HF-BF16 (safetensors) export of the marin Grug 67B-A2B MoE model after SFT Stage 2 (thinking). Architecture GrugMoeForCausalLM (model_type: grug_moe); serve with the marin vLLM fork (--enable-expert-parallel), not upstream vLLM.

Total params: ~67B (A2B active MoE: 256 experts, 4 experts/token, 26 layers, hidden 2560, 20 heads / 5 KV heads).
Base: june-67b-a2b step-42150 (July 2T cooldown) -> SFT Stage 1 wildchat (penfever/grug-67b-a2b-sft-s1-wildchat-step257).
SFT Stage 2: laion/llama-nemotron-science-reasoning-on-canonical-think-full (thinking-on, canonical <|start_think|>...<|end_think|>), 1 packed epoch = 630 steps, chained (weights-only init) from Stage 1's step-257 checkpoint, seq_len 32768, global batch 64, optimizer AdamH (fp32), cut cross-entropy (batched_xla).
Tokenizer: marin-community/marin-tokenizer; chat template delphi_v0.jinja2.

Export provenance

Exported from the native Levanter/Orbax checkpoint via marin's sanctioned GrugModelConfig.hf_checkpoint_converter().with_config_overrides({"dtype":"bfloat16"}) path (experiments/grug/moe/model.py), reproducing tests/vllm/e2e/test_june_67b_a2b_hf_bf16_export.py. pending_qb_betas is baked into the router bias before export (required for correct logits). All tensors BF16.

Companion s3 export: s3://marin-us-east-02a/marin/exports/grug/june-67b-a2b-sft-s2-thinking/step-630/hf-bf16-vllm/

penfever/grug-67b-a2b-sft-s2-thinking-step630

Grug 67B-A2B SFT (Stage 2, Nemotron Science Reasoning / thinking) - step 630

Export provenance

penfever/
grug-67b-a2b-sft-s2-thinking-step630