allenai / allenai/molmo2

Video Evaluation with max_frames=384 Failed

Đang mở
#20 0 bình luận 0 reaction 0 người được giao Xem trên GitHub
Ngôn ngữ chính
Python
Star
726
Fork
48
Chỉ số merge pull request
Không có pull request nào được merge trong 30 ngày

Mô tả

When perform evaluation on video benchmarks (like LongVideoBench), an error occurs if setting --model.mm_preprocessor.video.max_frames=384.
Details:

[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> Traceback (most recent call last):
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/molmo2/launch_scripts/eval_molmo2.py", line 317, in
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> main()
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/molmo2/launch_scripts/eval_molmo2.py", line 312, in main
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> cfg.build().run()
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/molmo2/olmo/eval/model_evaluator.py", line 439, in run
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> metrics = evaluator.run(
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^^^^^^^^^^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/molmo2/olmo/eval/inf_evaluator.py", line 398, in run
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> olmo_gen_output = model.generate(
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^^^^^^^^^^^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/molmo2/olmo/models/molmo2/molmo2.py", line 998, in generate
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> token_ids, scores = beam_search.search(initial_preds, state, step)
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/molmo2/olmo/nn/beam_search.py", line 1063, in search
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> return self._search(start_predictions, start_state, new_step)
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/molmo2/olmo/nn/beam_search.py", line 1093, in _search
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> start_class_log_probabilities, state = step(start_predictions, start_state, 0)
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/molmo2/olmo/nn/beam_search.py", line 1061, in new_step
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> return old_step(last_predictions, state)
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/molmo2/olmo/models/molmo2/molmo2.py", line 969, in step
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> output = self(
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/miniconda3/envs/molmo/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1773, in _wrapped_call_impl
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> return self._call_impl(*args, **kwargs)
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/miniconda3/envs/molmo/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1784, in _call_impl
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> return forward_call(*args, **kwargs)
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/molmo2/olmo/models/molmo2/molmo2.py", line 765, in forward
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> x, cache = block(
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/miniconda3/envs/molmo/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1773, in _wrapped_call_impl
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> return self._call_impl(*args, **kwargs)
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/miniconda3/envs/molmo/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1784, in _call_impl
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> return forward_call(*args, **kwargs)
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/molmo2/olmo/nn/llm.py", line 2271, in forward
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> att, cache = self.attention(
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^^^^^^^^^^^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/molmo2/olmo/nn/llm.py", line 1934, in attention
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> q, k = self.rotary_emb(q, k, position_ids=position_ids, cp_enabled=False)
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/miniconda3/envs/molmo/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1773, in _wrapped_call_impl
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> return self._call_impl(*args, **kwargs)
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/miniconda3/envs/molmo/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1784, in _call_impl
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> return forward_call(*args, **kwargs)
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/molmo2/olmo/nn/llm.py", line 1454, in forward
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> q_ = self.apply_rotary_pos_emb(
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^^^^^^^^^^^^^^^^^^^^^^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/molmo2/olmo/nn/llm.py", line 1339, in apply_rotary_pos_emb
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> return ((t * pos_cos) + (self.rotate_half(t) * pos_sin)).to(t.dtype)
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^^^^^^^^^^^^^^^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> File "/molmo2/olmo/nn/llm.py", line 1329, in rotate_half
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> return torch.cat((-x2, x1), dim=-1)
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> ^^^
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> torch.AcceleratorError: CUDA error: device-side assert triggered
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> For debugging consider passing CUDA_LAUNCH_BLOCKING=1
[2026-04-24T16:27:11+08:00] molmo-longvideobench-eval-4b-19629094-b5c76-rmw8k >> Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

Hướng dẫn đóng góp

Chưa lập chỉ mục được hướng dẫn đóng góp cho kho mã nguồn này

Đánh giá

Issue này chưa được đánh giá.

Nhận issue mới trong hộp thư của bạn

Bản tóm tắt ngắn những issue GitHub phù hợp với người mới.