Views
No views yet
This is a numerical W8A16 quantization of orcarouter/Qwen3.8-27B-Uncensored, an abliterated (refusal-removed) version of Qwen3.8-27B. All model credit belongs to Qwen and OrcaRouter; refer to the upstream model cards for architecture, capabilities, and usage guidance.
| Component | Precision | Reason |
|---|---|---|
| MLP projections | INT8 W8A16 | Largest dense GEMMs |
| Full-attention projections | INT8 W8A16 | Low output-distribution error |
GDN in_proj_qkv, in_proj_z, out_proj | INT8 W8A16 | ~4 GB memory recovery |
GDN in_proj_a, in_proj_b | BF16 | Tiny recurrent gates; precision safeguard |
| Vision tower | BF16 | Preserve multimodal fidelity |
lm_head | BF16 | Preserve final-logit fidelity |
| MTP head | BF16 | Keep speculative drafter close to target |
Norms, conv1d, A_log, dt_bias | BF16/FP32 | Non-Linear; never packed |
1# recipe.yaml
2default_stage:
3 default_modifiers:
4 QuantizationModifier:
5 targets: [Linear]
6 ignore:
7 - lm_head
8 - re:.*visual.*
9 - re:.*mtp.*
10 - re:.*linear_attn[.]in_proj_a$
11 - re:.*linear_attn[.]in_proj_b$
12 scheme: W8A16sm_86. They do not provide native FP8 tensor-core execution. W8A16 uses INT8 weights (Marlin kernel) with BF16 activations — the correct format for Ampere/Ada inference with predictable behavior and near-lossless quality.| Configuration | BF16 Source | This Quant |
|---|---|---|
| Weights (disk) | 55.6 GB | 29.4 GB |
| Loaded VRAM (1 GPU) | 56+ GB | ~29 GB |
| Minimum GPU | × H100 80 GB | 1× RTX 4090 / A6000 |
| Dual GPU | 2× 48 GB | 2× RTX 3090 24 GB |
1vllm serve morikomorizz/Qwen3.8-27B-Uncensored-INT8-W8A16-MTP \
2 --served-model-name qwen3.8-27b-uncensored-w8a16 \
3 --dtype bfloat16 \
4 --gpu-memory-utilization 0.92 \
5 --max-model-len 262144 \
6 --trust-remote-code \
7 --reasoning-parser qwen3 \
8 --tool-call-parser qwen3_coder \
9 --speculative-config '{"method":"mtp","num_speculative_tokens":3}'1export NCCL_P2P_DISABLE=1
2export VLLM_WORKER_MULTIPROC_METHOD=spawn
3
4vllm serve morikomorizz/Qwen3.8-27B-Uncensored-INT8-W8A16-MTP \
5 --tensor-parallel-size 2 \
6 --dtype bfloat16 \
7 --gpu-memory-utilization 0.92 \
8 --max-model-len 262144 \
9 --max-num-batched-tokens 8192 \
10 --kv-cache-dtype fp8_e4m3 \
11 --enable-chunked-prefill \
12 --trust-remote-code \
13 --reasoning-parser qwen3 \
14 --tool-call-parser qwen3_coder \
15 --speculative-config '{"method":"mtp","num_speculative_tokens":3}'1from openai import OpenAI
2client = OpenAI(base_url="http://localhost:8000/v1", api_key="none")
3resp = client.chat.completions.create(
4 model="qwen3.8-27b-uncensored-w8a16",
5 messages=[{"role": "user", "content": "Hello!"}],
6 max_tokens=512,
7)
8print(resp.choices[0].message.content)| Property | Value |
|---|---|
| Quantization | Data-free symmetric RTN W8A16, group size 128 |
| Runtime format | compressed-tensors / pack-quantized |
| Kernel dispatch | CompressedTensorsWNA16 → MarlinLinearKernel |
| Preserved precision | BF16 vision tower, lm_head, MTP, GDN gates |
| MTP | BF16 draft model; embeddings and lm_head shared with target |
| Runtime | vLLM; this is not a GGUF checkpoint |
| File | Purpose |
|---|---|
model-00001-of-00002.safetensors | Packed W8A16 language + GDN weights |
model-00002-of-00002.safetensors | Packed W8A16 + BF16 vision tower |
model_mtp.safetensors | BF16 MTP head, 15 tensors, ~0.79 GB |
model.safetensors.index.json | Shard-to-tensor mapping |
recipe.yaml | Exact llm-compressor W8A16 recipe |