biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Refactor] Define MultiModalKwargsItems separate from MultiModalKwargs (#23053)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-08-18 17:52:00 +08:00
committed by GitHub
parent 5c79b0d648
commit 27e8d1ea3e
77 changed files with 431 additions and 383 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
tests/models/multimodal/processing/test_mllama.py
View File

@@ -49,18 +49,18 @@ def test_profiling(
encoder_seq_lens = [len(dummy_encoder_data.prompt_token_ids)
] * max_num_seqs
mm_kwargs = processor.apply(
mm_data = processor.apply(
prompt=dummy_mm_data.prompt,
mm_data=dummy_mm_data.mm_data,
hf_processor_mm_kwargs=dict(),
)["mm_kwargs"]
)["mm_kwargs"].get_data()
# Get the actual number of encoder tokens for each sample.
# Because attn_metadata.encoder_seq_lens only counts the last
# group of images for each sample, which is used to cheat the
# block manager to allocate blocks for those images only.
# See MllamaMultiModalProcessor for more details.
num_tiles = [[t] for t in mm_kwargs.pop("num_tiles")]
num_tiles = [[t] for t in mm_data.pop("num_tiles")]
num_tokens_per_tile = calc_token_per_chunk(image_size)
actual_encoder_seq_lens = [
sum(num_tile) * num_tokens_per_tile for num_tile in num_tiles