[BugFix] fix multimodal hasher hash collision risk when ndarray shape or dtype differs

[3em0]

Summary

Fixes #7196

MultimodalHasher.hash_features() uses np.ndarray.tobytes() to compute SHA-256 digests for multimodal cache keys. However, tobytes() only serializes raw element bytes — it does not encode shape or dtype metadata.

This means:

• Arrays with different shapes but identical flattened bytes (e.g. (6,4) vs (4,6)) produce the same hash
• Arrays with different dtypes but coincidentally same byte patterns (e.g. float32 vs uint8 reinterpretation) also collide

While the current inference pipeline uses a fixed dtype (float32) and deterministic reshape paths — making real-world collision extremely unlikely — the hash function itself is fundamentally unsafe and could silently return wrong cached results if any upstream change alters the preprocessing path.

Fix

Prepend a "{shape}|{dtype}|" header to the byte payload before hashing:

header = f"{obj.shape}|{obj.dtype}|".encode()
return hashlib.sha256(header + obj.tobytes()).hexdigest()

Test plan

• Updated test_hash_features_ndarray to match new hash format
• Added test_hash_features_ndarray_shape_sensitivity — verifies (6,4) vs (4,6) produce different hashes
• Added test_hash_features_ndarray_dtype_sensitivity — verifies float32 vs float64 produce different hashes
• Existing test_hash_features_object unchanged and passing

[paddle-bot]

Thanks for your contribution!

[CLAassistant]

All committers have signed the CLA.

[fastdeploy-bot]

🤖 AI Code Review | 2026-04-03 17:28 CST

📋 Review 摘要

PR 概述：修复 MultimodalHasher 在 ndarray shape 或 dtype 不同时的 hash 碰撞风险
变更范围：fastdeploy/multimodal/hasher.py、tests/multimodal/test_hasher.py
影响面 Tag：BugFix

问题

未发现阻塞性问题。

总体评价

这是一个高质量的 BugFix PR。修复方案通过在 hash 前添加 {shape}|{dtype}| header 来区分不同 layout 的数组，实现简洁有效。测试覆盖充分，新增了 shape 敏感性和 dtype 敏感性测试用例。PR 描述详细说明了问题根因和修复思路，符合规范。