[MM][CG] Support --enable-vit-cuda-graph option for VLM examples (#40580)
Signed-off-by: shen-shanshan <467638484@qq.com>
This commit is contained in:
@@ -1802,7 +1802,11 @@ class Qwen3VLForConditionalGeneration(
|
||||
# spatial_merge_size=2 → 8x8 = 64 tokens
|
||||
min_budget = 64
|
||||
# Max: capped by max_num_batched_tokens
|
||||
max_budget = vllm_config.scheduler_config.max_num_batched_tokens
|
||||
# TODO(shen-shanshan): the max_budget auto-infer needs to be optimized later.
|
||||
max_budget = min(
|
||||
vllm_config.scheduler_config.max_num_batched_tokens,
|
||||
self.model_config.max_model_len,
|
||||
)
|
||||
return (min_budget, max_budget)
|
||||
|
||||
def _get_pixel_values_by_modality(
|
||||
|
||||
Reference in New Issue
Block a user