Feature Support Matrix

Status of vLLM features on Qualcomm Cloud AI hardware.

Status Legend

Symbol	Meaning
✅	Supported and validated
🧪	Experimental
📝	Planned
❌	Not supported

Feature Matrix

Feature	AOT Mode	Eager Mode	Notes
Inference
Text generation	✅	✅	Core serving capability
Continuous batching	✅	✅
Quantization
mxfp6	✅	❌	Hardware-native compute quantization
mxint8 KV cache	✅	❌	--kv-cache-dtype mxint8
Speculative Decoding
N-gram	✅	❌
Suffix	✅	❌
Draft model	✅	❌	Separate DLM on same device
Advanced Features
LoRA adapters	✅	❌	Hot-swap adapters
Disaggregated serving	✅	❌	xEyPzD prefill/decode split
Multimodal (VLM)	✅	🧪	kv_offload architecture
Embedding models	✅	❌	Pooling tasks (Score, embed, classify, rerank)
Encoder-decoder	✅	❌	Whisper
Tensor parallelism	✅	✅	Across QIDs
Pipeline parallelism	✅	❌	Across QIDs

Known Constraints

Constraint	Scope	Notes
Prefix caching	Both modes	📝 Planned
MLA attention	Both modes	Multi-head Latent Attention — 📝 Planned
Async output	Both modes	supports_async_output is False — 📝 Planned

Feature Combinations

Some features cannot be used together. Here is the compatibility matrix for common combinations:

Combination	AOT	Eager	Notes
SpD + Disaggregated	✅	❌	Supported on decode node
LoRA + Disaggregated	❌	❌	Not supported
Multimodal + CCL (Compiled Context Lengths)	✅	🧪	Supported (vision encoder on separate device)
Tensor Parallel + Disaggregated	✅	❌	Supported (TP within each node)
Quantization + LoRA	✅	❌	Supported (quantized base + LoRA adapters)