feat: share multimodal hash helpers

[CUHKSZzxy]

Summary

• Compute multimodal content hashes in the shared VL preprocessing path after get_expanded_mm_items, so the hash is based on processed, expanded multimodal items instead of PyTorch-engine-local data.
• Attach content_hash to each expanded multimodal item while excluding prompt-position offset, keeping the same content reusable at different prompt positions.
• Preserve the attached content_hash through PyTorch multimodal input processors into MultiModalData, and remove the PyTorch engine-side hash population path while keeping the scheduler fallback for defensive compatibility.

Validation

• Focused pre-commit checks for the touched implementation and test files passed.
• Focused VL hasher, VL preprocess, PyTorch multimodal hash preservation, and block-trie prefix-cache unit tests passed.
• Qwen3-VL processor tests passed.
• Real PyTorch VL server repeated-image prefix-cache check matched the baseline main branch for prompt/cached-token behavior and showed cache hits on repeated multimodal requests.

Assistance

Assisted with Codex + GPT-5.5 xHigh Fast, reviewed manually

[Copilot]

Pull request overview

This PR centralizes multimodal content hashing into a shared lmdeploy.vl helper module, then updates PyTorch prefix-cache code paths (and tests) to use the shared implementation while keeping existing cache-key behavior stable.

Changes:

• Added lmdeploy/vl/hasher.py with deterministic hashing helpers for both dataclass-style and dict-style multimodal payloads.
• Rewired PyTorch prefix-cache hashing call sites to use the shared VL hasher (including unit test monkeypatch targets).
• Added focused unit tests covering hash stability, sensitivity to content/meta/mRoPE, and ignoring position-only keys for dict-style items.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
lmdeploy/vl/hasher.py	Introduces shared deterministic multimodal hashing + “ensure content_hash” helpers for two multimodal representations.
lmdeploy/pytorch/multimodal/data_type.py	Removes local hashing implementation and re-exports shared hashing helpers for compatibility.
lmdeploy/pytorch/messages.py	Updates prefix-cache meta hashing fallback to call the shared VL hasher.
lmdeploy/pytorch/engine/engine.py	Ensures multimodal content hashes are populated after preprocessing when prefix caching is enabled.
tests/test_lmdeploy/test_vl/test_hasher.py	Adds unit tests validating hash determinism and correct inclusion/exclusion rules.
tests/pytorch/paging/test_block_trie.py	Adjusts monkeypatching to target the shared hasher module instead of the previous PyTorch-local symbol.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[grimoire]

Since we have alias make_multimodal_content_hash in pytorch/multimodal/data_type.py, the change in engine.py and messages.py might not be necessary.

[lvhan028]

It does not match what I had in mind for multimodal fingerprints.

What I expect

• Compute a stable content fingerprint early in the request path (before engine-specific preprocessing).
• Attach it to each multimodal item and pass it through to all backends
• Treat the fingerprint as general multimodal metadata, not something that only exists when enable_prefix_caching is on.

requirement mismatched

[Copilot]

Pull request overview

Copilot reviewed 23 out of 23 changed files in this pull request and generated 1 comment.

[CUHKSZzxy]

This is intentional. The human review guidance for this PR is to

treat the multimodal fingerprint as general multimodal metadata, not as prefix-cache-only state.

Prefix caching is one consumer of content_hash, but the hash is attached in the shared VL path, so it can be passed through consistently and reused by other backends later.

[grimoire]

has preprocess been override in other classes?

[CUHKSZzxy]

Nice catch. Now I have moved the hash attachment to a more general place outside the base vision model class.