3em0 3749457476 [BugFix] fix multimodal hasher hash collision risk when ndarray shape or dtype differs (#7185) ...

numpy tobytes() only serializes raw element bytes without encoding shape
or dtype metadata. This means arrays with identical raw bytes but
different shapes (e.g. (6,4) vs (4,6)) or different dtypes (e.g.
float32 vs uint8 reinterpretation of same memory) produce the same
SHA-256 digest, leading to silent cache collisions in
ProcessorCacheManager / EncoderCacheManager / PrefixCacheManager.

Prepend a "{shape}|{dtype}|" header to the byte payload before hashing
so that shape and dtype participate in the digest.

Added test cases for shape and dtype sensitivity.

2026-04-08 04:26:02 -07:00

..

__init__.py

[Feature] remove dependency on enable_mm and refine multimodal's code (#3014)

2025-08-01 20:01:18 +08:00

audio.py

[Feature] mm support prefix cache (#4134)

2025-10-27 17:39:51 +08:00

base.py

[Feature] mm support prefix cache (#4134)

2025-10-27 17:39:51 +08:00

hasher.py

[BugFix] fix multimodal hasher hash collision risk when ndarray shape or dtype differs (#7185)

2026-04-08 04:26:02 -07:00

image.py

[Bug Fix] process transparent image (#4807)

2025-11-05 17:15:24 +08:00

registry.py

[Feature] mm support prefix cache (#4134)

2025-10-27 17:39:51 +08:00

utils.py

[BugFix][Optimization] Replace silent failures with catchable exceptions and informative error messages (#6533)

2026-03-16 21:32:43 +08:00

video.py

[Feature] mm support prefix cache (#4134)

2025-10-27 17:39:51 +08:00