Adding original pi07 with gemma3 backbone and space-time siglip video… by akshay18iitg · Pull Request #197 · TensorAuto/OpenTau

akshay18iitg · 2026-04-29T01:34:19Z

… encoder

What this does

Adds a new pi07 policy built on the Gemma 3 VLM backbone and SpaceTimeSiglip video encoder, replacing the PaliGemma + V-JEPA2 stack from pi07_paligemma. (🗃️ Feature)

New files

File | Purpose -- | -- src/opentau/policies/pi07/__init__.py | Package init src/opentau/policies/pi07/gemma3_with_expert.py | Gemma3WithExpertConfig / Gemma3WithExpertModel — Gemma 3 VLM interleaved layer-wise with a Gemma-v1 action expert (AdaRMS, shared attention, knowledge insulation) src/opentau/policies/pi07/high_level_planner/ | High-level planner (memory + subtask prediction) using the Gemma 3 backbone src/opentau/policies/pi07/low_level_planner/ | Low-level planner (flow-matching continuous actions + FAST discrete tokens) using Gemma 3 + SpaceTimeSiglip src/opentau/policies/pi07/low_level_planner/video_encoder.py | SpaceTimeSiglipVideoEncoder — wraps Gemma 3's SigLIP vision tower with space-time separable attention for temporal video encoding tests/policies/test_pi07_high_level_planner.py | GPU integration test for the high-level planner (forward + autoregressive inference) tests/policies/test_pi07_low_level_planner.py | GPU integration test for the low-level planner (forward + select_action + 4 regression tests)

Key architectural changes vs `pi07_paligemma`

Gemma 3 (4B) replaces PaliGemma (3B) as the VLM backbone — larger text hidden size (2560 vs 2048), 34 layers, GQA with per-head Q/K RMSNorm, sliding/global attention pattern
SpaceTimeSiglip replaces V-JEPA2 as the video encoder — reuses the Gemma 3 SigLIP vision tower (no weight duplication) with causal temporal self-attention injected at every layer
448×448 image resolution (up from 224×224) — resize_imgs_with_padding defaults updated, Gemma3MultiModalProjector outputs 256 tokens per image
proj_width default changed to 1280 to match the larger action expert's hidden_size
embed_image rewritten to directly call vision tower + projector, returning a plain tensor (Gemma 3's get_image_features returns BaseModelOutputWithPooling, not a tensor)
_lm_head() accessor added — Gemma3ForConditionalGeneration owns lm_head directly (unlike PaliGemma where it's on the text model)
SpaceTime short-circuit for non-video inputs — the shared wrapped tower gracefully skips temporal attention when batch_size % num_frames != 0 (e.g. single subgoal images via embed_image)

## How it was tested Added tests/policies/test_pi07_high_level_planner.py — TestPI07HighLevelPlannerIntegration exercises the full training forward pass (MSE == 0, CE finite, mask/position-id layout verification) and autoregressive inference (step count, prefix growth, output shapes) Added tests/policies/test_pi07_low_level_planner.py — TestPI07LowLevelPlannerIntegration exercises training forward (VLM + action expert attention masks, normalize/unnormalize round-trip, loss finiteness) and inference via select_action; TestPI07LowLevelPlannerRegression pins 4 regression cases (prepare_metadata returns tensors, zip strict catches mismatch, suffix att_masks are bool, embed_prefix signature has no defaults) Both test files use a tiny VLM config (2 layers, 512/256 hidden) to fit within 24 GB GPU memory while exercising the full code path Examples: - Added `test_something` in `tests/test_stuff.py`. - Added `new_feature` and checked that training converges with policy X on dataset/environment Y. - Optimized `some_function`, it now runs X times faster than previously.

How to checkout & try? (for the reviewer)

Run the high-level planner integration test

pytest -sx tests/policies/test_pi07_high_level_planner.py -m "gpu and slow"

Run the low-level planner integration test

pytest -sx tests/policies/test_pi07_low_level_planner.py -m "gpu and slow"

Run just the regression tests (faster)

pytest -sx tests/policies/test_pi07_low_level_planner.py::TestPI07LowLevelPlannerRegression -m "gpu and slow"

Checklist

I have added Google-style docstrings to important functions and ensured function parameters are typed.
My PR includes policy-related changes.
- If the above is checked: I have run the GPU pytests (pytest -m "gpu") and regression tests.

