74..6a276f368 100644 --- a/examples/recipes/README.md +++ b/examples/recipes/README.md @@ -14,7 +14,7 @@ Each *(model, task)* includes: ## Models -Total: **75** (model, task) tuples that pass fp16 eval on all 10 (EP, device) buckets. +Total: **77** (model, task) tuples with curated recipes. 75 pass fp16 eval on all 10 (EP, device) buckets; dell-research-harvard/lt-wikidata-comp-en (feature-extraction, sentence-similarity) ships fp16 + w8a16 mpnet recipes verified via build and CPU perf, with full-bucket fp16 eval pending. | Model | Task | |---|---| @@ -40,6 +40,8 @@ Total: **75** (model, task) tuples that pass fp16 eval on all 10 (EP, device) bu | cardiffnlp/twitter-roberta-base-sentiment-latest | text-classification | | dbmdz/bert-large-cased-finetuned-conll03-english | token-classification | | deepset/bert-large-uncased-whole-word-masking-squad2 | question-answering | +| dell-research-harvard/lt-wikidata-comp-en | feature-extraction | +| dell-research-harvard/lt-wikidata-comp-en | sentence-similarity | | deepset/roberta-base-squad2 | question-answering | | deepset/tinyroberta-squad2 | question-answering | | dima806/fairface_age_image_detection | image-classification | diff --git a/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/feature-extraction_fp16_config.json b/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/feature-extraction_fp16_config.json new file mode 100644 index 000000000..e731b2a2a --- /dev/null +++ b/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/feature-extraction_fp16_config.json @@ -0,0 +1,64 @@ +{ + "export": { + "opset_version": 17, + "batch_size": 1, + "export_params": true, + "do_constant_folding": true, + "verbose": false, + "dynamo": false, + "enable_hierarchy_tags": true, + "clean_onnx": false, + "hierarchy_tag_format": "full", + "input_tensors": [ + { + "name": "input_ids", + "dtype": "int32", + "shape": [ + 1, + 512 + ], + "value_range": [ + 0, + 30527 + ] + }, + { + "name": "attention_mask", + "dtype": "int32", + "shape": [ + 1, + 512 + ], + "value_range": [ + 0, + 2 + ] + } + ], + "output_tensors": [ + { + "name": "last_hidden_state" + } + ] + }, + "optim": {}, + "quant": null, + "loader": { + "task": "feature-extraction", + "model_class": "AutoModel", + "model_type": "mpnet" + }, + "eval": { + "task": "feature-extraction", + "dataset": { + "path": "mteb/stsbenchmark-sts", + "split": "test", + "samples": 100, + "columns_mapping": { + "input_column_1": "sentence1", + "input_column_2": "sentence2", + "score_column": "score" + } + } + } +} diff --git a/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/feature-extraction_w8a16_config.json b/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/feature-extraction_w8a16_config.json new file mode 100644 index 000000000..229af3273 --- /dev/null +++ b/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/feature-extraction_w8a16_config.json @@ -0,0 +1,82 @@ +{ + "export": { + "opset_version": 17, + "batch_size": 1, + "export_params": true, + "do_constant_folding": true, + "verbose": false, + "dynamo": false, + "enable_hierarchy_tags": true, + "clean_onnx": false, + "hierarchy_tag_format": "full", + "input_tensors": [ + { + "name": "input_ids", + "dtype": "int32", + "shape": [ + 1, + 512 + ], + "value_range": [ + 0, + 30527 + ] + }, + { + "name": "attention_mask", + "dtype": "int32", + "shape": [ + 1, + 512 + ], + "value_range": [ + 0, + 2 + ] + } + ], + "output_tensors": [ + { + "name": "last_hidden_state" + } + ] + }, + "optim": {}, + "quant": { + "mode": "qdq", + "samples": 10, + "calibration_method": "minmax", + "weight_type": "uint8", + "activation_type": "uint16", + "per_channel": false, + "symmetric": false, + "save_calibration": false, + "distribution": "uniform", + "seed": null, + "calibration_load_path": null, + "calibration_save_path": null, + "op_types_to_quantize": null, + "nodes_to_exclude": null, + "task": "feature-extraction", + "model_id": "dell-research-harvard/lt-wikidata-comp-en" + }, + "compile": null, + "loader": { + "task": "feature-extraction", + "model_class": "AutoModel", + "model_type": "mpnet" + }, + "eval": { + "task": "feature-extraction", + "dataset": { + "path": "mteb/stsbenchmark-sts", + "split": "test", + "samples": 