Add VLM inference infrastructure: engine, protocol, and CLI support#65
Merged
Conversation
carinapeng
reviewed
Jun 29, 2026
carinapeng
reviewed
Jun 29, 2026
carinapeng
reviewed
Jun 29, 2026
carinapeng
reviewed
Jun 29, 2026
carinapeng
reviewed
Jun 29, 2026
carinapeng
reviewed
Jun 29, 2026
Runtime: - MultimodalInferenceEngine protocol with encodeImage() and generate() - CoreAISequentialVLMEngine: vision encoder + projector + embed_tokens + LLM decoder with scatter-merge of image embeddings at placeholder positions - EmbeddedInput type wrapping NDArray embeddings with position metadata - VisionConfig in LanguageConfig for image_size, patch_size, token count/id - LanguageBundle parses top-level "vision" block from metadata.json CLI (llm-runner): - --image flag routes through VLM engine when bundle kind is .vlm - Chat template detection with generic fallback for prompt construction - Accumulated token decode for correct spacing - Stop sequence support in VLM path Supports any VLM that exports 3 components (vision.aimodel, embed.aimodel, model.aimodel) with a vision config block in metadata.json. Model-family- specific export code lives in internal/python.
Read per-channel normalization from metadata instead of hardcoding. Fields are optional — bundles without them default to CLIP values (the most common across VLMs). Gemma/SigLIP bundles specify their own [0.5, 0.5, 0.5] values explicitly in metadata.json.
…bfloat16 - LanguageBundle init throws if a .vlm bundle omits the vision block - Rename imageTokenPositions → embeddingPositions (supports future audio/multi-modal embedding injection, not just images) - Accept bfloat16 logits in addition to float16 - Scatter merge uses UInt16 view (type-agnostic for f16/bf16)
Contributor
carinapeng
approved these changes
Jun 29, 2026
…arnings - Guard against currentKVCapacity==0 in growKVCache (would loop forever) - Prefix unused protocol args with _ in warmup()
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Runtime:
CLI (llm-runner):
Supports any VLM that exports 3 components (vision.aimodel, embed.aimodel, model.aimodel) with a vision config block in metadata.json. Model-family- specific export code lives in internal/python.