Qwen3.5-122B-A10B-abliterated-NVFP4

NVFP4 (4-bit floating point) quantized derivative of wangzhang/Qwen3.5-122B-A10B-abliterated, which itself is derived from Qwen/Qwen3.5-122B-A10B.

This repository provides a modified derivative checkpoint for local inference and serving. The primary changes in this repository are NVFP4 quantization, weight repacking / export formatting, and serving compatibility adjustments.

Model Details

Property	Value
Intermediate Base Model	wangzhang/Qwen3.5-122B-A10B-abliterated
Original Base Model	Qwen/Qwen3.5-122B-A10B
Architecture	Qwen3.5 MoE (256 routed experts, 10B active)
Quantization	NVFP4 (compressed-tensors, nvfp4-pack-quantized)
Original Size	228 GB (BF16)
Quantized Size	71.2 GB (69% reduction)
Format	safetensors (2 shards)

Quantization Method

This model was quantized using a template-based weight replacement approach:

1. Reference Template: RedHatAI/Qwen3.5-122B-A10B-NVFP4 — a calibrated NVFP4 checkpoint of the original (non-abliterated) Qwen3.5-122B-A10B, produced by llm-compressor with proper calibration data.
2. Weight Replacement: Each quantized tensor (weight_packed and weight_scale) was regenerated from the abliterated BF16 weights using the reference checkpoint's weight_global_scale and input_global_scale values.
3. Format Preservation: The reference checkpoint's config.json, quantization_config, global scales, and all metadata were preserved unchanged, ensuring full compatibility with vLLM's CUTLASS NVFP4 MoE kernel.

What is Quantized

Component	Format	Notes
Routed experts (gate/up/down_proj)	NVFP4	256 experts × 48 layers × 3 projections
Shared experts	NVFP4	48 layers × 3 projections
Self-attention (q/k/v/o_proj)	NVFP4	12 full-attention layers
Linear attention	BF16	36 layers, kept at full precision
Embeddings, norms, gates	BF16	Kept at full precision

Serving with vLLM

This model requires a text-only compatibility patch for vLLM since Qwen3.5 MoE is a multimodal architecture but this checkpoint contains only text weights.

Quick Start

# 1. Download the model
huggingface-cli download bjk110/Qwen3.5-122B-A10B-abliterated-NVFP4

# 2. Apply the text-only patch before starting vLLM
python vllm_patches/patch_qwen35_moe_text.py

# 3. Serve with vLLM
vllm serve /path/to/model \
    --served-model-name Qwen3.5-122B-A10B-abliterated-NVFP4 \
    --max-model-len 131072 \
    --max-num-seqs 4 \
    --gpu-memory-utilization 0.90 \
    --trust-remote-code \
    --enable-prefix-caching \
    --enable-chunked-prefill \
    --reasoning-parser qwen3

### Docker Compose (Recommended)

A complete Docker Compose setup is provided in the `serving/` directory:

```bash
# Copy serving files
cp -r serving/ /path/to/your/vllm-setup/

# Edit .env to set MODEL_PATH
vim serving/.env

# Start
cd serving && docker compose --profile head up -d

See serving/ directory for:

• docker-compose.yml — Full vLLM serving configuration
• .env.example — Environment variables template
• entrypoint.sh — Entrypoint with automatic patch application

Hardware Requirements

Configuration	Memory	max_model_len	Notes
1× NVIDIA DGX Spark (GB10)	121 GiB unified	131,072 (128K)	Tested and verified
1× GPU with 80+ GB VRAM	80 GiB	~65,536	Estimated

Performance (DGX Spark, TP=1)

Metric	Value
Throughput	14.5 tok/s average, 16.8 tok/s peak
KV Cache	222K tokens (20.4 GiB)
Max Concurrency	6.16× at 128K context
Model Loading	~13 min (2 shards)

Referenced Models

• Base model: wangzhang/Qwen3.5-122B-A10B-abliterated — Abliterated (uncensored) version of Qwen3.5-122B-A10B
• Original model: Qwen/Qwen3.5-122B-A10B — Official Qwen3.5 MoE model
• Quantization template: RedHatAI/Qwen3.5-122B-A10B-NVFP4 — Used as format reference for NVFP4 quantization structure
• FP8 variant: bjk110/Qwen3.5-122B-A10B-abliterated-FP8 — FP8 block-wise quantized version (116 GB, requires TP=2)