Gemma 3 VL (Vision-Language)#
Google’s Gemma 3 VL is a family of vision-language models built on the same research and technology used to create Gemini models. The Gemma 3 VL architecture combines the text-generation capabilities of Gemma 3 with a SigLIP vision encoder for robust visual understanding.
Gemma 3 VL models support multimodal tasks including image captioning, visual question answering, OCR, and general vision-language understanding.
Gemma family models are supported via the Bridge system with auto-detected configuration and weight mapping.
Verified configurations#
Choose an exact recorded configuration to see its command and expected result. These selectors are generated from the authoritative verification cards and never synthesize combinations.
Run a configuration#
Choose a workflow, precision, and exact recorded combination. The command and expected result update below.
Import · CPU
✓ VerifiedExact command
./scripts/conversion/convert.sh import --executor slurm --device cpu --nodes 1 --hf-model google/gemma-3-4b-it --hf-revision 093f9f388b31de276ce2de164bdc2081324b9767 --megatron-path work/model-verification/gemma-3-4b-it/cpu-megatron --torch-dtype bfloat16Expected result
Exact-revision BF16 CPU import loaded all 781 Bridge tensors and saved one reloadable DCP shard. An independent Gloo reload mapped all 883 source tensors with exact keys, shapes, dtypes, and torch.equal values, including all 437 legacy vision-tower keys.
Import · GPU
✓ VerifiedExact command
./scripts/conversion/convert.sh import --executor slurm --device gpu --nodes 1 --gpus-per-node 4 --hf-model google/gemma-3-4b-it --hf-revision 093f9f388b31de276ce2de164bdc2081324b9767 --megatron-path work/model-verification/gemma-3-4b-it/imported-megatron --torch-dtype bfloat16 --tp 2 --pp 2 --ep 1 --etp 1Expected result
Exact-revision distributed BF16 import completed with TP2/PP2/EP1/ETP1. All 781 Bridge tensors loaded and all four ranks saved a reloadable checkpoint with the vision tower and multimodal projector on the owning pipeline stage. Independent TP2/PP2 and Gloo reloads retained FusedAttention and matched all 883 mapped tensors exactly.
Export · CPU
✓ VerifiedExact command
./scripts/conversion/convert.sh export --executor slurm --device cpu --nodes 1 --hf-model google/gemma-3-4b-it --hf-revision 093f9f388b31de276ce2de164bdc2081324b9767 --megatron-path work/model-verification/gemma-3-4b-it/cpu-megatron/iter_0000000 --hf-path work/model-verification/gemma-3-4b-it/cpu-hf-export --torch-dtype bfloat16Expected result
The exact-revision CPU export exits successfully and writes all 883 tensors, 4,300,079,472 values, and 8,600,158,944 tensor-payload bytes. Keys, shapes, dtypes, and values are bitwise identical to the pinned BF16 source with no dtype conversions, including all 437 legacy vision_tower.vision_model.* tensors. Transformers strictly reloads the output natively as Gemma3ForConditionalGeneration with all 884 state entries and no loading discrepancies.
Export · GPU
✓ VerifiedExact command
./scripts/conversion/convert.sh export --executor slurm --device gpu --nodes 1 --gpus-per-node 4 --hf-model google/gemma-3-4b-it --hf-revision 093f9f388b31de276ce2de164bdc2081324b9767 --megatron-path work/model-verification/gemma-3-4b-it/imported-megatron/iter_0000000 --hf-path work/model-verification/gemma-3-4b-it/hf-export --torch-dtype bfloat16 --export-weight-dtype bfloat16 --distributed-save --tp 2 --pp 2 --ep 1 --etp 1Expected result
TP2/PP2 distributed export completed without a missing PP owner. Source and export contained the same 883 keys with exact shapes, dtypes, and torch.equal values, including all 437 legacy vision_tower.vision_model.* tensors. Transformers exposes 884 logical state entries after reload because tied lm_head.weight aliases the single stored language_model.model.embed_tokens.weight tensor; no extra tensor is serialized. Native Gemma3ForConditionalGeneration strict reload reported no missing, unexpected, mismatched, or errored tensors.
