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, 22:35 UTC
[Tests] Fixing bug inside MultiModalProfiler. (#21842)
Signed-off-by: Varun Shenoy <varun.vinayak.shenoy@oracle.com>
Varun Vinayak Shenoy committed
Jul 30, 2025 at 00:44 UTC
547795232de307e0bff5e779530f01d7e6f4a9ad
2 files changed
+70
-1
tests/models/multimodal/processing/test_mllama4.py
new
+67
@@ -0,0 +1,67 @@
1
+# SPDX-License-Identifier: Apache-2.0
2
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
+"""Tests for mllama's multimodal preprocessing and profiling."""
4
+import pytest
5
+from torch import prod
6
+from transformers import Llama4Config
7
+
8
+from vllm.multimodal import MULTIMODAL_REGISTRY
9
+from vllm.multimodal.profiling import MultiModalProfiler
10
+
11
+from ...utils import build_model_context
12
+
13
+
14
+@pytest.mark.parametrize("model_id", ["meta-llama/Llama-Guard-4-12B"])
15
+@pytest.mark.parametrize("max_model_len", [4096, 8192, 25600, 131072])
16
+def test_profiling(model_id: str, max_model_len: int):
17
+ model_config_kwargs = {
18
+ "max_model_len": max_model_len,
19
+ }
20
+ ctx = build_model_context(
21
+ model_id,
22
+ model_config_kwargs=model_config_kwargs,
23
+ limit_mm_per_prompt={"image": 1},
24
+ )
25
+
26
+ mm_config = ctx.get_mm_config()
27
+ processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config)
28
+ profiler = MultiModalProfiler(processor)
29
+
30
+ decoder_dummy_data = profiler.get_decoder_dummy_data(
31
+ max_model_len,
32
+ mm_counts=mm_config.limit_per_prompt,
33
+ )
34
+ dummy_mm_data = processor.dummy_inputs.get_dummy_processor_inputs(
35
+ max_model_len,
36
+ mm_counts=mm_config.limit_per_prompt,
37
+ )
38
+
39
+ hf_config = ctx.get_hf_config(Llama4Config)
40
+
41
+ mm_kwargs = processor.apply(
42
+ prompt=dummy_mm_data.prompt,
43
+ mm_data=dummy_mm_data.mm_data,
44
+ hf_processor_mm_kwargs=dict(),
45
+ )["mm_kwargs"]
46
+
47
+ image_size = hf_config.vision_config.image_size
48
+ patch_size = hf_config.vision_config.patch_size
49
+ downsample_ratio = int(
50
+ round(1.0 / (hf_config.vision_config.pixel_shuffle_ratio**2)))
51
+ tokens_per_patch = ((image_size // patch_size)**2) // downsample_ratio
52
+ chunks_per_image = prod(mm_kwargs["patches_per_image"])
53
+ total_num_patches = chunks_per_image * tokens_per_patch
54
+ num_tiles = mm_kwargs["aspect_ratios"][0][0] * mm_kwargs["aspect_ratios"][
55
+ 0][1] # x-y seperator tokens
56
+ total_tokens = total_num_patches.item() + num_tiles.item(
57
+ ) + 3 # image start, image, image end
58
+
59
+ profiled_tokens = profiler.get_mm_max_contiguous_tokens(
60
+ max_model_len,
61
+ mm_counts=mm_config.limit_per_prompt,
62
+ )
63
+
64
+ assert total_tokens == profiled_tokens["image"]
65
+ assert total_tokens == sum(
66
+ placeholder.length for placeholder in
67
+ decoder_dummy_data.multi_modal_placeholders["image"])
tests/models/registry.py
+3
-1
@@ -391,7 +391,9 @@ _MULTIMODAL_EXAMPLE_MODELS = {
391
extras={"thinking": "moonshotai/Kimi-VL-A3B-Thinking"}, # noqa: E501
392
trust_remote_code=True),
393
"Llama4ForConditionalGeneration": _HfExamplesInfo("meta-llama/Llama-4-Scout-17B-16E-Instruct", # noqa: E501
394
- max_model_len=10240),
394
+ max_model_len=10240,
395
+ extras={"llama-guard-4": "meta-llama/Llama-Guard-4-12B"}, # noqa: E501
396
+ ),
397
"LlavaForConditionalGeneration": _HfExamplesInfo("llava-hf/llava-1.5-7b-hf",
398
extras={"mistral": "mistral-community/pixtral-12b", # noqa: E501
399
"mistral-fp8": "nm-testing/pixtral-12b-FP8-dynamic"}), # noqa: E501