biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Model] Add support for YARN in NemotronNAS models (#18427)

Browse Source 
Signed-off-by: Nave Assaf <nassaf@nvidia.com>
This commit is contained in:
Naveassaf
2025-05-26 13:31:49 +03:00
committed by GitHub
parent 5a2c76cbe1
commit 6d68030f1c
2 changed files with 67 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
vllm/model_executor/models/llama.py
View File

@@ -162,20 +162,9 @@ class LlamaAttention(nn.Module):
prefix=f"{prefix}.o_proj",
)
is_neox_style = True
is_gguf = quant_config and quant_config.get_name() == "gguf"
if is_gguf and config.model_type == "llama":
is_neox_style = False
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position_embeddings,
base=rope_theta,
rope_scaling=rope_scaling,
is_neox_style=is_neox_style,
partial_rotary_factor=self.partial_rotary_factor,
)
self._init_rotary_emb(config,
rope_scaling=rope_scaling,
quant_config=quant_config)
if hasattr(config, "interleaved_sliding_window"):
interleaved_sliding_window = config.interleaved_sliding_window
@@ -214,6 +203,24 @@ class LlamaAttention(nn.Module):
output, _ = self.o_proj(attn_output)
return output
def _init_rotary_emb(self, config: LlamaConfig,
rope_scaling: Optional[dict[str, Any]],
quant_config: Optional[QuantizationConfig]) -> None:
is_neox_style = True
is_gguf = quant_config and quant_config.get_name() == "gguf"
if is_gguf and self.config.model_type == "llama":
is_neox_style = False
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=self.max_position_embeddings,
base=self.rope_theta,
rope_scaling=rope_scaling,
is_neox_style=is_neox_style,
partial_rotary_factor=self.partial_rotary_factor,
)
class LlamaDecoderLayer(nn.Module):