Skip to content

Add Nemotron 3 to tests via tiny model #5278

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
sergiopaniego wants to merge 9 commits into
base: main
Choose a base branch
from
+60 −1
Open

Add Nemotron 3 to tests via tiny model #5278

Show file tree
Hide file tree
Changes from 5 commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
dd4f3a6
Add Nemotron 3 to tests via tiny model
sergiopaniego Mar 12, 2026
529ef04
Code quality
sergiopaniego Mar 12, 2026
875a46c
Merge branch 'main' into nemotron3-tiny-tests
sergiopaniego Mar 12, 2026
b79861f
Updated
sergiopaniego Mar 12, 2026
27cef67
Merge branch 'nemotron3-tiny-tests' of github.com:huggingface/trl int…
sergiopaniego Mar 12, 2026
4f9b1f8
Update
sergiopaniego Mar 12, 2026
8122d6d
Merge branch 'main' into nemotron3-tiny-tests
sergiopaniego Mar 13, 2026
c555040
Update
sergiopaniego Mar 13, 2026
3c8f9d4
Cursor review
sergiopaniego Mar 13, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
28 changes: 28 additions & 0 deletions scripts/generate_tiny_models.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -59,6 +59,8 @@
LlavaNextForConditionalGeneration,
MistralConfig,
MistralForCausalLM,
NemotronHConfig,
NemotronHForCausalLM,
OPTConfig,
OPTForCausalLM,
PaliGemmaForConditionalGeneration,
Expand Down Expand Up @@ -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=4,
mamba_head_dim=8,
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)
Copy link
Member

@qgallouedec qgallouedec Mar 13, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can you cast backbone.layers.[N].mixer.D and backbone.layers.[N].mixer.A_log to fp32?: it seems like these two layers are in fp32, and we want to be as close as possible to the reference model

check how we do here for Qwen3.5 https://github.com/huggingface/trl/pull/5278/changes#diff-dd3349f840a26de373fc88378e6fcded0b75423da8a34f7cfa6ac573b7398b8bL404

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")
Expand Down
11 changes: 11 additions & 0 deletions tests/test_dpo_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -139,17 +139,28 @@ 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 does not support gradient checkpointing
gradient_checkpointing = "NemotronH" not in model_id

# 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",
gradient_checkpointing=gradient_checkpointing,
)
trainer = DPOTrainer(model=model_id, args=training_args, train_dataset=dataset)

Expand Down
14 changes: 13 additions & 1 deletion tests/test_sft_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -284,14 +284,26 @@ 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 does not support gradient checkpointing
gradient_checkpointing = "NemotronH" not in model_id

# 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", gradient_checkpointing=gradient_checkpointing
)
trainer = SFTTrainer(model=model_id, args=training_args, train_dataset=dataset)

# Save the initial parameters to compare them later
Expand Down
Loading