[Bugfix] Skip Qwen3 deepstack buffers without vision (#49397)

Signed-off-by: WEI CHENG CHIU <waynehacking8@gmail.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
Wei-Cheng (Wayne) Chiu
2026-08-05 12:46:23 +08:00
committed by GitHub
co-authored by OpenAI Codex
parent eb3dce9757
commit 7794b1e08b
3 changed files with 66 additions and 63 deletions
+23 -21
View File
@@ -142,6 +142,7 @@ from .qwen3 import Qwen3ForCausalLM, Qwen3Model
from .utils import (
AutoWeightsLoader,
PPMissingLayer,
StageMissingLayer,
WeightsMapper,
_merge_multimodal_embeddings,
maybe_prefix,
@@ -1769,15 +1770,6 @@ class Qwen3VLForConditionalGeneration(
self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data"
self._init_video_pruning(multimodal_config)
self.use_deepstack = hasattr(config.vision_config, "deepstack_visual_indexes")
self.deepstack_num_level = (
len(config.vision_config.deepstack_visual_indexes)
if self.use_deepstack
else 0
)
self.visual_dim = config.vision_config.out_hidden_size
self.multiscale_dim = self.visual_dim * self.deepstack_num_level
with self._mark_tower_model(vllm_config, {"image", "video"}):
self.visual = Qwen3_VisionTransformer(
config.vision_config,
@@ -1786,18 +1778,28 @@ class Qwen3VLForConditionalGeneration(
prefix=maybe_prefix(prefix, "visual"),
)
# register buffer for deepstack
if self.use_deepstack:
self.deepstack_input_embeds = [
torch.zeros(
vllm_config.scheduler_config.max_num_batched_tokens,
config.text_config.hidden_size,
)
for _ in range(self.deepstack_num_level)
]
# Tracks the valid token span currently stored in the buffer.
# Zero means there is no active deepstack payload to consume.
self.deepstack_input_embeds_num_tokens = 0
self.use_deepstack = hasattr(
config.vision_config, "deepstack_visual_indexes"
) and not isinstance(self.visual, StageMissingLayer)
self.deepstack_num_level = (
len(config.vision_config.deepstack_visual_indexes)
if self.use_deepstack
else 0
)
self.visual_dim = config.vision_config.out_hidden_size
self.multiscale_dim = self.visual_dim * self.deepstack_num_level
if self.use_deepstack:
self.deepstack_input_embeds = [
torch.zeros(
vllm_config.scheduler_config.max_num_batched_tokens,
config.text_config.hidden_size,
)
for _ in range(self.deepstack_num_level)
]
# Tracks the valid token span currently stored in the buffer.
# Zero means there is no active deepstack payload to consume.
self.deepstack_input_embeds_num_tokens = 0
with self._mark_language_model(vllm_config):
self.language_model = Qwen3LLMForCausalLM(