[BugFix]Add MiniCPM-o 4.5 end-to-end test suite (offline, online serving) (#5237)

Signed-off-by: y_null <y_null@qq.com>
Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com>
Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
This commit is contained in:
authored and GitHub committed 2026-07-25 17:00:13 -07:00
1 parent 76e4a78afc
commit b6c4578dac
27 files changed
+1338 -46

No files matched your search

+19 -6
View File
@@ -2424,6 +2424,10 @@ class OmniRunner:
video_padding_token = "<|VIDEO|>"
image_padding_token = "<|IMAGE|>"
audio_padding_token = "<|AUDIO|>"
# Default wrapping for Qwen-style models (bos/eos around the placeholder).
audio_fmt = "<|audio_bos|>{p}<|audio_eos|>"
image_fmt = "<|vision_bos|>{p}<|vision_eos|>"
video_fmt = "<|vision_bos|>{p}<|vision_eos|>"
if "Qwen3-Omni-30B-A3B-Instruct" in self.model_name:
video_padding_token = "<|video_pad|>"
image_padding_token = "<|image_pad|>"
@@ -2432,6 +2436,15 @@ class OmniRunner:
video_padding_token = "<VIDEO>"
image_padding_token = "<IMAGE>"
audio_padding_token = "<AUDIO>"
elif "MiniCPM" in self.model_name:
# MiniCPM-o expects the bare placeholder literals (with parens and ./),
# not Qwen-style bos/eos wrapping.
video_padding_token = "(<video>./</video>)"
image_padding_token = "(<image>./</image>)"
audio_padding_token = "(<audio>./</audio>)"
audio_fmt = "{p}"
image_fmt = "{p}"
video_fmt = "{p}"
if isinstance(prompts, str):
prompts = [prompts]
@@ -2492,28 +2505,28 @@ class OmniRunner:
if audio is not None:
if isinstance(audio, list):
for _ in audio:
user_content += f"<|audio_bos|>{audio_padding_token}<|audio_eos|>"
user_content += audio_fmt.format(p=audio_padding_token)
multi_modal_data["audio"] = audio
else:
user_content += f"<|audio_bos|>{audio_padding_token}<|audio_eos|>"
user_content += audio_fmt.format(p=audio_padding_token)
multi_modal_data["audio"] = audio
image = images_list[i]
if image is not None:
if isinstance(image, list):
for _ in image:
user_content += f"<|vision_bos|>{image_padding_token}<|vision_eos|>"
user_content += image_fmt.format(p=image_padding_token)
multi_modal_data["image"] = image
else:
user_content += f"<|vision_bos|>{image_padding_token}<|vision_eos|>"
user_content += image_fmt.format(p=image_padding_token)
multi_modal_data["image"] = image
video = videos_list[i]
if video is not None:
if isinstance(video, list):
for _ in video:
user_content += f"<|vision_bos|>{video_padding_token}<|vision_eos|>"
user_content += video_fmt.format(p=video_padding_token)
multi_modal_data["video"] = video
else:
user_content += f"<|vision_bos|>{video_padding_token}<|vision_eos|>"
user_content += video_fmt.format(p=video_padding_token)
multi_modal_data["video"] = video
user_content += prompt_text