m8than/DeepSeek-R1-0528-E2M1

DeepSeek-R1-0528 E2M1

Furnace gfx942 conversion of deepseek-ai/DeepSeek-R1-0528 at revision 4236a6af538feda4548eca9ab308586007567f52.

Storage

• Attention, dense MLP, and shared-expert linear weights retain the official FP8 E4M3 bytes with FP32 weight_scale_inv block scales.
• Routed experts, including the NextN layer, use canonical low-nibble-first E2M1: U32 weight tensors and FP32 weight_scale tensors.
• Weight blocks are 128×128. quantization_config.quant_method is moe_e2m1, and dense_weight_format is fp8_block128.

Runtime

Requires Furnace SGLang with dense-FP8 E2M1 support, AITER/CK kernels, and AMD gfx942 GPUs. Activations use the E2M1/FP8 block-128 path and outputs are BF16. This is not native MXFP4 and is incompatible with loaders that expect MXFP4 group-32 scales.

Conversion verification

The completed manifest reports status=complete and servable=true. Every routed tensor, including NextN, was checked for canonical U32/F32 dtype and shape. All 496 dense/shared FP8 weights and all 496 FP32 scales were hash-verified as byte-identical to the source, including cross-shard scale pairs.

E2M1 routed-expert conversion is lossy. No accuracy benchmark or inference run was performed while the GPUs were occupied by another workload.