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 14, 05:30 UTC
[Model] Allow passing custom number of max tiles to Nano 2 VL (#26403)
Signed-off-by: Eugene Khvedchenia <ekhvedchenia@nvidia.com>
Eugene Khvedchenya committed
Oct 8, 2025 at 14:19 UTC
f9582fd8f4f0e03a1b0d5b959bbf15d297336775
1 file changed
+19
-6
vllm/model_executor/models/nano_nemotron_vl.py
+19
-6
@@ -93,6 +93,7 @@ IMG_CONTEXT = "<image>"
93
94
# Profiling
95
MAX_FRAMES = 16
96
+DEFAULT_NUM_TILES = 12
97
98
99
class NanoNemotronVLImagePixelInputs(TypedDict):
@@ -227,6 +228,8 @@ def video_to_pixel_values(
228
max_num_tiles: int = 1,
229
use_thumbnail: bool,
230
) -> torch.Tensor:
231
+ assert max_num_tiles == 1, "Video modality always uses one tile"
232
+
233
# Convert each frame to a single resized tile tensor consistent
234
# with image path
235
frames_tensors: list[torch.Tensor] = []
@@ -255,13 +258,19 @@ class BaseNanoNemotronVLProcessor(ABC):
258
"""
259
260
def __init__(
258
- self, config: PretrainedConfig, tokenizer: AnyTokenizer, *args, **kwargs
261
+ self,
262
+ config: PretrainedConfig,
263
+ tokenizer: AnyTokenizer,
264
+ *args,
265
+ max_num_tiles: Optional[int] = None,
266
+ **kwargs,
267
) -> None:
268
super().__init__()
269
270
self.config = config
271
self.tokenizer = tokenizer
272
273
+ self.max_num_tiles = max_num_tiles or DEFAULT_NUM_TILES
274
image_size: int = config.force_image_size
275
patch_size: int = config.patch_size
276
@@ -361,7 +370,7 @@ class BaseNanoNemotronVLProcessor(ABC):
370
) -> BatchFeature:
371
# Use default if not provided
372
if max_num_tiles is None:
364
- max_num_tiles = 12
373
+ max_num_tiles = self.max_num_tiles
374
375
text, images = [self._make_batch_input(x) for x in (text, images)]
376
@@ -390,6 +399,7 @@ class NanoNemotronVLProcessor(BaseNanoNemotronVLProcessor):
399
config: PretrainedConfig,
400
tokenizer: AnyTokenizer,
401
*,
402
+ max_num_tiles: Optional[int] = None,
403
min_dynamic_patch: Optional[int] = None,
404
max_dynamic_patch: Optional[int] = None,
405
dynamic_image_size: Optional[bool] = None,
@@ -399,6 +409,7 @@ class NanoNemotronVLProcessor(BaseNanoNemotronVLProcessor):
409
super().__init__(
410
config=config,
411
tokenizer=tokenizer,
412
+ max_num_tiles=max_num_tiles,
413
min_dynamic_patch=min_dynamic_patch,
414
max_dynamic_patch=max_dynamic_patch,
415
dynamic_image_size=dynamic_image_size,
@@ -506,7 +517,7 @@ class NanoNemotronVLProcessor(BaseNanoNemotronVLProcessor):
517
) -> BatchFeature:
518
# Use default if not provided
519
if max_num_tiles is None:
509
- max_num_tiles = 12
520
+ max_num_tiles = self.max_num_tiles
521
522
text, images, videos = [
523
self._make_batch_input(x) for x in (text, images, videos)
@@ -521,7 +532,7 @@ class NanoNemotronVLProcessor(BaseNanoNemotronVLProcessor):
532
text, video_inputs = self._preprocess_video(
533
text=text,
534
videos=videos,
524
- max_num_tiles=max_num_tiles,
535
+ max_num_tiles=1,
536
dynamic_image_size=dynamic_image_size,
537
)
538
@@ -635,7 +646,7 @@ class BaseNanoNemotronVLProcessingInfo(BaseProcessingInfo):
646
def get_max_image_tokens(self) -> int:
647
processor = self.get_hf_processor()
648
# Use default max_num_tiles for max tokens calculation
638
- max_num_tiles = 12
649
+ max_num_tiles = processor.max_num_tiles
650
target_width, target_height = self.get_image_size_with_most_features(
651
max_num_tiles
652
)
@@ -768,7 +779,9 @@ class NanoNemotronBaseVLMultiModalProcessor(BaseMultiModalProcessor[_I]):
779
else:
780
image_size = images.get_image_size(item_idx)
781
# Extract max_num_tiles from kwargs, default to 12
771
- max_num_tiles = hf_processor_mm_kwargs.get("max_num_tiles", 12)
782
+ max_num_tiles = hf_processor_mm_kwargs.get(
783
+ "max_num_tiles", hf_processor.max_num_tiles
784
+ )
785
feature_size = self.info.get_num_image_tokens(
786
image_width=image_size.width,
787
image_height=image_size.height,