From c7a065a7531d0c4281883dffa156bec9959cc440 Mon Sep 17 00:00:00 2001 From: Shiyi Zheng Date: Mon, 13 Jul 2026 12:33:09 +0800 Subject: [PATCH 1/4] Add deepset/roberta-large-squad2 question-answering recipe Add fp16 recipe for deepset/roberta-large-squad2 (question-answering), mirroring the existing deepset/roberta-base-squad2 recipe. Bumps the recipe README total 75 -> 76. Validated CPU-only: genuine fp16 (FLOAT16=394, FLOAT32=0), build 331.4s, perf P50 1002.91ms. All-10-bucket coverage inferred from the certified roberta-base-squad2 QA recipe and certified roberta-large backbone; per-EP CI eval to confirm. --- examples/recipes/README.md | 3 +- .../question-answering_fp16_config.json | 70 +++++++++++++++++++ 2 files changed, 72 insertions(+), 1 deletion(-) create mode 100644 examples/recipes/deepset_roberta-large-squad2/question-answering_fp16_config.json diff --git a/examples/recipes/README.md b/examples/recipes/README.md index 86aa01869..33c10d167 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: **76** (model, task) tuples that pass fp16 eval on all 10 (EP, device) buckets. | Model | Task | |---|---| @@ -42,6 +42,7 @@ Total: **75** (model, task) tuples that pass fp16 eval on all 10 (EP, device) bu | dbmdz/bert-large-cased-finetuned-conll03-english | token-classification | | deepset/bert-large-uncased-whole-word-masking-squad2 | question-answering | | deepset/roberta-base-squad2 | question-answering | +| deepset/roberta-large-squad2 | question-answering | | deepset/tinyroberta-squad2 | question-answering | | dima806/fairface_age_image_detection | image-classification | | distilbert/distilbert-base-cased-distilled-squad | question-answering | diff --git a/examples/recipes/deepset_roberta-large-squad2/question-answering_fp16_config.json b/examples/recipes/deepset_roberta-large-squad2/question-answering_fp16_config.json new file mode 100644 index 000000000..1c8a70440 --- /dev/null +++ b/examples/recipes/deepset_roberta-large-squad2/question-answering_fp16_config.json @@ -0,0 +1,70 @@ +{ + "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, + 50265 + ] + }, + { + "name": "attention_mask", + "dtype": "int32", + "shape": [ + 1, + 512 + ], + "value_range": [ + 0, + 2 + ] + } + ], + "output_tensors": [ + { + "name": "start_logits" + }, + { + "name": "end_logits" + } + ] + }, + "optim": { + "clamp_constant_values": true + }, + "quant": null, + "loader": { + "task": "question-answering", + "model_class": "AutoModelForQuestionAnswering", + "model_type": "roberta" + }, + "eval": { + "task": "question-answering", + "dataset": { + "path": "rajpurkar/squad_v2", + "split": "validation", + "samples": 100, + "columns_mapping": { + "question_column": "question", + "context_column": "context", + "id_column": "id", + "label_column": "answers" + } + } + } +} From 26d3c4f69a71064cbad0c30e8654f4af91338393 Mon Sep 17 00:00:00 2001 From: Shiyi Zheng Date: Mon, 13 Jul 2026 13:47:36 +0800 Subject: [PATCH 2/4] recipe(roberta-large-squad2): nest under cpu/cpu/ per tested-EP layout (_meta-058) --- .../{ => cpu/cpu}/question-answering_fp16_config.json | 0 1 file changed, 0 insertions(+), 0 deletions(-) rename examples/recipes/deepset_roberta-large-squad2/{ => cpu/cpu}/question-answering_fp16_config.json (100%) diff --git a/examples/recipes/deepset_roberta-large-squad2/question-answering_fp16_config.json b/examples/recipes/deepset_roberta-large-squad2/cpu/cpu/question-answering_fp16_config.json similarity index 100% rename from examples/recipes/deepset_roberta-large-squad2/question-answering_fp16_config.json rename to examples/recipes/deepset_roberta-large-squad2/cpu/cpu/question-answering_fp16_config.json From 6106a96115bec8da8d04b6d436bd8ca9bb568486 Mon Sep 17 00:00:00 2001 From: Shiyi Zheng Date: Mon, 13 Jul 2026 20:12:59 +0800 Subject: [PATCH 3/4] recipe: drop README index-row change (recipe-only PR; not certified for all-10-bucket README) --- examples/recipes/README.md | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/examples/recipes/README.md b/examples/recipes/README.md index 33c10d167..86aa01869 100644 --- a/examples/recipes/README.md +++ b/examples/recipes/README.md @@ -14,7 +14,7 @@ Each *(model, task)* includes: ## Models -Total: **76** (model, task) tuples that pass fp16 eval on all 10 (EP, device) buckets. +Total: **75** (model, task) tuples that pass fp16 eval on all 10 (EP, device) buckets. | Model | Task | |---|---| @@ -42,7 +42,6 @@ Total: **76** (model, task) tuples that pass fp16 eval on all 10 (EP, device) bu | dbmdz/bert-large-cased-finetuned-conll03-english | token-classification | | deepset/bert-large-uncased-whole-word-masking-squad2 | question-answering | | deepset/roberta-base-squad2 | question-answering | -| deepset/roberta-large-squad2 | question-answering | | deepset/tinyroberta-squad2 | question-answering | | dima806/fairface_age_image_detection | image-classification | | distilbert/distilbert-base-cased-distilled-squad | question-answering | From 40e9971b2398ce32567fe0f9ecbf373b5f98359e Mon Sep 17 00:00:00 2001 From: "Shiyi Zheng (from Dev Box)" Date: Thu, 16 Jul 2026 15:44:31 +0800 Subject: [PATCH 4/4] recipe(roberta-large-squad2): add verified CPU fp32 tuple --- .../cpu/question-answering_fp32_config.json | 70 +++++++++++++++++++ 1 file changed, 70 insertions(+) create mode 100644 examples/recipes/deepset_roberta-large-squad2/cpu/cpu/question-answering_fp32_config.json diff --git a/examples/recipes/deepset_roberta-large-squad2/cpu/cpu/question-answering_fp32_config.json b/examples/recipes/deepset_roberta-large-squad2/cpu/cpu/question-answering_fp32_config.json new file mode 100644 index 000000000..1c8a70440 --- /dev/null +++ b/examples/recipes/deepset_roberta-large-squad2/cpu/cpu/question-answering_fp32_config.json @@ -0,0 +1,70 @@ +{ + "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, + 50265 + ] + }, + { + "name": "attention_mask", + "dtype": "int32", + "shape": [ + 1, + 512 + ], + "value_range": [ + 0, + 2 + ] + } + ], + "output_tensors": [ + { + "name": "start_logits" + }, + { + "name": "end_logits" + } + ] + }, + "optim": { + "clamp_constant_values": true + }, + "quant": null, + "loader": { + "task": "question-answering", + "model_class": "AutoModelForQuestionAnswering", + "model_type": "roberta" + }, + "eval": { + "task": "question-answering", + "dataset": { + "path": "rajpurkar/squad_v2", + "split": "validation", + "samples": 100, + "columns_mapping": { + "question_column": "question", + "context_column": "context", + "id_column": "id", + "label_column": "answers" + } + } + } +}