100, + "columns_mapping": { + "input_column_1": "sentence1", + "input_column_2": "sentence2", + "score_column": "score" + } + } + } +} diff --git a/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/sentence-similarity_fp16_config.json b/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/sentence-similarity_fp16_config.json new file mode 100644 index 000000000..25b22809c --- /dev/null +++ b/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/sentence-similarity_fp16_config.json @@ -0,0 +1,64 @@ +{ + "export": { + "opset_version": 17, + "batch_size": 1, + "export_params": true, + "do_constant_folding": true, + "verbose": false, + "dynamo": false, + "enable_hierarchy_tags": true, + "clean_onnx": false, + "hierarchy_tag_format": "full", + "input_tensors": [ + { + "name": "input_ids", + "dtype": "int32", + "shape": [ + 1, + 512 + ], + "value_range": [ + 0, + 30527 + ] + }, + { + "name": "attention_mask", + "dtype": "int32", + "shape": [ + 1, + 512 + ], + "value_range": [ + 0, + 2 + ] + } + ], + "output_tensors": [ + { + "name": "last_hidden_state" + } + ] + }, + "optim": {}, + "quant": null, + "loader": { + "task": "sentence-similarity", + "model_class": "AutoModel", + "model_type": "mpnet" + }, + "eval": { + "task": "sentence-similarity", + "dataset": { + "path": "mteb/stsbenchmark-sts", + "split": "test", + "samples": 100, + "columns_mapping": { + "input_column_1": "sentence1", + "input_column_2": "sentence2", + "score_column": "score" + } + } + } +} diff --git a/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/sentence-similarity_w8a16_config.json b/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/sentence-similarity_w8a16_config.json new file mode 100644 index 000000000..0de709cc2 --- /dev/null +++ b/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/sentence-similarity_w8a16_config.json @@ -0,0 +1,82 @@ +{ + "export": { + "opset_version": 17, + "batch_size": 1, + "export_params": true, + "do_constant_folding": true, + "verbose": false, + "dynamo": false, + "enable_hierarchy_tags": true, + "clean_onnx": false, + "hierarchy_tag_format": "full", + "input_tensors": [ + { + "name": "input_ids", + "dtype": "int32", + "shape": [ + 1, + 512 + ], + "value_range": [ + 0, + 30527 + ] + }, + { + "name": "attention_mask", + "dtype": "int32", + "shape": [ + 1, + 512 + ], + "value_range": [ + 0, + 2 + ] + } + ], + "output_tensors": [ + { + "name": "last_hidden_state" + } + ] + }, + "optim": {}, + "quant": { + "mode": "qdq", + "samples": 10, + "calibration_method": "minmax", + "weight_type": "uint8", + "activation_type": "uint16", + "per_channel": false, + "symmetric": false, + "save_calibration": false, + "distribution": "uniform", + "seed": null, + "calibration_load_path": null, + "calibration_save_path": null, + "op_types_to_quantize": null, + "nodes_to_exclude": null, + "task": "sentence-similarity", + "model_id": "dell-research-harvard/lt-wikidata-comp-en" + }, + "compile": null, + "loader": { + "task": "sentence-similarity", + "model_class": "AutoModel", + "model_type": "mpnet" + }, + "eval": { + "task": "sentence-similarity", + "dataset": { + "path": "mteb/stsbenchmark-sts", + "split": "test", + "samples": 1000, + "columns_mapping": { + "input_column_1": "sentence1", + "input_column_2": "sentence2", + "score_column": "score" + } + } + } +} From 225e93bba08f57f88626ccfc3771c6fd1f566f94 Mon Sep 17 00:00:00 2001 From: "github-actions[bot]" Date: Wed, 15 Jul 2026 11:46:21 +0800 Subject: [PATCH 2/4] recipe(mpnet): nest lt-wikidata-comp-en recipes under cpu/cpu (tested EP/device) Move the 4 feature-extraction/sentence-similarity fp16+w8a16 recipes from the flat model dir into cpu/cpu/, reflecting the only EP/device combo actually validated (build + CPU perf; GPU/NPU host-blocked). Follows the _meta-051 one-folder-per-EP+device layout. Note: the e2e eval discover_recipe_variants uses a flat, non-recursive glob, so these nested recipes are not auto-discovered until the harness is taught to recurse. Contents unchanged; verified a recipe-driven fp16 build still succeeds from the new path. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com> --- .../{ => cpu/cpu}/feature-extraction_fp16_config.json | 0 .../{ => cpu/cpu}/feature-extraction_w8a16_config.json | 0 .../{ => cpu/cpu}/sentence-similarity_fp16_config.json | 0 .../{ => cpu/cpu}/sentence-similarity_w8a16_config.json | 0 4 files changed, 0 insertions(+), 0 deletions(-) rename examples/recipes/dell-research-harvard_lt-wikidata-comp-en/{ => cpu/cpu}/feature-extraction_fp16_config.json (100%) rename examples/recipes/dell-research-harvard_lt-wikidata-comp-en/{ => cpu/cpu}/feature-extraction_w8a16_config.json (100%) rename examples/recipes/dell-research-harvard_lt-wikidata-comp-en/{ => cpu/cpu}/sentence-similarity_fp16_config.json (100%) rename examples/recipes/dell-research-harvard_lt-wikidata-comp-en/{ => cpu/cpu}/sentence-similarity_w8a16_config.json (100%) diff --git a/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/feature-extraction_fp16_config.json b/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/feature-extraction_fp16_config.json similarity index 100% rename from examples/recipes/dell-research-harvard_lt-wikidata-comp-en/feature-extraction_fp16_config.json rename to examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/feature-extraction_fp16_config.json diff --git a/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/feature-extraction_w8a16_config.json b/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/feature-extraction_w8a16_config.json similarity index 100% rename from examples/recipes/dell-research-harvard_lt-wikidata-comp-en/feature-extraction_w8a16_config.json rename to examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/feature-extraction_w8a16_config.json diff --git a/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/sentence-similarity_fp16_config.json b/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/sentence-similarity_fp16_config.json similarity index 100% rename from examples/recipes/dell-research-harvard_lt-wikidata-comp-en/sentence-similarity_fp16_config.json rename to examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/sentence-similarity_fp16_config.json diff --git a/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/sentence-similarity_w8a16_config.json b/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/sentence-similarity_w8a16_config.json similarity index 100% rename from examples/recipes/dell-research-harvard_lt-wikidata-comp-en/sentence-similarity_w8a16_config.json rename to examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/sentence-similarity_w8a16_config.json From 3d65e747b268a74e86c239f9179786324c0deb1d Mon Sep 17 00:00:00 2001 From: "github-actions[bot]" Date: Wed, 15 Jul 2026 17:03:51 +0800 Subject: [PATCH 3/4] recipe(mpnet): revert examples/recipes/README.md per review Reviewer requested reverting the README changes (will be emphasized on the skill later). Restores the Total line to 75 and removes the two dell-research-harvard/lt-wikidata-comp-en rows. The four recipe JSON configs under cpu/cpu/ are unchanged. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com> --- examples/recipes/README.md | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) diff --git a/examples/recipes/README.md b/examples/recipes/README.md index 6a276f368..1077e4f74 100644 --- a/examples/recipes/README.md +++ b/examples/recipes/README.md @@ -14,7 +14,7 @@ Each *(model, task)* includes: ## Models -Total: **77** (model, task) tuples with curated recipes. 75 pass fp16 eval on all 10 (EP, device) buckets; dell-research-harvard/lt-wikidata-comp-en (feature-extraction, sentence-similarity) ships fp16 + w8a16 mpnet recipes verified via build and CPU perf, with full-bucket fp16 eval pending. +Total: **75** (model, task) tuples that pass fp16 eval on all 10 (EP, device) buckets. | Model | Task | |---|---| @@ -40,8 +40,6 @@ Total: **77** (model, task) tuples with curated recipes. 