Pretrain
× UnsupportedRecorded metrics
- Initial loss
- None
- Final loss
- None
- Step time · last 10 avg
- None ms
- Model throughput · last 10 avg
- None TFLOP/s/GPU
- Token throughput · last 10 avg
- None tokens/s/GPU
Exact command
No runnable command is recorded for this status.
Expected result
Megatron Bridge does not provide a public Gemma 3 VL 4B pretraining recipe. The available exact-model recipes cover SFT and PEFT, so a supported pretraining command and checkpoint contract cannot be stated.
SFT · H100
✓ VerifiedRecorded metrics
- Initial loss
- 1.16385
- Final loss
- 0.143927
- Step time · last 10 avg
- 5,461.010 ms
- Model throughput · last 10 avg
- 22.450 TFLOP/s/GPU
- Token throughput · last 10 avg
- 3,000.178 tokens/s/GPU
Exact command
./scripts/training/train.sh --nodes 1 --gpus-per-node 8 --recipe gemma3_vl_4b_sft_config --mode sft --dataset cord-v2 --pretrained_checkpoint work/model-verification/gemma-3-4b-it/imported-megatron/iter_0000000 --max_steps 10 --seq_length 4096 'dataset.source.load_kwargs={revision:"7f0115a4b758a71d6473b8d085751692da2fef98"}' validation.eval_iters=0 validation.eval_interval=0 checkpoint.load=null --save_dir work/model-verification/gemma-3-4b-it/sft-checkpoints --save_interval 10 logger.log_interval=1 logger.log_throughput=trueExpected result
On 8 H100s, the TP1/PP1 recipe completes exactly ten exact-revision CORD-v2 full-SFT steps at global batch size 32 and sequence length 4096. Loss is finite from 1.163850 to 0.143927 with no skipped or NaN iterations. The ten-step metric window includes the first-step cold-start cost by policy, and the final iteration-10 checkpoint saves successfully.
Long Context · H100
× UnsupportedRecorded metrics
- Initial loss
- None
- Final loss
- None
- Step time · last 10 avg
- None ms
- Model throughput · last 10 avg
- None TFLOP/s/GPU
- Token throughput · last 10 avg
- None tokens/s/GPU
Exact command
No runnable command is recorded for this status.
Expected result
Gemma 3 VL image-bidirectional attention uses dense local/global additive biases with FusedAttention. The provider rejects context parallelism greater than one, and the public recipes disable in-batch THD packing, so there is no supported exact-model long-context SFT contract.
LoRA · H100
â—‹ UnverifiedRecorded metrics
- Initial loss
- None
- Final loss
- None
- Step time · last 10 avg
- None ms
- Model throughput · last 10 avg
- None TFLOP/s/GPU
- Token throughput · last 10 avg
- None tokens/s/GPU
Exact command
./scripts/training/train.sh --nodes 1 --gpus-per-node 1 --recipe gemma3_vl_4b_peft_config --mode lora --dataset cord-v2 --pretrained_checkpoint work/model-verification/gemma-3-4b-it/imported-megatron/iter_0000000 --max_steps 10 --seq_length 4096 'dataset.source.load_kwargs={revision:"7f0115a4b758a71d6473b8d085751692da2fef98"}' validation.eval_iters=0 validation.eval_interval=0 checkpoint.load=null --save_dir work/model-verification/gemma-3-4b-it/peft-checkpoints --save_interval 10 logger.log_interval=1 logger.log_throughput=trueExpected result
Pending exact-revision CORD-v2 LoRA verification with the public Gemma 3 VL 4B H100 recipe. The run must complete at least ten optimizer steps with finite loss, no skipped or NaN iterations, all five metrics, and a reloadable adapter checkpoint.