Note: Before submitting this PR, please read the contributor guideline.

… encoder

shuheng-liu · 2026-04-29T04:59:31Z

Thanks for putting this together — the structural integration of #167 + #178 + #171 is correct, but on closer reading there are a few issues we'd like to flag, and we'll open a fix-up PR (claude/review-pr-197-xg92L → feat/pi07) to help resolve them.

Findings

1. Critical — text-embedding double-scale fix from #178 was not carried over.
PR #178's central architectural correction was that Gemma 3's embed_tokens is a Gemma3TextScaledWordEmbedding which already multiplies by √hidden_size internally, so the pi05-style manual * math.sqrt(emb_dim) must be removed; otherwise text tokens are scaled to ~51× the image-token magnitude and both the bidirectional prefix attention and the FAST/response cross-entropy heads are silently miscalibrated. PR #197 swapped the backbone to Gemma 3 but kept the manual scaling at all 14 sites copied from #167:

pi07/low_level_planner/modeling_pi07_low_level.py lines 1169, 1185, 1217, 1228, 1244, 1283, 1296, 1309, 1331
pi07/high_level_planner/modeling_pi07_high_level.py lines 921, 933, 946, 965, 980, 1015

2. Test coverage is GPU-only — none of the architectural invariants from #178/#171 are checked.
PR #178 shipped 26 CPU-only tests (RoPE θ symmetry across the two streams, no-sliding-window enforcement, projector-at-448, vision_image_size invariant, embedding-magnitude invariant). PR #171 shipped 46 CPU-only tests (state-dict invariance, single-frame byte-exact invariance, wrapped-layer indices, temporal-PE boundary). PR #197 has only @pytest.mark.gpu @pytest.mark.slow integration tests — neither the double-scale nor a regression to the RoPE-θ pairing or SigLIP weight sharing would be caught on the CPU CI.

3. available_policies not updated.
src/opentau/__init__.py:152 still reads ["pi0", "pi05", "pi05_mem", "value"]. The new pi07_high_level / pi07_low_level registry entries in factory.py are unreachable from that list, and the tests/test_available.py invariant test won't pick them up.

4. SpaceTimeEncoderLayerWrapper short-circuit can silently mis-trigger.
forward bypasses temporal attention when bt % t != 0 so the same wrapped tower can serve subgoal images via embed_image. But if a caller batches subgoals such that B*K happens to be divisible by num_frames, the temporal causal SDPA will fire over images that aren't temporally related, with no warning. Better to plumb an explicit temporal_active flag (we'll wire one through a context manager on the encoder).

5. _fix_pytorch_state_dict_keys only matches the new gemma3_with_expert.* prefix.
A user with a pi07_paligemma checkpoint can't warm-start pi07. PR #178 handled the equivalent by accepting both prefixes in its hook.

Tier-3 cosmetic: pi07/gemma3_with_expert.py:418 says "fa2 attention is not supported for pi06 yet" (copy-paste leftover, should say "pi07"); the embed_image docstring claims Gemma3ForConditionalGeneration.get_image_features returns a BaseModelOutputWithPooling — in current transformers it returns the projected tensor directly.

Open question: spacetime_layer_stride defaults to 1 in PI07LowLevelPlannerConfig (every SigLIP layer wrapped) vs 4 in #171 (every 4th, MEM paper). Was wrapping every layer intentional? We'll leave the default unchanged in the fix-up PR and just call it out.

The fix-up PR will land Tier-1 and Tier-2 fixes plus ported CPU tests, leaving GPU integration unchanged. We'll link it here once it's open.

Generated by Claude Code

shuheng-liu · 2026-04-29T05:14:36Z

Fix-up PR opened: #198 (draft, targets feat/pi07).

