Files
sglang/test/registered/vlm/test_vlm_input_format.py
2026-03-23 00:18:45 -07:00

694 lines
26 KiB
Python

import json
import unittest
from io import BytesIO
from typing import Optional
import requests
import torch
# Compatibility shim: Kimi-VL dynamic module expects PytorchGELUTanh which may
# be missing in transformers==4.57.1. Inject a lightweight implementation so
# the model can import successfully without downgrading transformers.
import transformers.activations as _hf_activations
from PIL import Image
from transformers import (
AutoModel,
AutoProcessor,
Gemma3ForConditionalGeneration,
Qwen2_5_VLForConditionalGeneration,
)
from sglang.test.ci.ci_register import register_cuda_ci
if not hasattr(_hf_activations, "PytorchGELUTanh"):
class PytorchGELUTanh(torch.nn.Module):
def forward(self, x):
return torch.nn.functional.gelu(x, approximate="tanh")
_hf_activations.PytorchGELUTanh = PytorchGELUTanh
_hf_activations.ACT2FN.setdefault(
"pytorch_gelu_tanh",
lambda x: torch.nn.functional.gelu(x, approximate="tanh"),
)
from sglang import Engine
from sglang.srt.entrypoints.openai.protocol import ChatCompletionRequest
from sglang.srt.parser.conversation import generate_chat_conv
from sglang.srt.utils.hf_transformers_utils import _fix_added_tokens_encoding
register_cuda_ci(est_time=447, suite="stage-b-test-1-gpu-large")
IMAGE_MAN_IRONING_URL = "https://raw.githubusercontent.com/sgl-project/sgl-test-files/refs/heads/main/images/man_ironing_on_back_of_suv.png"
IMAGE_SGL_LOGO_URL = "https://raw.githubusercontent.com/sgl-project/sgl-test-files/refs/heads/main/images/sgl_logo.png"
class VLMInputTestBase:
model_path = None
chat_template = None
processor = None
visual = None # Should be a callable for precomputed embeddings
@classmethod
def setUpClass(cls):
assert cls.model_path is not None, "Set model_path in subclass"
assert cls.chat_template is not None, "Set chat_template in subclass"
cls.image_urls = [IMAGE_MAN_IRONING_URL, IMAGE_SGL_LOGO_URL]
cls.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
cls.main_image = []
for image_url in cls.image_urls:
response = requests.get(image_url)
cls.main_image.append(Image.open(BytesIO(response.content)))
cls.processor = AutoProcessor.from_pretrained(
cls.model_path, trust_remote_code=True, use_fast=True
)
_fix_added_tokens_encoding(cls.processor.tokenizer)
cls._init_visual()
@classmethod
def _init_visual(cls):
"""Override in subclass to set up cls.visual as a callable for precomputed embeddings."""
raise NotImplementedError
def setUp(self):
self.engine = Engine(
model_path=self.model_path,
chat_template=self.chat_template,
device=self.device.type,
mem_fraction_static=0.8,
enable_multimodal=True,
disable_cuda_graph=True,
trust_remote_code=True,
)
def tearDown(self):
self.engine.shutdown()
def verify_response(self, output):
# The goal is to check that the model roughly understands:
# - image 1: taxi / car scene
# - image 2: SGL logo / company
# We intentionally keep the check keyword-based and loose to avoid
# overfitting to a specific phrasing.
out_text = output["text"].lower()
assert any(w in out_text for w in ("taxi", "cab", "car")), out_text
has_sg_or_logo_side = any(
kw in out_text
for kw in (
"sg ",
"sgl",
" sgl",
"logo",
"software guidance",
"labs",
"laborator",
"company",
" text",
)
)
assert has_sg_or_logo_side, out_text
def get_completion_request(self) -> ChatCompletionRequest:
json_structure = {
"model": self.model_path,
"messages": [
{
"role": "user",
"content": [
{"type": "image_url", "image_url": {"url": self.image_urls[0]}},
{"type": "image_url", "image_url": {"url": self.image_urls[1]}},
{
"type": "text",
"text": "Describe both the first image and the second image in detail separately.", # update prompt, ensure kimi-vl understands the images separately.
