[Leonccaa]

Change

ModelConfig.compute_hash ignored limit_mm_per_prompt. A zero limit changes the traced language-model graph (the vision tower is replaced when image and video are both zero; input IDs are fed instead of embeddings when every modality is zero and precomputed embeddings are off), so a text-only config and a multimodal config sharing one compile-cache root could load each other's compiled graphs.

Upstream vLLM #54642 hashes every limit value. Non-zero values do not change the graph, so this change hashes only the sorted set of zero-limit modalities together with enable_mm_embeds, and only when some limit is zero. Configs without a zero limit keep exactly the previous hash and their existing compile caches. A comment in mm_limit_hash_factors notes that a model which uses a non-zero limit value inside the compiled graph would have to be added there.

No new switch.

Risk and rollback: configs with a zero limit get a new compile-cache key once (one cold compile). Revert this commit to return to the previous key.

Validation

• tests/config/test_mm_limit_compute_hash.py: 9 passed: factor helper cases, and ModelConfig hashes on a local Qwen2-VL config (no zero limit gives exactly the previous hash; text-only, image-off and embeds variants all differ).
• The hash of a config without zero limits was compared byte for byte with the unpatched code.
• GPU: part of the 0.2.0rc4 release (DFlash2 production configuration on 4x V100 with TP4, multimodal limits all non-zero, so its key is unchanged), now running in production.
• pre-commit on the changed files, including mypy.

Attribution

Adapted from upstream vLLM #54642 (conditional on zero limits, see above); upstream authorship credited. AI-assisted change (Claude); the submitter reviewed every changed line. Commits are signed off.

🤖 Generated with Claude Code