[BugFix]: Slice image features correctly when prefix cache hits on multimodal requests (qwen2.5-vl)

[Smallhucaptain]

问题描述

enable_prefix_caching=True 时，发送两次相同的多模态请求，第二次崩溃：
ValueError: cannot copy sequence with size 3577 to array axis with dimension 14

根本原因

input_ids 被正确切片到 [prefill_start, prefill_end)，
但 image_features 没有对应切片，仍然返回完整的 [3577, 3584]，
导致 get_input_embeddings() 中形状不匹配。

修复方案

新增 _calc_image_feature_range()，根据 mm_positions 计算
prefill 范围对应的图像特征切片范围，在 extract_vision_features()
之后进行切片。

测试验证

• 两次相同请求均正常返回，无崩溃
• sliced_feature_count == image_tokens_in_prefill (14 == 14)
• 两次请求的图像特征 Last row 完全一致

[paddle-bot]

Thanks for your contribution!

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
17 out of 18 committers have signed the CLA.

✅ liyonghua0910
✅ EmmonsCurse
✅ kevincheng2
✅ iosmers
✅ Deleter-D
✅ gongshaotian
✅ yuanlehome
✅ fxyfxy777
✅ a31413510
✅ yangjianfengo1
✅ plusNew001
✅ Jiang-Jia-Jun
✅ qwes5s5
✅ luukunn
✅ ApplEOFDiscord
✅ xiaoxiaohehe001
✅ ZhangYulongg
❌ huzesen

---

huzesen seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}