EXAONE#
Megatron Bridge supports Hugging Face conversion for EXAONE 4.0 dense language models, EXAONE 4.5 vision-language models, and K-EXAONE MoE language models. Checked-in H100 recipes cover EXAONE 4.0 1.2B, EXAONE 4.5 VL 33B, K-EXAONE-236B-A23B, and K-EXAONE 2.0 750B-A37B.
Use the model-specific examples for commands and supported parallelism:
The EXAONE 4.0 and EXAONE 4.5 examples use the shared
scripts/conversion/convert.sh import entry point. The K-EXAONE examples use
their model-specific distributed round-trip workflows. K-EXAONE 2.0 also has a checked-in
model verification card
that records the current verification status of individual workflows.
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
â—‹ UnverifiedExact command
No runnable command is recorded for this status.
Expected result
No CPU Hugging Face-to-Megatron conversion result is claimed. Future verification must import the complete immutable-revision checkpoint, persist iter_0000000, reload it, and audit all mapped keys, shapes, dtypes, and values.
Import · GPU
✓ VerifiedExact command
./scripts/conversion/convert.sh import --executor slurm --device gpu --nodes 6 --gpus-per-node 8 --hf-model LGAI-EXAONE/K-EXAONE-2.0-750B-A37B --hf-revision 4bb394fd12f57be7174be7c80c30cc05472bf9e1 --megatron-path work/model-verification/k-exaone-2/imported-megatron --torch-dtype bfloat16 --tp 1 --pp 3 --ep 16 --etp 1 --distributed-timeout-minutes 180 --low-memory-saveExpected result
The pinned 59,396-tensor BF16 checkpoint imported and persisted iter_0000000 on 48 H100 GPUs, and that artifact subsequently reloaded for a full distributed GPU re-export audit. The audit compared all 59,396 tensors and 1,498,715,010,176 payload bytes with zero missing, unexpected, duplicate, shape, dtype, or value mismatches.
Export · CPU
â—‹ UnverifiedExact command
No runnable command is recorded for this status.
Expected result
No CPU Megatron-to-Hugging-Face conversion result is claimed. Future verification depends on a complete CPU import and must export, strictly reload, and compare all lossless tensors against the pinned source revision.
Export · GPU
â—‹ UnverifiedExact command
./scripts/conversion/convert.sh export --executor slurm --device gpu --nodes 6 --gpus-per-node 8 --hf-model LGAI-EXAONE/K-EXAONE-2.0-750B-A37B --hf-revision 4bb394fd12f57be7174be7c80c30cc05472bf9e1 --megatron-path work/model-verification/k-exaone-2/imported-megatron/iter_0000000 --hf-path work/model-verification/k-exaone-2/hf-export --torch-dtype bfloat16 --tp 1 --pp 3 --ep 16 --etp 1 --distributed-timeout-minutes 180 --distributed-save --save-every-n-ranks 1 --no-progressExpected result
A complete distributed export produced 59,396 indexed keys in 17 safetensors shards. Every exported tensor matches the pinned source bitwise across 1,498,715,008,512 payload bytes, and the corrected config/tokenizer reload successfully. The run exposed and fixed preservation of one physical repeated MTP layer across four speculative steps. The config fix is committed; this item remains unverified until the clean-commit export and exact audit rerun complete.
Pretrain · 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
No runnable command is recorded for this status.
Expected result
The exported 512-GPU H100 recipe must complete a bounded public-data pretraining run at TP4/PP16/EP32 with natural routing, finite loss, no skipped or NaN iterations, all five required metrics, and reloadable middle and final checkpoints.
SFT · 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
No runnable command is recorded for this status.
Expected result
The exported 512-GPU H100 full-SFT recipe must load the pinned checkpoint and complete a bounded immutable-dataset run at TP4/PP16/EP32 with finite loss, no skipped or NaN iterations, all four required metrics, and a reloadable final full-model checkpoint.
Long Context · 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
No runnable command is recorded for this status.
Expected result
No long-context SFT result is claimed. Future verification must first resolve a supported long-context topology, then verify sequence packing and context parallelism together with finite loss, no skipped or NaN iterations, and all five required metrics.
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
No runnable command is recorded for this status.
Expected result
The exported 128-GPU H100 PEFT recipe must load the pinned checkpoint, verify that LoRA targets the model-native attention projections, and complete a bounded immutable-dataset run at TP4/PP8/EP16 with finite loss, no skipped or NaN iterations, all five required metrics, and a reloadable adapter checkpoint.