},
],
}
],
}
json_str = json.dumps(json_structure)
return ChatCompletionRequest.model_validate_json(json_str)
def get_processor_output(self, req: Optional[ChatCompletionRequest] = None):
if req is None:
req = self.get_completion_request()
conv = generate_chat_conv(req, template_name=self.chat_template)
text = conv.get_prompt()
# Process inputs using processor
inputs = self.processor(
text=[text],
images=self.main_image,
return_tensors="pt",
).to(self.device)
return inputs, text
async def test_accepts_image(self):
req = self.get_completion_request()
conv = generate_chat_conv(req, template_name=self.chat_template)
text = conv.get_prompt()
output = await self.engine.async_generate(
prompt=text,
image_data=self.main_image,
sampling_params=dict(temperature=0.0, max_new_tokens=512),
)
self.verify_response(output)
async def test_accepts_precomputed_embeddings(self):
req = self.get_completion_request()
processor_output, _ = self.get_processor_output(req=req)
with torch.inference_mode():
precomputed_embeddings = self.__class__.visual(processor_output)
output = await self.engine.async_generate(
input_ids=processor_output["input_ids"][0].detach().cpu().tolist(),
image_data=[
self._precomputed_image_data(processor_output, precomputed_embeddings)
],
sampling_params=dict(temperature=0.0, max_new_tokens=512),
)
self.verify_response(output)
async def test_accepts_processor_output(self):
req = self.get_completion_request()
processor_output, prompt = self.get_processor_output(req=req)
output = await self.engine.async_generate(
input_ids=processor_output["input_ids"][0].detach().cpu().tolist(),
image_data=[self._processor_output_image_data(processor_output)],
sampling_params=dict(temperature=0.0, max_new_tokens=512),
)
self.verify_response(output)
def _precomputed_image_data(self, processor_output, precomputed_embeddings):
"""This should not be overridden."""
return dict(
processor_output,
format="precomputed_embedding",
feature=precomputed_embeddings,
)
def _processor_output_image_data(self, processor_output):
"""Override in subclass to pass the correct set of arguments."""
raise NotImplementedError
class TestQwenVLUnderstandsImage(VLMInputTestBase, unittest.IsolatedAsyncioTestCase):
model_path = "Qwen/Qwen2.5-VL-3B-Instruct"
chat_template = "qwen2-vl"
@classmethod
def _init_visual(cls):
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
cls.model_path, torch_dtype=torch.bfloat16
).eval()
# In transformers v5, .visual moved under .model
visual = model.model.visual
cls.visual_model = visual.to(cls.device)
# In transformers v5, the visual encoder returns BaseModelOutputWithPooling;
# pooler_output has the spatially-merged embeddings we need.
def visual(processor_output):
out = cls.visual_model(
processor_output["pixel_values"], processor_output["image_grid_thw"]
)
return out.pooler_output if hasattr(out, "pooler_output") else out
cls.visual = visual
def _processor_output_image_data(self, processor_output):
return dict(processor_output, format="processor_output")
class TestGemmaUnderstandsImage(VLMInputTestBase, unittest.IsolatedAsyncioTestCase):
model_path = "google/gemma-3-4b-it"
chat_template = "gemma-it"
@classmethod
def _init_visual(cls):
model = Gemma3ForConditionalGeneration.from_pretrained(
cls.model_path, torch_dtype=torch.bfloat16
)
base_model = model.model
cls.vision_tower = base_model.vision_tower.eval().to(cls.device)
if hasattr(base_model, "multi_modal_projector"):
cls.mm_projector = base_model.multi_modal_projector.eval().to(cls.device)
else:
cls.mm_projector = model.multi_modal_projector.eval().to(cls.device)
cls.visual = lambda processor_output: cls.mm_projector(
cls.vision_tower(
pixel_values=processor_output["pixel_values"]
).last_hidden_state
)
def _processor_output_image_data(self, processor_output):
return dict(processor_output, format="processor_output")
