Is it actually possible to train a VLM with multiple image inputs using GRPOTrainer?

[Randomdude11]

According to the documentation, it is possible by having an 'images' key in the datapoints, containing a list of images. However, when I attempt this, on accelerate I get the error

[rank2]: Parameter indices which did not receive grad for rank 2: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 ...

If I try to run it without accelerate, training does start, but there are no images passed to the input whatsoever. Has anybody managed to successfully run this use case?

[weathon]

+1

[Randomdude11]

It is possible, the latest commit which fixed it was around 1 week ago. You need to pull the main branch, it’s not in PyPI’s latest release yet. Then it works as in the documentation.