Summary of what landed there:

Tier 1 — correctness (blocker). Removes the manual * math.sqrt(*_dim) double-scale at all 14 sites in pi07/{low,high}_level_planner/modeling_pi07_*.py.
Tier 2. pi07_high_level / pi07_low_level added to available_policies (+ tests/test_available.py); bt % t != 0 short-circuit replaced with an explicit suppress_spacetime_temporal(...) context manager wired into embed_image; paligemma_with_expert.* → gemma3_with_expert.* rewrite added to _fix_pytorch_state_dict_keys for both planners.
Tier 3. pi06 → pi07 in the fa2 NotImplementedError; embed_image docstring corrected.
Tests. 8 CPU tests for the Gemma 3 backbone (test_pi07_cpu.py) + 17 CPU tests for the SpaceTime SigLIP video encoder (test_pi07_video_encoder_cpu.py); all 27 pass locally and pre-commit is clean.

The one open question — spacetime_layer_stride default of 1 vs. 4 (#171 / MEM-paper) — is left unchanged in #198. Happy to flip in a follow-up if you confirm 1 was intentional.

Generated by Claude Code

shuheng-liu · 2026-04-29T05:16:27Z

@claude spacetime_layer_stride should default to 4.

shuheng-liu · 2026-04-29T05:19:39Z

Updated #198 to use 4 as the default spacetime_layer_stride in accordance with pi-mem paper.

claude

Superseded — see follow-up review and summary comment below.

claude

[claude-review] Spot review — see PR-level summary comment for details.

claude · 2026-04-29T21:26:03Z

[claude-review] summary for commit e24ba88

Re-reviewing on commit e24ba88 fix(pi07): align text attention with paper §VI.B + Fig. 19 (#236). New since the previously-reviewed 62a5677: two upstream commits (6f256ec SDPA + gradient checkpointing, e24ba88 attention-layout fix in pi07_paligemma) plus the [claude-fix] address review feedback commit b383089. Diff vs 62a5677 is +904/-243 across 12 files. Substantive changes in gemma3_with_expert.py (SDPA forward + _run_layer extraction for ckpt), the two pi07 configurations (config plumbing into vlm_config), pi07_paligemma/{high,low}_level_planner (paper §VI.B + Fig. 19 layout reorder), plus matching tests.

Previous review carryover resolved.

The five suggestions from the 62a5677 review remain resolved.
Two of the four new findings on 62a5677 are addressed in b383089/e24ba88:
- HL planner gate semantics tightened to metadata_tokens is not None and metadata_masks.any() (pi07/high_level_planner/modeling_pi07_high_level.py:986-989), matching LL behavior.
- Batch-wide gate documented in pi07/low_level_planner/modeling_pi07_low_level.py:1165-1175 docstring.

Carryover from earlier reviews still applicable.

The all-optional-blocks layout in tests/policies/test_pi07_low_level_planner.py:198-222 still hard-codes prefix_end + comma; a GPU regression for the no-optionals path remains open.
No CPU test pins the RoPE-θ pairing across the two streams (pi07/gemma3_with_expert.py apply_rope sites); a future regression there would only be caught under GPU integration.
test_pi06.py:516-532 actions stats reshape still flattens per-step variance — small fixture-quality nit.

New findings on this round.

