Fix red CI: benchmark script (#34351)

* dont'trigger always

* fux

* oups

* update

* ??

* ?

* aie

.github/workflows/benchmark.yml

@@ -18,21 +18,17 @@ jobs:
     name: Benchmark
     runs-on:
       group: aws-g5-4xlarge-cache
+    if: |
+      (github.event_name == 'pull_request' && contains( github.event.pull_request.labels.*.name, 'run-benchmark') )||
+      (github.event_name == 'push' && github.ref == 'refs/heads/main')
     container:
       image: huggingface/transformers-pytorch-gpu
       options: --gpus all --privileged --ipc host
     steps:
       - name: Get repo
-        if: github.event_name == 'pull_request'
         uses: actions/checkout@v4
         with:
-          ref: ${{ github.event.pull_request.head.sha }}
-
-      - name: Get repo
-        if: github.event_name == 'push'
-        uses: actions/checkout@v4
-        with:
-          ref: ${{ github.sha }}
+          ref: ${{ github.event.pull_request.head.sha || github.sha }}
 
       - name: Install libpq-dev & psql
         run: |

scripts/deberta_scrtipt.py

@@ -0,0 +1,82 @@
+import torch
+from transformers import pipeline, AutoTokenizer, AutoModel, AutoModelForMaskedLM
+import time
+
+test_sentence = 'Do you [MASK] the muffin man?'
+
+# for comparison
+bert = pipeline('fill-mask', model = 'bert-base-uncased')
+print('\n'.join([d['sequence'] for d in bert(test_sentence)]))
+
+
+deberta = pipeline('fill-mask', model = 'microsoft/deberta-v3-base', model_kwargs={"legacy": False})
+print('\n'.join([d['sequence'] for d in deberta(test_sentence)]))
+
+
+tokenizer = AutoTokenizer.from_pretrained("microsoft/deberta-v3-base")
+
+tokenized_dict = tokenizer(
+    ["Is this working",], ["Not yet",],
+    return_tensors="pt"
+)
+
+deberta.model.forward = torch.compile(deberta.model.forward)
+start=time.time()
+deberta.model(**tokenized_dict)
+end=time.time()
+print(end-start)
+
+
+start=time.time()
+deberta.model(**tokenized_dict)
+end=time.time()
+print(end-start)
+
+
+start=time.time()
+deberta.model(**tokenized_dict)
+end=time.time()
+print(end-start)
+
+
+model = AutoModel.from_pretrained('microsoft/deberta-base')
+model.config.return_dict = False
+model.config.output_hidden_states=False
+input_tuple = (tokenized_dict['input_ids'], tokenized_dict['attention_mask'])
+
+
+start=time.time()
+traced_model = torch.jit.trace(model, input_tuple)
+end=time.time()
+print(end-start)
+
+
+start=time.time()
+traced_model(tokenized_dict['input_ids'], tokenized_dict['attention_mask'])
+end=time.time()
+print(end-start)
+
+
+start=time.time()
+traced_model(tokenized_dict['input_ids'], tokenized_dict['attention_mask'])
+end=time.time()
+print(end-start)
+
+
+start=time.time()
+traced_model(tokenized_dict['input_ids'], tokenized_dict['attention_mask'])
+end=time.time()
+print(end-start)
+
+
+start=time.time()
+traced_model(tokenized_dict['input_ids'], tokenized_dict['attention_mask'])
+end=time.time()
+print(end-start)
+
+
+torch.jit.save(traced_model, "compiled_deberta.pt")
+
+
+
+# my_script_module = torch.jit.script(model)