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

[Bugfix] Fix Qwen3.5 config loading (#34554)

Browse Source 
Signed-off-by: Roger Wang <hey@rogerw.io>
This commit is contained in:
Roger Wang
2026-02-14 03:56:11 -08:00
committed by GitHub
parent 342a7cda2d
commit 2f186635cb
2 changed files with 18 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
vllm/transformers_utils/configs/qwen3_5.py
View File

@@ -72,10 +72,6 @@ class Qwen3_5TextConfig(PretrainedConfig):
"mrope_section",
"mrope_interleaved",
]
self.pad_token_id = pad_token_id
self.bos_token_id = bos_token_id
self.eos_token_id = eos_token_id
self.tie_word_embeddings = tie_word_embeddings
self.vocab_size = vocab_size
self.max_position_embeddings = max_position_embeddings
self.hidden_size = hidden_size
@@ -111,6 +107,13 @@ class Qwen3_5TextConfig(PretrainedConfig):
self.linear_num_key_heads = linear_num_key_heads
self.linear_num_value_heads = linear_num_value_heads
super().__init__(**kwargs)
# Set these AFTER super().__init__() because transformers v4's
# PretrainedConfig.__init__ has these as explicit params with different
# defaults (e.g. tie_word_embeddings=True) that would overwrite our values.
self.pad_token_id = pad_token_id
self.bos_token_id = bos_token_id
self.eos_token_id = eos_token_id
self.tie_word_embeddings = tie_word_embeddings
class Qwen3_5VisionConfig(PretrainedConfig):
@@ -182,8 +185,9 @@ class Qwen3_5Config(PretrainedConfig):
self.video_token_id = video_token_id
self.vision_start_token_id = vision_start_token_id
self.vision_end_token_id = vision_end_token_id
self.tie_word_embeddings = tie_word_embeddings
super().__init__(**kwargs)
# Set after super().__init__() to avoid v4 PretrainedConfig overwrite
self.tie_word_embeddings = tie_word_embeddings
__all__ = ["Qwen3_5Config", "Qwen3_5TextConfig"]

14
vllm/transformers_utils/configs/qwen3_5_moe.py
View File

@@ -79,10 +79,6 @@ class Qwen3_5MoeTextConfig(PretrainedConfig):
"mrope_section",
"mrope_interleaved",
]
self.pad_token_id = pad_token_id
self.bos_token_id = bos_token_id
self.eos_token_id = eos_token_id
self.tie_word_embeddings = tie_word_embeddings
self.vocab_size = vocab_size
self.max_position_embeddings = max_position_embeddings
self.hidden_size = hidden_size
@@ -123,6 +119,13 @@ class Qwen3_5MoeTextConfig(PretrainedConfig):
self.output_router_logits = output_router_logits
self.router_aux_loss_coef = router_aux_loss_coef
super().__init__(**kwargs)
# Set these AFTER super().__init__() because transformers v4's
# PretrainedConfig.__init__ has these as explicit params with different
# defaults (e.g. tie_word_embeddings=True) that would overwrite our values.
self.pad_token_id = pad_token_id
self.bos_token_id = bos_token_id
self.eos_token_id = eos_token_id
self.tie_word_embeddings = tie_word_embeddings
class Qwen3_5MoeVisionConfig(PretrainedConfig):
@@ -194,8 +197,9 @@ class Qwen3_5MoeConfig(PretrainedConfig):
self.video_token_id = video_token_id
self.vision_start_token_id = vision_start_token_id
self.vision_end_token_id = vision_end_token_id
self.tie_word_embeddings = tie_word_embeddings
super().__init__(**kwargs)
# Set after super().__init__() to avoid v4 PretrainedConfig overwrite
self.tie_word_embeddings = tie_word_embeddings
__all__ = ["Qwen3_5MoeConfig", "Qwen3_5MoeTextConfig"]