[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
@@ -23,6 +23,7 @@ from vllm.model_executor.layers.linear import QKVParallelLinear, RowParallelLine
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.models.qwen3_next import Qwen3NextAttention
from vllm.model_executor.models.qwen3_vl import mrope_positions_factor
from vllm.platforms import current_platform
from vllm.utils.torch_utils import (
LayerNameType,
@@ -248,6 +249,7 @@ class Qwen4ExpQSAAttention(Qwen3NextAttention, AttentionLayerBase):
head_size=self.head_dim,
max_position=config.max_position_embeddings,
rope_parameters=config.rope_parameters,
mrope_positions_factor=mrope_positions_factor(vllm_config),
)
self.q_norm = GemmaRMSNorm(self.head_dim, eps=config.rms_norm_eps)
self.k_norm = GemmaRMSNorm(self.head_dim, eps=config.rms_norm_eps)
+2
View File
@@ -29,6 +29,7 @@ from vllm.model_executor.layers.linear import (
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding, get_rope
from vllm.model_executor.models.qwen3_next import Qwen3NextAttention
from vllm.model_executor.models.qwen3_vl import mrope_positions_factor
from vllm.model_executor.models.utils import (
AutoWeightsLoader,
WeightsMapper,
@@ -427,6 +428,7 @@ class Qwen4ExpQSAAttention(Qwen3NextAttention, AttentionLayerBase):
head_size=self.head_dim,
max_position=config.max_position_embeddings,
rope_parameters=config.rope_parameters,
mrope_positions_factor=mrope_positions_factor(vllm_config),
)
self.q_norm = GemmaRMSNorm(self.head_dim, eps=config.rms_norm_eps)
self.k_norm = GemmaRMSNorm(self.head_dim, eps=config.rms_norm_eps)