75 pass fp16 eval on al | cardiffnlp/twitter-roberta-base-sentiment-latest | text-classification | | dbmdz/bert-large-cased-finetuned-conll03-english | token-classification | | deepset/bert-large-uncased-whole-word-masking-squad2 | question-answering | -| dell-research-harvard/lt-wikidata-comp-en | feature-extraction | -| dell-research-harvard/lt-wikidata-comp-en | sentence-similarity | | deepset/roberta-base-squad2 | question-answering | | deepset/tinyroberta-squad2 | question-answering | | dima806/fairface_age_image_detection | image-classification | From 1f8fff4101f0eaef2fc2d56a16e5df8edc6abedf Mon Sep 17 00:00:00 2001 From: "github-actions[bot]" Date: Wed, 15 Jul 2026 18:51:39 +0800 Subject: [PATCH 4/4] recipe(mpnet): cpu/cpu ships fp32+fp16, drop w8a16 for lt-wikidata-comp-en Per CPU precision policy: the cpu/cpu bucket ships both float precisions (fp32 + fp16, quant:null) and no w8a16 - quantization regressed vs float on CPU with no int8/VNNI benefit. - add feature-extraction_fp32_config.json + sentence-similarity_fp32_config.json (byte-identical to the fp16 recipes; quant:null) - remove feature-extraction_w8a16_config.json + sentence-similarity_w8a16_config.json - fp16 recipes unchanged Verified: recipe-driven build of feature-extraction_fp32 completed in 181.6s, model.onnx 427.5 MB (fp32 initializers on CPU by design). Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com> --- ...on => feature-extraction_fp32_config.json} | 20 +---------------- ...n => sentence-similarity_fp32_config.json} | 22 ++----------------- 2 files changed, 3 insertions(+), 39 deletions(-) rename examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/{feature-extraction_w8a16_config.json => feature-extraction_fp32_config.json} (69%) rename examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/{sentence-similarity_w8a16_config.json => sentence-similarity_fp32_config.json} (67%) diff --git a/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/feature-extraction_w8a16_config.json b/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/feature-extraction_fp32_config.json similarity index 69% rename from examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/feature-extraction_w8a16_config.json rename to examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/feature-extraction_fp32_config.json index 229af3273..e731b2a2a 100644 --- a/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/feature-extraction_w8a16_config.json +++ b/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/feature-extraction_fp32_config.json @@ -42,25 +42,7 @@ ] }, "optim": {}, - "quant": { - "mode": "qdq", - "samples": 10, - "calibration_method": "minmax", - "weight_type": "uint8", - "activation_type": "uint16", - "per_channel": false, - "symmetric": false, - "save_calibration": false, - "distribution": "uniform", - "seed": null, - "calibration_load_path": null, - "calibration_save_path": null, - "op_types_to_quantize": null, - "nodes_to_exclude": null, - "task": "feature-extraction", - "model_id": "dell-research-harvard/lt-wikidata-comp-en" - }, - "compile": null, + "quant": null, "loader": { "task": "feature-extraction", "model_class": "AutoModel", diff --git a/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/sentence-similarity_w8a16_config.json b/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/sentence-similarity_fp32_config.json similarity index 67% rename from examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/sentence-similarity_w8a16_config.json rename to examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/sentence-similarity_fp32_config.json index 0de709cc2..25b22809c 100644 --- a/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/sentence-similarity_w8a16_config.json +++ b/examples/recipes/dell-research-harvard_lt-wikidata-comp-en/cpu/cpu/sentence-similarity_fp32_config.json @@ -42,25 +42,7 @@ ] }, "optim": {}, - "quant": { - "mode": "qdq", - "samples": 10, - "calibration_method": "minmax", - "weight_type": "uint8", - "activation_type": "uint16", - "per_channel": false, - "symmetric": false, - "save_calibration": false, - "distribution": "uniform", - "seed": null, - "calibration_load_path": null, - "calibration_save_path": null, - "op_types_to_quantize": null, - "nodes_to_exclude": null, - "task": "sentence-similarity", - "model_id": "dell-research-harvard/lt-wikidata-comp-en" - }, - "compile": null, + "quant": null, "loader": { "task": "sentence-similarity", "model_class": "AutoModel", @@ -71,7 +53,7 @@ "dataset": { "path": "mteb/stsbenchmark-sts", "split": "test", - "samples": 1000, + "samples": 100, "columns_mapping": { "input_column_1": "sentence1", "input_column_2": "sentence2",