[RyanPCo]

Turns QwenVLA into a chassis that accepts any backbone, then adds the second one.

• 0d1a5c3f3 rename QwenVLA to FlowVLA. The chassis never mentioned Qwen outside its constructor — prompt assembly, batch plumbing, norm stats and the eval contract are all backbone-agnostic. Drops FlowVLAModel's QwenVLBackbone type hint in favour of the documented duck-typed contract and relabels the LR-group dict backbone instead of vlm. No behaviour change.
• 4e27d41a1 ResNetTextBackbone — the only new module of the ResNet+text arm. One ResNet per canonical camera role (front is the same module for every vendor, since VIZ_IMAGE_KEY is), each fed a single frame because ResNet.forward's reshape is only channel-correct at F == 1; the frozen Qwen3-Embedding per-token features and their mask concatenated after them; modality embeddings so a patch token is distinguishable from a word. It returns the one context once per DiT block — the head's per-layer projectors are what give each block its own view.
• 4a2afe70f recipe config — train_zarr_mecka_flagship_kp_flowvla is train_zarr_mecka_flagship_kp_hpt with one line changed, so the three arms share data, evaluators, val heads and step budget by construction. embodiment_label off makes the prompt exactly the HPT arm's bare "fold clothes"; dit_hidden 768 != the backbone's 1024 keeps the head's per-layer projectors real Linears instead of nn.Identity.
• 2f54b0542 train-step arm — two optimizer steps through the real trainHydra.train() path with the text encoder swapped for the fixtures stub, so no 596M checkpoint is loaded. The proprio override records a real difference: FlowVLA concatenates every proprio key while HPT selects them with stem_specs, so plain data/mecka is 164-wide here where the flagship keypoint config (ee_pose off) is 144.

Testing

83 passed across test_flowvla_algo.py, test_flowvla_nets.py, test_flowvla_configs.py and test_pi05_norm_rot6d.py.

🤖 Generated with Claude Code

[RyanPCo]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• FlowVLA: backbone-neutral chassis and a ResNet+text backbone #661 👈 (View in Graphite)
• QwenVLA: a Qwen 3.5 VLM policy with a layer-wise cross-attention flow head #660
• hpt: Qwen 3.5-0.8B VLM stem pair (one joint image+text forward) #648
• deps: transformers 5.17.0 (Qwen 3.5), with openpi's patch ported #647
• hpt: cache frozen Qwen text features per unique prompt #646
• hpt: never re-initialise pretrained stems; hash-check them; mask text padding #645 : 1 other dependent PR (#664 )
• obs: proprio history (K proprio steps, one token each, learned time embedding) #650
• loss: mask the repeat-last action padding out of the HPT loss #649
• resnet: keep a frozen backbone in eval mode; give backbones their own LR group #652
• augs: per-sample image augmentation for HPT train-mode augs #651
• eval: deterministic validation (seeded prompts + flow sampling); log-only grad-norm MAD #644
• data: mecka_fold_flagship_top3_hpt_6d, the 3-operator flagship train split #643
• val: subsampled metric loaders + pinned contiguous video loaders #642
• infra: mecka 6D / keypoint recipes, 300M models, defaults, and the rollout revert #640
• eval: layout-aware frame-invariant metrics, and a train-split viz head #659
• data: stop letting one bad episode or a rotation channel take a run down #658
• norm: guard degenerate channels, horizon pooling, and no rot6d normalization #657
• keypoints: the 144-D wrist-first hand-keypoint action becomes the human default #656
• mecka: fix the double-mirrored LEFT wrist frame, and correct existing zarrs #655
• 6d: PI goes 6D-only; Eva and Human gain the continuous-rotation modes #654
• 6d: continuous rotation representation, bimanual layouts, shared norm helpers #653
• train: give each Slurm slurm job gets its own cache dir to avoid conflict #638 : 1 other dependent PR (#639 )
• write git version into the run dir for reproducibility #637
• train: don't load PI base weights on resume and verify that checkpoint path exists #636
• configs: declare proprio and actions dims once per HPT model config #634 : 1 other dependent PR (#635 )
• cache norm stats for data configs to minimize recomputation #633
• import cloud packages lazily to support local training #631 : 1 other dependent PR (#632 )
• cache dataset resolver for faster training initialization #630
• validate episode hashes in data config before training #628
• modify pi camera map to match data #629
• add tests to ensure training works between every model and data vendor #626
• run CI tests on every PR #625
• add env variables for user-specific paths #624
• build: set pyproject and uv as the dependency managers #623
• fix(configs): fix paths for data hydra yaml configs #622
• refactor(algo): delete ACT #627 : 1 other dependent PR (#621 )
• tests: create dedicated test folder with unit and integration tests #620
• add pyright type checker #619
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.