[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
@@ -91,6 +91,7 @@ from .qwen3_vl import (
Qwen3VLForConditionalGeneration,
Qwen3VLMultiModalProcessor,
Qwen3VLProcessingInfo,
mrope_positions_factor,
)
from .utils import (
AutoWeightsLoader,
@@ -161,6 +162,7 @@ class Qwen3_5DecoderLayer(Qwen3NextDecoderLayer):
quant_config=quant_config,
prefix=f"{prefix}.self_attn",
reduce_results=not self.use_attn_reduce_scatter_for_moe,
mrope_positions_factor=mrope_positions_factor(vllm_config),
)
else:
raise ValueError(f"Invalid layer_type {self.layer_type}")