# Updated Kimi-VL test to use the new input format.
class TestKimiVLImageUnderstandsImage(
VLMInputTestBase, unittest.IsolatedAsyncioTestCase
):
model_path = "moonshotai/Kimi-VL-A3B-Instruct"
chat_template = "kimi-vl"
@classmethod
def _init_visual(cls):
import inspect
from transformers import AutoConfig
from transformers.dynamic_module_utils import get_class_from_dynamic_module
config = AutoConfig.from_pretrained(cls.model_path, trust_remote_code=True)
# Transformers v5 auto-populates rope_scaling with
# {"rope_theta": ..., "rope_type": "default"} even when the original
# config had rope_scaling: null. The remote KimiVL code branches on
# `if self.config.rope_scaling is None` so we must reset it.
tc = getattr(config, "text_config", None)
if tc is not None:
rs = getattr(tc, "rope_scaling", None)
if isinstance(rs, dict) and rs.get("rope_type") == "default":
tc.rope_scaling = None
# Transformers v5 calls tie_weights(recompute_mapping=False) in
# post_init, but KimiVL's tie_weights doesn't accept that kwarg.
auto_map = getattr(config, "auto_map", {})
model_ref = auto_map.get("AutoModel")
if model_ref:
model_cls = get_class_from_dynamic_module(model_ref, cls.model_path)
orig_tie = model_cls.tie_weights
if "recompute_mapping" not in inspect.signature(orig_tie).parameters:
def _patched_tie(self, **kwargs):
return orig_tie(self)
model_cls.tie_weights = _patched_tie
model = AutoModel.from_pretrained(
cls.model_path, config=config, trust_remote_code=True
)
cls.vision_tower = model.vision_tower.eval().to(cls.device)
cls.mm_projector = model.multi_modal_projector.eval().to(cls.device)
_vt_dtype = next(cls.vision_tower.parameters()).dtype
cls.visual = lambda tokenizer_output: cls.mm_projector(
cls.vision_tower(
pixel_values=tokenizer_output["pixel_values"].to(_vt_dtype),
grid_hws=tokenizer_output["image_grid_hws"],
)
)
def _processor_output_image_data(self, processor_output):
return dict(processor_output, format="processor_output")
# not for CI: too large
# class TestLlama4ImageUnderstandsImage(
# VLMInputTestBase, unittest.IsolatedAsyncioTestCase
# ):
# # Allow overriding via env for local/offline runs.
# model_path = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
# chat_template = "llama-4"
# def setUp(self):
# if torch.cuda.device_count() < 4:
# self.skipTest("Skipping Llama-4 test: requires 4 GPUs for TP=4")
# self.engine = Engine(
# model_path=self.model_path,
# trust_remote_code=True,
# chat_template=self.chat_template,
# enable_multimodal=True,
# mem_fraction_static=0.8,
# tp_size=4,
# attention_backend="fa3",
# context_length=65536,
# )
# @classmethod
# def _init_visual(cls):
# model = AutoModel.from_pretrained(
# cls.model_path,
# trust_remote_code=True,
# torch_dtype="auto",
# force_download=True,
# )
# cls.vision_tower = model.vision_model.eval().to(cls.device)
# cls.mm_projector = model.multi_modal_projector.eval().to(cls.device)
# cls.visual = lambda tokenizer_output: cls.mm_projector(
# cls.vision_tower(
# pixel_values=tokenizer_output["pixel_values"],
# ).last_hidden_state.flatten(0, -2)
# )
# def _processor_output_image_data(self, processor_output):
# # Llama-4 vision expects processor_output format with pixel_values
# return dict(processor_output, format="processor_output")
# class TestLlavaUnderstandsImage(VLMInputTestBase, unittest.IsolatedAsyncioTestCase):
# model_path = "llava-hf/llava-1.5-7b-hf"
# chat_template = "vicuna_v1.1"
# @classmethod
# def _init_visual(cls):
# from transformers import LlavaForConditionalGeneration
# model = LlavaForConditionalGeneration.from_pretrained(
# cls.model_path,
# torch_dtype=torch.float16,
# low_cpu_mem_usage=True,
# )
# cls.vision_tower = model.vision_tower.eval().to(cls.device)
# cls.multi_modal_projector = model.multi_modal_projector.eval().to(cls.device)
# cls.config = model.config
# def visual_func(processor_output):
# pixel_values = processor_output["pixel_values"].to(
# cls.device, dtype=torch.float16
# )
# vision_outputs = cls.vision_tower(pixel_values, output_hidden_states=True)
# image_features = vision_outputs.hidden_states[-2]
# if cls.config.vision_feature_select_strategy == "default":
# image_features = image_features[:, 1:]
# elif cls.config.vision_feature_select_strategy == "full":
# image_features = image_features
# image_features = cls.multi_modal_projector(image_features)
# return image_features
# cls.visual = visual_func
# def _processor_output_image_data(self, processor_output):
# return dict(processor_output, format="processor_output")
class TestInternVLUnderstandsImage(VLMInputTestBase, unittest.IsolatedAsyncioTestCase):
model_path = "OpenGVLab/InternVL2-2B"
chat_template = "internvl-2-5"
@classmethod
def setUpClass(cls):
assert cls.model_path is not None, "Set model_path in subclass"
assert cls.chat_template is not None, "Set chat_template in subclass"
cls.image_urls = [IMAGE_MAN_IRONING_URL, IMAGE_SGL_LOGO_URL]
cls.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
cls.main_image = []
for image_url in cls.image_urls:
response = requests.get(image_url)
cls.main_image.append(Image.open(BytesIO(response.content)))
# InternVL models (2, 3, 3.5) do not ship a standard HuggingFace
# Processor; AutoProcessor.from_pretrained returns a bare tokenizer.
# Use AutoTokenizer explicitly so the intent is clear.
from transformers import AutoTokenizer
cls.processor = AutoTokenizer.from_pretrained(
cls.model_path, trust_remote_code=True
)
cls._init_visual()
@classmethod
def _init_visual(cls):
try:
model = AutoModel.from_pretrained(
cls.model_path,
trust_remote_code=True,
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=False,
)
except RuntimeError as e:
if "meta" not in str(e):
raise
# Transformers v5 always uses meta tensors for init, which breaks
# models calling .item() in __init__ (e.g. InternVL's drop_path_rate).
# Fall back to from_config + manual weight loading.
import gc
import glob
import os
from huggingface_hub import snapshot_download
from safetensors.torch import load_file
from transformers import AutoConfig
config = AutoConfig.from_pretrained(cls.model_path, trust_remote_code=True)
with torch.device("cpu"):
model = AutoModel.from_config(
config,
trust_remote_code=True,
torch_dtype=torch.bfloat16,
)
model_dir = snapshot_download(cls.model_path)
for f in sorted(glob.glob(os.path.join(model_dir, "*.safetensors"))):
shard = load_file(f)
model.load_state_dict(shard, strict=False)
del shard
gc.collect()
cls.vision_model = model.vision_model.eval().to(cls.device)
cls.mlp1 = model.mlp1.eval().to(cls.device)
config = model.config
cls.internvl_config = config
image_size = getattr(config, "force_image_size", None) or (
config.vision_config.image_size
)
patch_size = config.vision_config.patch_size
cls.num_image_token = int(
(image_size // patch_size) ** 2 * (config.downsample_ratio**2)
)
cls.internvl_image_size = image_size
cls.internvl_downsample_ratio = config.downsample_ratio
cls.internvl_ps_version = config.ps_version
cls.internvl_select_layer = config.select_layer
del model
def pixel_shuffle(x, scale_factor):
n, w, h, c = x.size()
x = x.view(n, w, int(h * scale_factor), int(c / scale_factor))
x = x.permute(0, 2, 1, 3).contiguous()
x = x.view(
n,
int(h * scale_factor),
int(w * scale_factor),
int(c / (scale_factor * scale_factor)),
)
if cls.internvl_ps_version != "v1":
x = x.permute(0, 2, 1, 3).contiguous()
return x
def visual_func(processor_output):
pixel_values = processor_output["pixel_values"].to(
cls.device, dtype=torch.bfloat16
