CoolFace
Modelpublic

YuYu1015/Huihui-Gemma-4-E2B-it-abliterated-NVFP4

sourceHugging Facegemmaupdated 6mo agoView on Hugging Face
0likes24downloads
Model Card

Huihui-gemma-4-E2B-it-abliterated-NVFP4

English | 繁體中文


English

NVFP4 quantization of huihui-ai/Huihui-gemma-4-E2B-it-abliterated, quantized using NVIDIA ModelOpt with NVFP4_MLP_ONLY strategy (only MLP layers quantized, attention preserved in higher precision).

Model Details

ItemValue
ArchitectureDense, Per-Layer Embeddings (PLE), ~2.3B effective parameters
Base modelgoogle/gemma-4-E2B-it
Fine-tuned byhuihui-ai (abliteration)
Quantized byYuYu1015
Model size~7.4 GB (NVFP4)
Context lengthUp to 128,000 tokens
MultimodalVision + Audio supported

Quantization Details

ItemValue
MethodNVIDIA ModelOpt v0.42.0
SchemeNVFP4 (E2M1 + FP8 per-group scaling, group size 16)
StrategyNVFP4_MLP_ONLY — only MLP/FFN layers quantized, all attention layers preserved
Calibration datasetabisee/cnn_dailymail
Calibration samples512
HardwareNVIDIA DGX Spark (GB10, 128GB unified memory)

Layers Preserved in Higher Precision

LayerReason
self_attn.* (all layers)Attention layers preserved for accuracy (MLP_ONLY strategy)
lm_headOutput head
vision_tower.*Vision encoder
audio_tower.*Audio encoder
multi_modal_projector.*Multimodal projection
embed_tokensInput embeddings

Serving with vLLM

bash
vllm serve /path/to/model \
    --quantization modelopt \
    --served-model-name gemma-4-e2b \
    --trust-remote-code \
    --gpu-memory-utilization 0.90 \
    --max-model-len 32768 \
    --enable-prefix-caching \
    --enable-chunked-prefill \
    --language-model-only

DGX Spark (SM121) Compatibility Notes

  • —NVFP4 on SM121 falls back to W4A16 (native W4A4 path not available, missing cvt.e2m1x2 instruction)
  • —Use --quantization modelopt (not compressed-tensors)
  • —--language-model-only skips vision/audio encoder profiling for text-only inference
  • —Clear page cache before starting on UMA: sudo sh -c 'echo 3 > /proc/sys/vm/drop_caches'

Safety Warning

This model has safety filtering removed (abliterated) and may generate inappropriate content. Users are solely responsible for all consequences arising from its use.

Credits


繁體中文

huihui-ai/Huihui-gemma-4-E2B-it-abliterated 的 NVFP4 量化版本,使用 NVIDIA ModelOpt 的 NVFP4_MLP_ONLY 策略量化(僅量化 MLP 層,attention 保留高精度)。

模型資訊

項目數值
架構Dense,Per-Layer Embeddings (PLE),約 2.3B 有效參數
基礎模型google/gemma-4-E2B-it
微調者huihui-ai(abliteration)
量化者YuYu1015
模型大小~7.4 GB(NVFP4)
Context 長度最高 128,000 tokens
多模態支援視覺 + 音訊

量化詳情

項目數值
方法NVIDIA ModelOpt v0.42.0
方案NVFP4(E2M1 + FP8 逐群縮放,群組大小 16)
策略NVFP4_MLP_ONLY — 僅量化 MLP/FFN 層,所有 attention 層保留高精度
校準資料集abisee/cnn_dailymail
校準樣本數512
量化硬體NVIDIA DGX Spark(GB10, 128GB 統一記憶體)

保留高精度的層

層原因
self_attn.*(所有層)Attention 層保留以確保精度(MLP_ONLY 策略)
lm_head輸出頭
vision_tower.*視覺編碼器
audio_tower.*音訊編碼器
multi_modal_projector.*多模態投影層
embed_tokens輸入嵌入

DGX Spark (SM121) 相容性說明

  • —NVFP4 在 SM121 上會退回 W4A16(原生 W4A4 路徑不可用,缺少 cvt.e2m1x2 指令)
  • —使用 --quantization modelopt(非 compressed-tensors)
  • —--language-model-only 跳過視覺/音訊編碼器 profiling,加速純文字推理
  • —UMA 架構啟動前請先清除 page cache:sudo sh -c 'echo 3 > /proc/sys/vm/drop_caches'

安全警告

此模型已移除安全過濾機制(abliterated),可能產生不當內容。使用者須自行承擔所有風險與法律責任。

致謝