[Bugfix][Multimodal] Keep Gemma 4 video frame counts on CPU (#52441)
Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
This commit is contained in:
@@ -286,6 +286,12 @@ MODEL_CONFIGS: dict[str, VitCudagraphTestConfig] = {
|
|||||||
"<bos><start_of_turn>user\n<|video|>\nDescribe this video in one sentence."
|
"<bos><start_of_turn>user\n<|video|>\nDescribe this video in one sentence."
|
||||||
"<end_of_turn>\n<start_of_turn>model\n"
|
"<end_of_turn>\n<start_of_turn>model\n"
|
||||||
),
|
),
|
||||||
|
# The 16-frame test video produces 1056 vision tokens. Capture only
|
||||||
|
# the smallest supported bucket that covers it instead of all default
|
||||||
|
# buckets through max_model_len, which adds unrelated memory pressure.
|
||||||
|
compilation_config_overrides={
|
||||||
|
"encoder_cudagraph_token_budgets": [1120],
|
||||||
|
},
|
||||||
needs_video_metadata=True,
|
needs_video_metadata=True,
|
||||||
marks=[pytest.mark.core_model],
|
marks=[pytest.mark.core_model],
|
||||||
),
|
),
|
||||||
|
|||||||
@@ -801,7 +801,7 @@ class Gemma4MultiModalProcessor(BaseMultiModalProcessor[Gemma4ProcessingInfo]):
|
|||||||
MultiModalFieldConfig.flat_from_sizes("video", vfc)
|
MultiModalFieldConfig.flat_from_sizes("video", vfc)
|
||||||
),
|
),
|
||||||
video_frame_counts=MultiModalFieldConfig.batched(
|
video_frame_counts=MultiModalFieldConfig.batched(
|
||||||
"video",
|
"video", keep_on_cpu=True
|
||||||
),
|
),
|
||||||
video_num_soft_tokens=MultiModalFieldConfig.batched(
|
video_num_soft_tokens=MultiModalFieldConfig.batched(
|
||||||
"video", keep_on_cpu=True
|
"video", keep_on_cpu=True
|
||||||
|
|||||||
Reference in New Issue
Block a user