[Glm46v] Bug fix for accuracy drop and unable to launch server (#14585)

Co-authored-by: yhyang201 <yhyang201@gmail.com>
Co-authored-by: zRzRzRzRzRzRzR <2448370773@qq.com>
Co-authored-by: Minglei Zhu <mingleizhu1122@gmail.com>
This commit is contained in:
Binyao Jiang
2025-12-07 23:45:02 -08:00
committed by GitHub
co-authored by yhyang201 zRzRzRzRzRzRzR Minglei Zhu
parent a2ca9bd4f1
commit cf0478d602
12 changed files with 308 additions and 29 deletions
-4
View File
@@ -1,6 +1,5 @@
from transformers import PretrainedConfig
from transformers.configuration_utils import layer_type_validation
from transformers.modeling_rope_utils import rope_config_validation
from sglang.utils import logger
@@ -168,7 +167,6 @@ class Qwen3OmniMoeTextConfig(PretrainedConfig):
# BC: if there is a 'type' field, move it to 'rope_type'.
if self.rope_scaling is not None and "type" in self.rope_scaling:
self.rope_scaling["rope_type"] = self.rope_scaling["type"]
rope_config_validation(self)
# MoE arguments
self.decoder_sparse_step = decoder_sparse_step
@@ -311,7 +309,6 @@ class Qwen3OmniMoeTalkerCodePredictorConfig(PretrainedConfig):
# BC: if there is a 'type' field, move it to 'rope_type'.
if self.rope_scaling is not None and "type" in self.rope_scaling:
self.rope_scaling["rope_type"] = self.rope_scaling["type"]
rope_config_validation(self)
self.layer_types = layer_types
if self.layer_types is None:
@@ -405,7 +402,6 @@ class Qwen3OmniMoeTalkerTextConfig(PretrainedConfig):
# BC: if there is a 'type' field, move it to 'rope_type'.
if self.rope_scaling is not None and "type" in self.rope_scaling:
self.rope_scaling["rope_type"] = self.rope_scaling["type"]
rope_config_validation(self)
# MoE arguments
self.decoder_sparse_step = decoder_sparse_step
-5
View File
@@ -1,5 +1,4 @@
from transformers import PretrainedConfig
from transformers.modeling_rope_utils import rope_config_validation
class Qwen3VLVisionConfig(PretrainedConfig):
@@ -187,8 +186,6 @@ class Qwen3VLTextConfig(PretrainedConfig):
self.attention_bias = attention_bias
self.attention_dropout = attention_dropout
rope_config_validation(self, ignore_keys={"mrope_section", "mrope_interleaved"})
super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
@@ -450,8 +447,6 @@ class Qwen3VLMoeTextConfig(PretrainedConfig):
self.rope_scaling = rope_scaling
self.head_dim = head_dim or hidden_size // num_attention_heads
rope_config_validation(self, ignore_keys={"mrope_section", "mrope_interleaved"})
# MoE arguments
self.decoder_sparse_step = decoder_sparse_step
self.moe_intermediate_size = moe_intermediate_size