Roadmap
Upcoming features and improvements for the vLLM QAIC plugin.
Last updated: June 2026
AOT Mode
| Feature |
Status |
Description |
| LMCache integration |
 Ongoing |
KV cache sharing across fleet via LMCache |
| llm-d orchestration |
Ongoing |
Kubernetes-native distributed inference |
| Additional model architectures |
Ongoing |
Expanding validated model list |
PYT (Eager / JIT) Mode
| Feature |
Status |
Description |
| Fused HexNN kernels |
Ongoing |
PagedAttention, MLA, GatedLinearAttention |
torch.compile support |
 Planned |
Wider model coverage via compiled graphs |
| Triton kernel support |
Planned |
Execution of Triton Kernels |
| Text-only LLM support |
Ongoing |
Extending beyond VLMs |
| Speculative decoding |
Planned |
SpD support in eager mode |
Status Legend
| Symbol |
Meaning |
| Ongoing |
Active development |
| Planned |
On the roadmap, not yet started |
 Experimental |
Available but not production-ready |