)
if cls.internvl_select_layer == -1:
vit_embeds = cls.vision_model(
pixel_values=pixel_values,
output_hidden_states=False,
return_dict=True,
).last_hidden_state
else:
vit_embeds = cls.vision_model(
pixel_values=pixel_values,
output_hidden_states=True,
return_dict=True,
).hidden_states[cls.internvl_select_layer]
vit_embeds = vit_embeds[:, 1:, :]
h = w = int(vit_embeds.shape[1] ** 0.5)
vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1)
vit_embeds = pixel_shuffle(
vit_embeds, scale_factor=cls.internvl_downsample_ratio
)
vit_embeds = vit_embeds.reshape(
vit_embeds.shape[0], -1, vit_embeds.shape[-1]
)
vit_embeds = cls.mlp1(vit_embeds)
return vit_embeds
cls.visual = visual_func
def get_processor_output(self, req=None):
"""Override to handle InternVL's custom preprocessing.
Uses shared ``image_to_pixel_values`` from ``internvl_utils`` for
image preprocessing (dynamic tiling + normalize) and expands
``<IMG_CONTEXT>`` placeholders into ``<img>`` + context tokens +
``</img>`` — mirroring the logic in
``InternVLProcessor.process_internlm2_mm_data_async``.
"""
from sglang.srt.multimodal.internvl_utils import image_to_pixel_values
from sglang.srt.multimodal.processors.internvl import InternVLProcessor
if req is None:
req = self.get_completion_request()
conv = generate_chat_conv(req, template_name=self.chat_template)
text = conv.get_prompt()
# Preprocess images using the shared utility (dynamic tiling +
# bicubic resize + ImageNet normalize), same pipeline as the engine.
all_pixel_values = []
num_patches_list = []
for img in self.main_image:
pv = image_to_pixel_values(
img,
input_size=self.internvl_image_size,
max_num_tiles=InternVLProcessor.IMAGE_MAX_NUM,
use_thumbnail=True,
)
all_pixel_values.append(pv)
num_patches_list.append(pv.shape[0])
pixel_values = torch.cat(all_pixel_values, dim=0).to(self.device)
# Expand each <IMG_CONTEXT> placeholder into <img> + <IMG_CONTEXT>*N + </img>.
# This mirrors InternVLProcessor.process_internlm2_mm_data_async.
ph = "<<<__IMG_PH__>>>"
expanded_text = text.replace(InternVLProcessor.IMG_CONTEXT, ph)
for num_patches in num_patches_list:
image_tokens = (
InternVLProcessor.IMG_START
+ InternVLProcessor.IMG_CONTEXT * (self.num_image_token * num_patches)
+ InternVLProcessor.IMG_END
)
expanded_text = expanded_text.replace(ph, image_tokens, 1)
# Remove any remaining placeholders (more placeholders than images)
expanded_text = expanded_text.replace(ph, "")
# Tokenize the expanded text
input_ids = self.processor(expanded_text, return_tensors="pt")["input_ids"]
return {
"input_ids": input_ids,
"pixel_values": pixel_values,
}, text
def _processor_output_image_data(self, processor_output):
return dict(processor_output, format="processor_output")
class TestMiniCPMVUnderstandsImage(VLMInputTestBase, unittest.IsolatedAsyncioTestCase):
model_path = "openbmb/MiniCPM-V-4"
chat_template = "minicpmv"
@classmethod
def setUpClass(cls):
assert cls.model_path is not None, "Set model_path in subclass"
assert cls.chat_template is not None, "Set chat_template in subclass"
cls.image_urls = [IMAGE_MAN_IRONING_URL, IMAGE_SGL_LOGO_URL]
cls.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
cls.main_image = []
for image_url in cls.image_urls:
response = requests.get(image_url)
cls.main_image.append(Image.open(BytesIO(response.content)))
cls.processor = AutoProcessor.from_pretrained(
cls.model_path, trust_remote_code=True
)
_fix_added_tokens_encoding(cls.processor.tokenizer)
cls._init_visual()
@classmethod
def _init_visual(cls):
try:
model = AutoModel.from_pretrained(
cls.model_path, trust_remote_code=True, torch_dtype=torch.bfloat16
)
except (AttributeError, RuntimeError) as e:
err = str(e)
if "all_tied_weights_keys" not in err and "meta" not in err:
raise
# Transformers v5: remote model code may lack all_tied_weights_keys
# or meta-tensor init may break .item() calls. Fall back to
# from_config + manual weight loading.
import gc
import glob
import os
from huggingface_hub import snapshot_download
from safetensors.torch import load_file
from transformers import AutoConfig
config = AutoConfig.from_pretrained(cls.model_path, trust_remote_code=True)
with torch.device("cpu"):
model = AutoModel.from_config(
config,
trust_remote_code=True,
torch_dtype=torch.bfloat16,
)
model_dir = snapshot_download(cls.model_path)
for f in sorted(glob.glob(os.path.join(model_dir, "*.safetensors"))):
shard = load_file(f)
model.load_state_dict(shard, strict=False)
del shard
gc.collect()
cls.vpm_model = model.vpm.eval().to(cls.device)
cls.resampler_model = model.resampler.eval().to(cls.device)
del model
def visual_func(processor_output):
pixel_values = processor_output["pixel_values"]
tgt_sizes = processor_output["tgt_sizes"]
pixel_values_flat = []
tgt_sizes_flat = []
for pixel_b, tgt_b in zip(pixel_values, tgt_sizes):
if isinstance(pixel_b, (list, tuple)):
for pixel_n, tgt_n in zip(pixel_b, tgt_b):
pixel_values_flat.append(pixel_n)
tgt_sizes_flat.append(tgt_n)
else:
pixel_values_flat.append(pixel_b)
tgt_sizes_flat.append(tgt_b)
tgt_sizes_tensor = torch.stack(tgt_sizes_flat, dim=0)
device = cls.vpm_model.embeddings.position_embedding.weight.device
dtype = cls.vpm_model.embeddings.position_embedding.weight.dtype
all_pixel_values_lst = [
i.flatten(end_dim=1).permute(1, 0) for i in pixel_values_flat
]
max_patches = int(
(tgt_sizes_tensor[:, 0] * tgt_sizes_tensor[:, 1]).max().item()
)
all_pixel_values = torch.nn.utils.rnn.pad_sequence(
all_pixel_values_lst, batch_first=True, padding_value=0.0
)
B, L, _ = all_pixel_values.shape
all_pixel_values = all_pixel_values.permute(0, 2, 1).reshape(B, 3, -1, L)
patch_attn_mask = torch.zeros(
(B, 1, max_patches), dtype=torch.bool, device=device
)
tgt_sizes_dev = tgt_sizes_tensor.to(device)
mask_shapes = tgt_sizes_dev[:, 0] * tgt_sizes_dev[:, 1]
patch_attn_mask[:, 0, :] = torch.arange(
max_patches, device=device
).unsqueeze(0) < mask_shapes.unsqueeze(1)
vision_output = cls.vpm_model(
all_pixel_values.type(dtype),
patch_attention_mask=patch_attn_mask,
tgt_sizes=tgt_sizes_tensor,
)
vision_embedding = vision_output.last_hidden_state
return cls.resampler_model(vision_embedding, tgt_sizes_tensor)
cls.visual = visual_func
def _processor_output_image_data(self, processor_output):
return dict(processor_output, format="processor_output")
if __name__ == "__main__":
unittest.main()