Add Nemotron 3 to tests via tiny model

scripts/generate_tiny_models.py

@@ -59,6 +59,8 @@
     LlavaNextForConditionalGeneration,
     MistralConfig,
     MistralForCausalLM,
+    NemotronHConfig,
+    NemotronHForCausalLM,
     OPTConfig,
     OPTForCausalLM,
     PaliGemmaForConditionalGeneration,
@@ -227,6 +229,32 @@ def init_weights_tiny_model(model):
     init_weights_tiny_model(model)
     push_to_hub(model, tokenizer, generation_config, "tiny", suffix)
 
+# Hybrid Mamba-Attention models
+tokenizer = AutoTokenizer.from_pretrained("nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16")
+generation_config = GenerationConfig.from_pretrained("nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16")
+config = NemotronHConfig(
+    vocab_size=len(tokenizer.vocab),
+    hidden_size=16,
+    num_attention_heads=4,
+    num_key_value_heads=2,
+    intermediate_size=32,
+    layers_block_type=["mamba", "attention"],  # 2 layers: one Mamba + one Attention
+    mamba_num_heads=8,
+    mamba_head_dim=4,
+    mamba_n_groups=1,
+    ssm_state_size=16,
+    mamba_d_conv=4,
+    mamba_expand=2,
+    n_routed_experts=4,
+    num_experts_per_tok=2,
+    moe_intermediate_size=32,
+    moe_shared_expert_intermediate_size=32,
+    use_mamba_kernels=False,  # CPU-friendly for testing
+)
+model = NemotronHForCausalLM(config).to(dtype=torch.bfloat16)
+init_weights_tiny_model(model)
+push_to_hub(model, tokenizer, generation_config, "tiny")
+
 # Two slightly bigger models, required for vLLM testing
 tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-32B-Instruct")
 generation_config = GenerationConfig.from_pretrained("Qwen/Qwen2.5-32B-Instruct")

tests/test_dpo_trainer.py

@@ -171,17 +171,33 @@ class TestDPOTrainer(TrlTestCase):
             "trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
             "trl-internal-testing/tiny-Qwen3MoeForCausalLM",
             "trl-internal-testing/tiny-GptOssForCausalLM",
+            pytest.param(
+                "trl-internal-testing/tiny-NemotronHForCausalLM",
+                marks=pytest.mark.skipif(
+                    Version(transformers.__version__) < Version("5.3.0"),
+                    reason="NemotronH models were introduced in transformers-5.3.0",
+                ),
+            ),
         ],
     )
     def test_train(self, model_id):
         # Get the dataset
         dataset = load_dataset("trl-internal-testing/zen", "standard_preference", split="train")
 
+        # NemotronH (hybrid Mamba-Attention) does not support gradient checkpointing. The Mamba CUDA
+        # kernels require strides to be multiples of 8, which is incompatible with tiny model dimensions.
+        # Force CPU so that the model uses the pure PyTorch path (works fine on GPU without kernels).
+        kwargs = {}
+        if "NemotronH" in model_id:
+            kwargs["gradient_checkpointing"] = False
+            kwargs["use_cpu"] = True
+
         # Initialize the trainer
         training_args = DPOConfig(
             output_dir=self.tmp_dir,
             learning_rate=0.1,  # use higher lr because gradients are tiny and default lr can stall updates
             report_to="none",
+            **kwargs,
         )
         trainer = DPOTrainer(model=model_id, args=training_args, train_dataset=dataset)
 

tests/test_sft_trainer.py

@@ -284,14 +284,29 @@ def test_init_with_training_arguments(self):
             "trl-internal-testing/tiny-GptOssForCausalLM",
             "trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
             "trl-internal-testing/tiny-Qwen3MoeForCausalLM",
+            pytest.param(
+                "trl-internal-testing/tiny-NemotronHForCausalLM",
+                marks=pytest.mark.skipif(
+                    Version(transformers.__version__) < Version("5.3.0"),
+                    reason="NemotronH models were introduced in transformers-5.3.0",
+                ),
+            ),
         ],
     )
     def test_train(self, model_id):
         # Get the dataset
         dataset = load_dataset("trl-internal-testing/zen", "standard_language_modeling", split="train")
 
+        # NemotronH (hybrid Mamba-Attention) does not support gradient checkpointing. The Mamba CUDA
+        # kernels require strides to be multiples of 8, which is incompatible with tiny model dimensions.
+        # Force CPU so that the model uses the pure PyTorch path (works fine on GPU without kernels).
+        kwargs = {}
+        if "NemotronH" in model_id:
+            kwargs["gradient_checkpointing"] = False
+            kwargs["use_cpu"] = True
+
         # Initialize the trainer
-        training_args = SFTConfig(output_dir=self.tmp_dir, report_to="none")
+        training_args = SFTConfig(output_dir=self.tmp_dir, report_to="none", **kwargs)
         trainer = SFTTrainer(model=model_id, args=training_args, train_dataset=dataset)
 
         # Save the initial parameters to compare them later