Repository navigation

QwenVLA: a Qwen 3.5 VLM policy with a layer-wise cross-attention flow head - #660

Closed

RyanPCo wants to merge 9 commits into

ryanco/qwen35-vlmfrom

Closed

QwenVLA: a Qwen 3.5 VLM policy with a layer-wise cross-attention flow head#660
RyanPCo wants to merge 9 commits into
ryanco/qwen35-vlmfrom
ryanco/qwenvla

Commits on Sep 17, 2026
- algo: DeterministicEvalMixin, the seeded eval helpers HPT and QwenVLA share
  
  RyanPCo
  and
  claude
  committed
  bdf2790
- tests: multi-image fake Qwen 3.5 fixture for QwenVLA
  
  RyanPCo
  and
  claude
  committed
  785b415
- models: layer-wise cross-attention DiT + LayerwiseFMHead (GR00T N1.5 / starVLA port, torch primitives)
  
  RyanPCo
  and
  claude
  committed
  9905b88
- models: QwenVLBackbone (multi-frame Qwen 3.5, all layers) and QwenVLAModel
  
  RyanPCo
  and
  claude
  committed
  1aeeb55
- algo: QwenVLA (embodiment prompt, padded shared action width, deterministic eval)
  
  RyanPCo
  and
  claude
  committed
  45d199f
- configs: QwenVLA model / recipe yamls; flagship keypoint data + recipes for the QwenVLA vs HPT comparison
  
  RyanPCo
  and
  claude
  committed
  d7c81f2
- trainer: ddp_qwenvla (unused-parameter detection + bf16 gradient compression)
  
  RyanPCo
  and
  claude
  committed
  80609cd
- qwenvla: build the algo on the CPU; the device setter moves the nets once Lightning assigns the rank's GPU
  
  RyanPCo
  and
  claude
  committed
  6acc494
- eval: overlay videos play at the source rate under DDP
  
  RyanPCo
  and
  claude
  committed
  c3d071e