suggestion — src/opentau/policies/pi07/low_level_planner/modeling_pi07_low_level.py:1257 (and 1274/1287/1305/1383/1411) — Commit e24ba88 is titled fix(pi07): align text attention with paper §VI.B + Fig. 19 but the diff only touches the legacy pi07_paligemma package. The new gemma3-backbone pi07 policy that this PR primarily ships still emits [0] * N (bidirectional) for language / State: / state tokens / state-end / subgoal-end / ";\n " and [1] + [0] * (N-1) (prefix-LM) for the response span — exactly the pre-fix pattern. The pi07_paligemma diff also reorders the prefix so subgoals sit AFTER metadata + ";\n " (Fig. 19); the gemma3 prefix here keeps the old order. Worth confirming this split is intentional vs. an oversight in cherry-picking.
suggestion — src/opentau/policies/pi07/high_level_planner/modeling_pi07_high_level.py:982 and :1017 — Same point as LL: language and ";\n " blocks are still bidirectional, while pi07_paligemma/.../modeling_pi07_high_level.py:932,962 were flipped to [1] * N per-token causal in e24ba88.
suggestion — src/opentau/policies/pi07/{high_level_planner,low_level_planner}/configuration_pi07_*.py:203-206 — __post_init__ plumbing uses asymmetric "non-default → propagate" logic. Setting --policy.attention_implementation=eager or --policy.gradient_checkpointing=False to explicitly revert a non-default vlm_config.* value silently no-ops. Documented as intentional via test_post_init_preserves_explicit_vlm_config_when_policy_default, but worth flagging as a footgun for sweep-style overrides.
suggestion — tests/policies/test_pi07_cpu.py:929 (TestPi07GradCkptEquivalence.test_grad_ckpt_forward_matches_no_ckpt) — Forward-only check; torch.utils.checkpoint.checkpoint is a no-op in forward and recomputes the body in backward, so saved-tensor-hook / RNG bugs in the _run_layer extraction would only manifest after loss.backward(). Per CLAUDE.md's deterministic-training rule, a loss.backward() + per-param grad-equality check would actually exercise the recompute path.
suggestion — tests/policies/test_pi07_paligemma_attention_layout.py — Tests exercise make_att_2d_masks with hand-coded att_masks arrays, not the post-fix embed_prefix output. A regression flipping att_masks += [1] * num_lang_embs back to [0] * num_lang_embs inside embed_prefix would still pass every assertion here. Pinning the call site (instantiate planner / shape-only stand-in, run embed_prefix, assert emitted att_masks) would close that gap.
nit — src/opentau/policies/pi07/gemma3_with_expert.py:597 (sdpa_attention_forward) — is_causal=False is correct given the precise bool mask, and the test pins fp32 eager↔SDPA equivalence to 1e-4. No bf16 equivalence test (the actual production dtype) — A100/H100 + bf16 path is unverified by CI. Not blocking; the math backend on CPU is fp32 either way, but a bf16 allclose with looser tolerance on a GPU runner would round out the coverage.

No blocking issues found.

claude

[claude-review] Spot review for commit 3c5ee90 — see PR-level summary comment for the full picture.

shuheng-liu · 2026-04-29T22:06:55Z

FYI: this will fix #171

akshay18iitg · 2026-04-29T22:54:00Z

@claude fix on all the agreed points

@claude

- addresses @claude[bot] (low-level is_pad defaults, low_level_planner modeling_pi07_low_level.py:984-986): switched the *_is_pad fallbacks to torch.ones to match the high-level planner — missing speed/quality/mistake no longer fabricate "Speed: 0.0" entries in the prompt. - addresses @claude[bot] (prepare_metadata docstrings): updated both planners' docstrings to reference Gemma 3 (not PaliGemma) and to enumerate robot_type / control_mode (string-valued, empty-string-as-pad). - addresses @claude[bot] (high-level all-empty guard, high_level_planner/modeling_pi07_high_level.py:715): mirrored the low-level "if segments else ''" guard so an all-padded sample emits "" instead of the literal "Metadata: ". Both planners now agree. - addresses @claude[bot] (CPU coverage for prepare_metadata): added TestPrepareMetadataSegments to tests/policies/test_pi07_cpu.py covering (a) robot+control populated, (b) both absent, (c) one populated and one empty, (d) per-sample all-empty emits "" (regression for both planners), and (e) low-level missing speed/quality/mistake never produces a fabricated "Speed: 0.0" segment. tests: passed — pytest -m "not gpu" tests/policies/test_pi07_cpu.py Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

shuheng-liu · 2026-04-30T23:04:59Z

FYI, the failed CI is fixable my merging latest main into this branch.

claude

[claude-review] Spot review for commit 2949e73 — see PR-level summary comment for the full picture.

shuheng-liu · 2026-05-02T03:43:39Z

@claude fix the remaining items in the review.

