GLM-5#
GLM-5, GLM-5.1, and GLM-5.2 are large sparse MoE language models with Multi-Latent Attention and Dynamic Sparse Attention. Megatron Bridge supports these checkpoints through the shared GLM5Bridge.
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
A CPU import of the pinned Hugging Face revision must complete every mapping and create a reloadable Megatron checkpoint. This workflow is deferred by this card.
Import · GPU
✓ VerifiedExact command
./scripts/conversion/convert.sh import --executor slurm --device gpu --nodes 4 --gpus-per-node 8 --hf-model zai-org/GLM-5 --hf-revision 4e6698ba8e85059d749020e3c4d2123719f23926 --megatron-path work/model-verification/glm5/gpu-megatron --torch-dtype bfloat16 --tp 1 --pp 2 --ep 8 --etp 2 --distributed-timeout-minutes 60 --low-memory-saveExpected result
The pinned 32-H100 import exits successfully at TP1/PP2/EP8/ETP2, completes all 6,201 distributed mapping tasks, and persists a reloadable iter_0000000 checkpoint. Reload plus exact export projection covers all 59,079 tensors and 1,487,822,475,264 tensor-payload bytes in the 78-layer inference graph with matching keys, shapes, dtypes, and values. The 791 source tensors under model.layers.78 belong only to the intentionally disabled appended MTP auxiliary layer and are outside this item.
Export · CPU
✓ VerifiedExact command
./scripts/conversion/convert.sh export --executor slurm --device cpu --nodes 4 --cpu-processes-per-node 8 --cpus-per-task 16 --mem 0 --exclusive --hf-model zai-org/GLM-5 --hf-revision 4e6698ba8e85059d749020e3c4d2123719f23926 --megatron-path work/model-verification/glm5/gpu-megatron/iter_0000000 --hf-path work/model-verification/glm5/cpu-hf-export --torch-dtype bfloat16 --tp 1 --pp 2 --ep 8 --etp 2 --distributed-timeout-minutes 240 --distributed-save --save-every-n-ranks 1 --no-progressExpected result
The 32-process distributed CPU export exits successfully and writes 280 safetensors shards. The exhaustive projection audit covers all 59,079 tensors and 1,487,822,475,264 tensor-payload bytes in the 78-layer inference graph with zero missing, unexpected, shape, dtype, or value mismatches. The 791 tensors under model.layers.78 belong only to the intentionally disabled appended MTP auxiliary layer and remain outside this item. Transformers 5.12.1 strictly reloads the output as GlmMoeDsaForCausalLM with 1,629 state tensors, 743,911,199,232 parameters, and no loading discrepancies.
Export · GPU
✓ VerifiedExact command
./scripts/conversion/convert.sh export --executor slurm --device gpu --nodes 4 --gpus-per-node 8 --hf-model zai-org/GLM-5 --hf-revision 4e6698ba8e85059d749020e3c4d2123719f23926 --megatron-path work/model-verification/glm5/gpu-megatron/iter_0000000 --hf-path work/model-verification/glm5/gpu-hf-export --torch-dtype bfloat16 --tp 1 --pp 2 --ep 8 --etp 2 --distributed-timeout-minutes 60 --distributed-save --save-every-n-ranks 1Expected result
The 32-H100 distributed export exits successfully after all 6,201 mapping tasks and writes 280 safetensors shards. Of those, 278 are byte-for-byte identical to the pinned source shards; the two shards that also contain excluded MTP keys match all 63 common tensors exactly. The composite audit covers all 59,079 inference-graph tensors and 1,487,822,475,264 tensor-payload bytes with zero key, shape, dtype, or value mismatch. Transformers 5.12.1 strictly reloads the output as GlmMoeDsaForCausalLM with 1,629 state tensors, 743,911,199,232 parameters, and no loading discrepancies.
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
A public GLM-5 H100 recipe must complete a bounded 100-step run with finite loss, no skipped or NaN iterations, all five metrics, and a reloadable final checkpoint. Training is deferred by this card.
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
A pinned-data 100-step full-SFT run must finish with finite loss, no skipped or NaN iterations, all five metrics, and a reloadable final checkpoint. Training is deferred by this card.
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
A dedicated packed long-context SFT run must complete a bounded run with finite loss, no skipped or NaN iterations, all five metrics, and a reloadable checkpoint. Training is deferred by this card.
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
A GLM-5 PEFT recipe with an audited adapter target set must complete a bounded run with finite loss, all five metrics, and a reloadable adapter checkpoint. Training is deferred by this card.