| name: fp8_dynamic_moe | |
| scheme: FP8_DYNAMIC # FP8 weights, FP8 dynamic per-token activations — data-free, vLLM-native | |
| engine: llmcompressor | |
| # MoE variant of fp8_dynamic.yaml for qwen3_5_moe bases (Ornith-1.0-35B, Qwen3.6-35B-A3B). | |
| # FP8_DYNAMIC needs no calibration data; this tiny placeholder is ignored by DataFreePipeline. | |
| calibration: | |
| dataset: neuralmagic/calibration | |
| config: LLM | |
| split: train | |
| num_samples: 4 | |
| max_seq_length: 512 | |
| ignore: | |
| - lm_head | |
| - "re:.*visual.*" | |
| - "re:.*linear_attn.*" # Mamba/SSM block stays in bf16 (landmine 5) | |
| - "re:.*mtp.*" # no-op on Ornith (base ships 0 mtp.*), harmless | |
| - "re:.*mlp.gate$" # MoE router — keep bf16 (MoE-only, landmine 24) | |
| - "re:.*mlp.shared_expert_gate$" # shared-expert gate — keep bf16 | |
| # The 256 routed mlp.experts.N.* and the mlp.shared_expert.* FFNs ARE quantised. | |
| export: | |
| save_compressed: true | |