@claude

- addresses @claude[bot] (subgoal opt-in docstring): updated _load_subgoal_frames docstring to match the always-on behaviour; replaced the stale test_missing_subgoals_key_in_info_returns_empty test (which encoded the old opt-in gate) with two pinned cases — no-cameras → {}, and "no info.subgoals key still loads subgoals". Added camera_keys / image_keys / episode_data_index attrs to the SimpleNamespace meta in the two existing video tests so they match the new attribute reads. - addresses @claude[bot] (image-dtype fallback row index): new test_image_dtype_fallback_uses_absolute_row_index in tests/datasets/ test_optional_keys.py stubs hf_dataset.__getitem__ and pins that the parquet-row lookup uses ep_start + subgoal_frame, never the within-episode index. - addresses @claude[bot] ("fully 0 padded" comment): rewrote the comment at modeling_pi07_low_level.py:946 to call out -1 ([-1, 1] SigLIP range) and the False-mask role of the placeholder. - addresses @claude[bot] (_action_indicator_len recompute): cached the Action-indicator length at PI07LowLevelPlannerFlowMatching.__init__; both forward sites now read self._action_indicator_len. - addresses @claude[bot] (embed_prefix CPU coverage): added TestEmbedPrefixConditionalGuards in tests/policies/test_pi07_cpu.py with a fake Gemma3WithExpert + tokenizer + state_proj + embed_video so the three guards (response_masks.any(), subgoal availability, metadata_masks.any()) are exercised without GPU. Cases: all-False optional masks → no spurious causal boundary; mixed-availability subgoal batch → header/footer pad mask zeroes the pad-only sample; response.any() → exactly one boundary. - addresses @claude[bot] (.base_layer. stale comment in video_encoder.py): rewrote the wrap comment at video_encoder.py:422 to explain that the wrapper adopts submodules by reference, so wrapped-layer state-dict keys are byte-for-byte identical to a vanilla SiglipEncoderLayer (no .base_layer. prefix). tests: passed — pytest tests/policies tests/datasets -m "not gpu" -n auto (452 passed, 12 skipped; the 1 collection error in tests/policies/test_pi07_paligemma_low_level_planner.py is pre-existing and unrelated — pi07_paligemma still imports VJEPA2VideoEncoder from pi05_mem, which was removed in #171). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

#229)

claude

[claude-review] Spot review for commit 62a5677 — see PR-level summary comment for the full picture.

shuheng-liu · 2026-05-02T06:57:46Z

@claude fix according to the new review

@claude

- addresses @claude (LL planner doc): added a docstring note in embed_prefix that has_any_optional is computed batch-wide so per-sample optional layout follows the batch-level OR (kept intentionally for prefix-length uniformity). - addresses @claude (HL planner gate): tightened embed_prefix metadata gate from `metadata_tokens is not None` to `metadata_masks.any()` to match the low-level planner's semantics; lets training cleanly drop both the metadata block and the trailing ";\n " prefix-end when an entire batch is metadata-free. - addresses @claude (GPU regression test): added test_no_optionals_path_on_real_gemma3 in TestPI07LowLevelPlannerIntegration that exercises empty response + all-padded metadata + subgoal_is_pad on the real Gemma 3 backbone, pinning the collapsed prefix length and the position of the first causal boundary. - addresses @claude (test_pi06 stats reshape): replaced `np.full((chunk_size, 32), float(stats.flatten()[0]))` with `np.tile(stats[:1], (config.chunk_size, 1))` so per-feature variance in future fixtures is preserved instead of collapsed to a scalar. tests: passed -- python -m pytest -m "not gpu" tests/policies/test_pi07_cpu.py tests/policies/test_pi07_video_encoder_cpu.py tests/policies/test_pi06.py Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

claude

[claude-review] Spot review for commit e24ba88 — see PR-level summary comment for the full picture.

…pi07) (#242)

shuheng-liu · 2026-05-04T06:09:21Z

To save Claude context, I'll open a fresh PR.

Brings 16 main commits onto feat/pi07 in preparation for opening the feat/pi07 → main PR. One conflict (tests/policies/test_pi06.py: np.tile vs np.full for chunk-size action stats) — kept HEAD's np.tile fix from claude review feedback on #197, which preserves per-feature variance instead of collapsing every step to stats.flatten()[0]. Also dropped stale --ignore=tests/policies/test_pi07_paligemma_low_level_planner.py from .github/workflows/cpu_test.yml and gpu_test.yml: main deleted the file in #234, but the --ignore line on feat/pi07 (added in #229) was on a different part of the file so git auto-merged without flagging it. Removing both now keeps the workflows clean.

