🐢 Fix slow tests (huggingface#2397)

* fix slow CI

* fix dpo

* formatting

* Apply suggestions from code review

* `setup_chat_format` may add a pad token

---------

Co-authored-by: Quentin Gallouédec <[email protected]>
Co-authored-by: Quentin Gallouédec <[email protected]>

Author: kashif

tests/slow/test_dpo_slow.py

@@ -62,6 +62,7 @@ def test_dpo_bare_model(self, model_id, loss_type, pre_compute_logits):
         """
         model = AutoModelForCausalLM.from_pretrained(model_id)
         tokenizer = AutoTokenizer.from_pretrained(model_id)
+        tokenizer.pad_token = tokenizer.eos_token if tokenizer.pad_token is None else tokenizer.pad_token
 
         with tempfile.TemporaryDirectory() as tmp_dir:
             training_args = DPOConfig(
@@ -116,6 +117,7 @@ def test_dpo_peft_model(self, model_id, loss_type, pre_compute_logits, gradient_
         """
         model = AutoModelForCausalLM.from_pretrained(model_id)
         tokenizer = AutoTokenizer.from_pretrained(model_id)
+        tokenizer.pad_token = tokenizer.eos_token if tokenizer.pad_token is None else tokenizer.pad_token
 
         with tempfile.TemporaryDirectory() as tmp_dir:
             training_args = DPOConfig(
@@ -180,6 +182,7 @@ def test_dpo_peft_model_qlora(self, model_id, loss_type, pre_compute_logits, gra
 
         model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=quantization_config)
         tokenizer = AutoTokenizer.from_pretrained(model_id)
+        tokenizer.pad_token = tokenizer.eos_token if tokenizer.pad_token is None else tokenizer.pad_token
 
         with tempfile.TemporaryDirectory() as tmp_dir:
             training_args = DPOConfig(

tests/slow/test_sft_slow.py

@@ -104,6 +104,7 @@ def test_sft_trainer_transformers(self, model_name, packing):
 
             model = AutoModelForCausalLM.from_pretrained(model_name)
             tokenizer = AutoTokenizer.from_pretrained(model_name)
+            tokenizer.pad_token = tokenizer.eos_token if tokenizer.pad_token is None else tokenizer.pad_token
 
             trainer = SFTTrainer(
                 model,
@@ -138,6 +139,7 @@ def test_sft_trainer_peft(self, model_name, packing):
 
             model = AutoModelForCausalLM.from_pretrained(model_name)
             tokenizer = AutoTokenizer.from_pretrained(model_name)
+            tokenizer.pad_token = tokenizer.eos_token if tokenizer.pad_token is None else tokenizer.pad_token
 
             trainer = SFTTrainer(
                 model,
@@ -174,6 +176,7 @@ def test_sft_trainer_transformers_mp(self, model_name, packing):
 
             model = AutoModelForCausalLM.from_pretrained(model_name)
             tokenizer = AutoTokenizer.from_pretrained(model_name)
+            tokenizer.pad_token = tokenizer.eos_token if tokenizer.pad_token is None else tokenizer.pad_token
 
             trainer = SFTTrainer(
                 model,
@@ -209,6 +212,7 @@ def test_sft_trainer_transformers_mp_gc(self, model_name, packing, gradient_chec
 
             model = AutoModelForCausalLM.from_pretrained(model_name)
             tokenizer = AutoTokenizer.from_pretrained(model_name)
+            tokenizer.pad_token = tokenizer.eos_token if tokenizer.pad_token is None else tokenizer.pad_token
 
             trainer = SFTTrainer(
                 model,
@@ -245,6 +249,7 @@ def test_sft_trainer_transformers_mp_gc_peft(self, model_name, packing, gradient
 
             model = AutoModelForCausalLM.from_pretrained(model_name)
             tokenizer = AutoTokenizer.from_pretrained(model_name)
+            tokenizer.pad_token = tokenizer.eos_token if tokenizer.pad_token is None else tokenizer.pad_token
 
             trainer = SFTTrainer(
                 model,
@@ -288,6 +293,7 @@ def test_sft_trainer_transformers_mp_gc_device_map(
 
             model = AutoModelForCausalLM.from_pretrained(model_name, device_map=device_map)
             tokenizer = AutoTokenizer.from_pretrained(model_name)
+            tokenizer.pad_token = tokenizer.eos_token if tokenizer.pad_token is None else tokenizer.pad_token
 
             trainer = SFTTrainer(
                 model,
@@ -327,6 +333,7 @@ def test_sft_trainer_transformers_mp_gc_peft_qlora(self, model_name, packing, gr
 
             model = AutoModelForCausalLM.from_pretrained(model_name, quantization_config=quantization_config)
             tokenizer = AutoTokenizer.from_pretrained(model_name)
+            tokenizer.pad_token = tokenizer.eos_token if tokenizer.pad_token is None else tokenizer.pad_token
 
             trainer = SFTTrainer(
                 model,
@@ -370,7 +377,9 @@ def test_sft_trainer_with_chat_format_qlora(self, model_name, packing):
             model = AutoModelForCausalLM.from_pretrained(model_name, quantization_config=quantization_config)
             tokenizer = AutoTokenizer.from_pretrained(model_name)
 
-            model, tokenizer = setup_chat_format(model, tokenizer)
+            if tokenizer.chat_template is None:
+                model, tokenizer = setup_chat_format(model, tokenizer)
+            tokenizer.pad_token = tokenizer.eos_token if tokenizer.pad_token is None else tokenizer.pad_token
 
             trainer = SFTTrainer(
                 model,