[Model] Drop the 4x M-RoPE cache headroom for Qwen3-VL, Qwen3.5 and Qwen3.8 (#59853)

Signed-off-by: lesj0610 <lesj0610@users.noreply.github.com>
Signed-off-by: lesj0610 <lesj0610@gmail.com>
Co-authored-by: lesj0610 <lesj0610@users.noreply.github.com>
This commit is contained in:
SeongJun Lee
2026-10-10 06:51:52 +00:00
committed by GitHub
co-authored by lesj0610
parent c1692c85f1
commit a98247ab4d
11 changed files with 56 additions and 6 deletions
+2
View File
@@ -280,6 +280,7 @@ class Qwen3NextAttention(nn.Module):
quant_config: QuantizationConfig | None = None,
reduce_results: bool = True,
prefix: str = "",
mrope_positions_factor: int = 4,
) -> None:
super().__init__()
self.config = config
@@ -331,6 +332,7 @@ class Qwen3NextAttention(nn.Module):
max_position=config.max_position_embeddings,
rope_parameters=config.rope_parameters,
dual_chunk_attention_config=self.dual_chunk_attention_config,
mrope_positions_factor=mrope_positions_factor,
)
# Late-interaction retrieval models (e.g. ColQwen3.5) run BIDIRECTIONAL