Adding original pi07 with gemma3 backbone and space-time siglip video…

c0cc4e9

… encoder

akshay18iitg self-assigned this Apr 29, 2026

akshay18iitg linked an issue Apr 29, 2026 that may be closed by this pull request

V-JEPA2 video encoder receives unnormalized [0,1] pixels — expects ImageNet normalization #170

Closed

akshay18iitg added feature New feature or request model new model or model request and removed feature New feature or request labels Apr 29, 2026

shuheng-liu mentioned this pull request Apr 29, 2026

fix(pi07): align with #178/#171 invariants & restore CPU tests #198

Merged

3 tasks

shuheng-liu and others added 2 commits April 29, 2026 11:45

fix(pi07): align with #178/#171 invariants & restore CPU tests (#198)

e135cb6

Merge main

90db9e2

claude Bot reviewed Apr 29, 2026

View reviewed changes

Comment thread src/opentau/policies/pi07/low_level_planner/video_encoder.py Outdated

Adding control_type and robot_type to metadata in policy code

3c5ee90

claude Bot reviewed Apr 29, 2026

View reviewed changes

Applying pi07_paligemma fixes to pi07

2949e73

claude Bot reviewed May 1, 2026

View reviewed changes

claude Bot mentioned this pull request May 2, 2026

fix(pi07_paligemma): byte-equivalent prefix to pi05 when optionals dropped (#226) #227

Merged

3 tasks

fix(pi07): gate optional prefix tokens in low- and high-level planners (

62a5677

#229)

claude Bot reviewed May 2, 2026

View reviewed changes

Comment thread src/opentau/policies/pi07/low_level_planner/modeling_pi07_low_level.py

Comment thread src/opentau/policies/pi07/high_level_planner/modeling_pi07_high_level.py Outdated

Comment thread tests/policies/test_pi07_low_level_planner.py

Comment thread tests/policies/test_pi06.py Outdated

shuheng-liu assigned shuheng-liu and unassigned akshay18iitg May 2, 2026

shuheng-liu added 2 commits May 2, 2026 01:11

feat(pi07): SDPA attention + gradient checkpointing options (#233)

6f256ec

fix(pi07): align text attention with paper §VI.B + Fig. 19 (#236)

e24ba88

shuheng-liu requested review from WilliamYue37 and shuheng-liu May 4, 2026 04:01

claude Bot reviewed May 4, 2026

View reviewed changes

fix(pi07): align text attention with paper §VI.B + Fig. 19 (apply to …

b65f972

…pi07) (#242)

shuheng-liu closed this May 4, 2026

shuheng-liu mentioned this pull request May 4, 2026

feat(pi07): π0.7 policy with Gemma 3 backbone and SpaceTime SigLIP video encoder #244

Merged

3 tasks

Conversation

akshay18iitg commented Apr 29, 2026

What this does

New files

Key architectural changes vs pi07_paligemma

How to checkout & try? (for the reviewer)

Run the high-level planner integration test

Run the low-level planner integration test

Run just the regression tests (faster)

Checklist

Note: Before submitting this PR, please read the contributor guideline.

Uh oh!

shuheng-liu commented Apr 29, 2026

Findings

Uh oh!

shuheng-liu commented Apr 29, 2026

Uh oh!

shuheng-liu commented Apr 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

shuheng-liu commented Apr 29, 2026

Uh oh!

claude Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

claude Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

claude Bot commented Apr 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

claude Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

shuheng-liu commented Apr 29, 2026

Uh oh!

akshay18iitg commented Apr 29, 2026

Uh oh!

shuheng-liu commented Apr 30, 2026

Uh oh!

claude Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

shuheng-liu commented May 2, 2026

Uh oh!

claude Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

shuheng-liu commented May 2, 2026

Uh oh!

claude Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

shuheng-liu commented May 4, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Key architectural changes vs `pi07_paligemma`

shuheng-liu commented Apr 29, 2026 •

edited

Loading

claude Bot left a comment •

edited

Loading

claude Bot commented Apr 29, 2026 •

edited

Loading