Skip to content

feat: Add support for YARN in NemotronNAS models #4906

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

Merged
merged 5 commits into from
Jun 29, 2025
Merged

feat: Add support for YARN in NemotronNAS models #4906

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
4b5e19b
feat: Add support for YARN in NemotronNAS models
amirkl94 Jun 4, 2025
c141457
Fix issues from tests
amirkl94 Jun 19, 2025
0b76688
Revert recursive weight loading
amirkl94 Jun 24, 2025
b223322
Remove wrong not
amirkl94 Jun 24, 2025
541b22b
Remove neox all together
amirkl94 Jun 26, 2025
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
3 changes: 1 addition & 2 deletions tensorrt_llm/_torch/attention_backend/interface.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -392,8 +392,7 @@ def create_rope_const_params(self, interleave: bool = True):
)

if self.scale_type == RotaryScalingType.yarn:
rope_inv_freq = None
_, rope_cos_sin = RopeEmbeddingUtils.create_sinusoidal_positions_yarn(
rope_inv_freq, rope_cos_sin = RopeEmbeddingUtils.create_sinusoidal_positions_yarn(
self.max_positions,
self.dim,
self.theta,
Expand Down
2 changes: 1 addition & 1 deletion tensorrt_llm/_torch/attention_backend/trtllm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -110,7 +110,7 @@ def __init__(
self.qk_rope_head_dim = None
self.v_head_dim = None

self.rotary_inv_freq, self.rotary_cos_sin = rope_params.create_rope_const_params(
self.rotary_inv_freq, self.rotary_cos_sin = self.rope_params.create_rope_const_params(
)

self.num_heads = num_heads
Expand Down
15 changes: 12 additions & 3 deletions tensorrt_llm/_torch/models/modeling_nemotron_nas.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,7 @@
from torch import nn
from transformers import PretrainedConfig

from tensorrt_llm.functional import PositionEmbeddingType
from tensorrt_llm.functional import PositionEmbeddingType, RotaryScalingType
from tensorrt_llm.lora_manager import HfLoraLoader
from tensorrt_llm.models.convert_utils import split_matrix_tp

Expand Down Expand Up @@ -48,19 +48,28 @@ def _create_linear_from_configs(model_config: ModelConfig[PretrainedConfig],


class NemotronNASAttention(Attention):
NON_NEOX_TYPES = ("mistral_yarn", "rope_llama4")

def __init__(self, model_config: ModelConfig[PretrainedConfig],
layer_idx: int):
config = model_config.pretrained_config
is_neox = getattr(model_config.pretrained_config,
"position_embedding_type",
None) not in self.NON_NEOX_TYPES
rope = RopeParams.from_config(config)
if rope.scale_type == RotaryScalingType.yarn:
rope.mscale_all_dim = 0.0

super().__init__(
hidden_size=config.hidden_size,
num_attention_heads=config.num_attention_heads,
num_key_value_heads=config.num_key_value_heads[layer_idx],
max_position_embeddings=config.max_position_embeddings,
bias=False,
pos_embd_params=PositionalEmbeddingParams(
type=PositionEmbeddingType.rope_gpt_neox,
rope=RopeParams.from_config(config),
type=PositionEmbeddingType.rope_gpt_neox
if is_neox else PositionEmbeddingType.rope_gptj,
rope=rope,
),
layer_idx=layer_idx,
dtype=config.torch_dtype,
Expand Down