DeepSeek-V4-Flash-0731-bf16

A bfloat16 re-export of deepseek-ai/DeepSeek-V4-Flash-0731. The weights are the published ones, dequantized: the dense linears from fp8 (e4m3 with ue8m0 scales over 128x128 blocks) and the routed experts from MXFP4, both to plain bfloat16. 43 layers, 114 safetensors shards. The upstream `config.json` carries `quantization_config` and `expert_dtype = "fp4"`. An inference engine reading those fields allocates fp8 linear parameters and MXFP4 expert parameters, which bfloat16 weights cannot fill. The routed experts at least fail loudly, in vLLM's `_load_w13`, where the packed MXFP4 shape does not match. The fp8 linears fail silently: a bfloat16 tensor has the same shape as the fp8 parameter it is loaded into, so `copy_` casts it elementwise and no dequantization scale is ever applied. Serving bfloat16 weights therefore requires a checkpoint whose config carries neither field, which is what this repo is. There is no Jinja chat template, by upstream design. Use the bundled `encoding/` directory, which is copied verbatim from upstream and is the reference implementation for turning OpenAI-format messages into an input string and for parsing completions back.

译文尚未就绪，显示来源原文