A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 4, 15:00 UTC
[Bugfix] Remove invalid deepstack boundary check for Qwen3-VL (#40932)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Isotr0py committed
Apr 27, 2026 at 15:27 UTC
22631f80a01a04b06398952e77d7890ab660ab10
2 files changed
-22
vllm/model_executor/models/qwen3_omni_moe_thinker.py
-11
@@ -1778,11 +1778,6 @@ class Qwen3OmniMoeThinkerForConditionalGeneration(
1778
return None # If vision tower is skipped
1779
if getattr(self, "deepstack_input_embeds_num_tokens", 0) == 0:
1780
return None
1781
- if num_tokens > self.deepstack_input_embeds_num_tokens:
1782
- raise ValueError(
1783
- "Requested more deepstack tokens than available in buffer: "
1784
- f"{num_tokens=} > {self.deepstack_input_embeds_num_tokens=}"
1785
- )
1781
1782
# get deepstack_input_embeds from buffer, and clear the buffer
1783
return IntermediateTensors(
@@ -1824,12 +1819,6 @@ class Qwen3OmniMoeThinkerForConditionalGeneration(
1819
1820
# clear deepstack_input_embeds in buffer
1821
if num_tokens > 0:
1827
- if num_tokens > self.deepstack_input_embeds_num_tokens:
1828
- raise ValueError(
1829
- "Requested to clear more deepstack tokens than available in "
1830
- "buffer: "
1831
- f"{num_tokens=} > {self.deepstack_input_embeds_num_tokens=}"
1832
- )
1822
for idx in range(self.deepstack_num_level):
1823
self.deepstack_input_embeds[idx][:num_tokens].zero_()
1824
self.deepstack_input_embeds_num_tokens = 0
vllm/model_executor/models/qwen3_vl.py
-11
@@ -1716,11 +1716,6 @@ class Qwen3VLForConditionalGeneration(
1716
return None # If vision tower is skipped
1717
if getattr(self, "deepstack_input_embeds_num_tokens", 0) == 0:
1718
return None
1719
- if num_tokens > self.deepstack_input_embeds_num_tokens:
1720
- raise ValueError(
1721
- "Requested more deepstack tokens than available in buffer: "
1722
- f"{num_tokens=} > {self.deepstack_input_embeds_num_tokens=}"
1723
- )
1719
1720
# get deepstack_input_embeds from buffer, and clear the buffer
1721
return IntermediateTensors(
@@ -1762,12 +1757,6 @@ class Qwen3VLForConditionalGeneration(
1757
1758
# clear deepstack_input_embeds in buffer
1759
if num_tokens > 0:
1765
- if num_tokens > self.deepstack_input_embeds_num_tokens:
1766
- raise ValueError(
1767
- "Requested to clear more deepstack tokens than available in "
1768
- "buffer: "
1769
- f"{num_tokens=} > {self.deepstack_input_embeds_num_tokens=}"
1770
- )
1760
for idx in range(self.deepstack_num_level):
1761
self.deepstack_input_embeds[idx][:num_tokens].zero_()
1762
self.deepstack_input_embeds_num_tokens = 0