Qwen3.8-27B-nvfp4full-NInfer

模型介绍

The repository contains a fuller-NVFP4 weight profile of Qwen/Qwen3.8-27B in the native NInfer `.ninfer` artifact format, with the z-lab/Qwen3.8-27B-DFlash2 block-diffusion speculative drafter embedded in the same image. It runs on NInfer **v3** engines — Text and MTP components through format/shape routes the upstream runtime already implements, the optional NVFP4-encoded DFlash2 companion through the fork's runtime support. It is not a Transformers checkpoint, Safetensors distribution, or GGUF file. Compared with the official `nvfp4` profile it extends NVFP4 from the layers 0–55 MLP to nearly the whole Text backbone: - every Text `mlp/gate_up` and `mlp/down` is NVFP4 (the layers 0–55 words are copied bit-exactly from unsloth/Qwen3.8-27B-NVFP4; layers 56–63 are quantized locally from the official BF16 checkpoint); - every GDN `query_key_value_z` is NVFP4, and GDN `output` is NVFP4 on 47 of 48 layers; - full-attention `query_key_gate_value` is NVFP4 on the ten deepest layers and `attention/output` on 14/16, with 9 BF16 exception parents retaining the registered Qwen3.6-27B NVFP4 exception pattern; - the token embedding and full output head use groupwise W8G32_F16S instead of row-scaled FP8. This yields 247 NVFP4 parents with 247 site-level input divisors. Locally quantized parents use the documented encoder profile `NVFP4_MAXABS_DIVISOR_RNE_V1` (per-tensor FP32 divisor, per-16 E4M3FN block scales, RNE E2M1 codes) with site divisors calibrated by streaming the official BF16 checkpoint. The complete contract, encoder, calibration corpus, and validation evidence are `docs/maintainer/qwen3.8-27b-artifact.md` §14 in the fork.

主要特点

特点摘要待依据来源补齐，不凭模型名称推断能力。

局限与注意事项

暂无足够来源支持具体缺点结论。请重点验证实际任务表现、时延、数据政策、许可与版本生命周期。

来源报价与费用分析

同一模型在不同渠道、地区和配置下可能采用不同价格。下表独立保留各渠道报价，不将渠道价格视为厂商官方统一价格。

公开评测与运行配置

暂无已采集的可核验公开评分。模型未上榜不代表能力较低。

常见问题

资料来源与可追溯记录

许可证：apache-2.0。模型介绍与第三方评测均以原来源的具体版本和方法为准。