upgrade transformers to 5.2.0 and torchao to 0.16.0 (#3407)

* upgrade transformers to 5.1.0 and torchao to 0.16.0

* upgrade trl for parity

* handle trl api changes

* orpo doesn't have max_prompt_len to check anymore

* cpoconfig doesn't take max_prompt_length and fix cpu offload

* slow fsdp1 test

* triton min 3.4.0 and liger to 0.7.0

* use transformers main for now for zero3 fix

* handle group_by_length change

* fix changes upstream

* mark skip flaky test

* use transformers latest release 5.2.0

tests/core/test_builders.py

@@ -300,7 +300,6 @@ class TestHFRLTrainerBuilder:
         self._test_common_training_arguments(training_arguments, rl=orpo_cfg.rl)
         # ORPO specific
         assert training_arguments.beta == 0.1  # maps from orpo_alpha
-        assert training_arguments.max_prompt_length == 512
 
     def test_kto_training_arguments(self, kto_cfg, model, tokenizer):
         builder = HFRLTrainerBuilder(kto_cfg, model, tokenizer)

tests/e2e/multigpu/test_fsdp1.py

@@ -186,6 +186,7 @@ class TestFSDP1:
 
         verify_training_success(temp_dir)
 
+    @pytest.mark.skip(reason="slow test, deprecate fsdp1 asap")
     def test_dpo_fft(self, temp_dir):
         cfg = DictDefault(
             {

tests/e2e/multigpu/test_fsdp2.py

@@ -365,6 +365,7 @@ class TestFSDP2:
 
         verify_training_success(temp_dir)
 
+    @pytest.mark.skip(reason="slow test w cu129 + torch 2.9.1 + py3.12")
     @require_torch_2_7_0
     def test_dpo_fft(self, temp_dir):
         cfg = DictDefault(