Downloads · 30 days
38
24% of all-time downloads
Danny-Dasilva/Qwen3-4B-antidoom-W4A16-GPTQ-g32-DSpark
Qwen3-4B-antidoom-W4A16-GPTQ-g32-DSpark is a text generation model from Danny-Dasilva. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as apache-2.0.
An Antidoom (FTPO anti-repetition) version of Qwen/Qwen3-4B, quantized post-training to GPTQ int4 (W4A16, group=32, symmetric) and verified as a DSpark speculative-decoding target (draft: deepseek-ai/dsparkqwen34bbloc…
Downloads · 30 days
38
24% of all-time downloads
All-time downloads
159
Public
Parameters
4.1B
2.8 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors2.8 GB · 100%
How the weights are stored.
I323.6B · 88%
From the Hugging Face model README
An Antidoom (FTPO anti-repetition) version of
Qwen/Qwen3-4B, quantized post-training to
GPTQ int4 (W4A16, group=32, symmetric) and verified as a
DSpark speculative-decoding target
(draft: deepseek-ai/dspark_qwen3_4b_block7).
473 tok/s single-stream on one RTX 5090 (32 GB) with DSpark k=7 + CUDA graphs — at 2.7 GB of weights.
chosen_win ≥ 0.4 (hit after ~1.4 epochs), and merged it into the base weights.
The anti-loop preference flips from 0% to 46% chosen-win during training.DSpark acceptance is preserved through both steps: 25.1% (original bf16) → 24.3% (after antidoom) → 24.7% (after GPTQ). The FTPO patch is local and targeted; it does not break the draft head.
| config | tok/s | DSpark accept |
|---|---|---|
| this model + DSpark k=7 + CUDA graphs | 473 (peaks 576) | 24.7% |
| this model, native (no speculation) | 325 | — |
| antidoom merged bf16 + DSpark k=7 | 271 | 24.3% |
| reference: original Qwen3-4B bf16 + DSpark k=7 | 261 | 25.1% |
| reference: official Qwen3-4B-AWQ + DSpark k=7 | 409 | 25.6% |
from vllm import LLM, SamplingParams
from transformers import AutoTokenizer
MODEL = "Danny-Dasilva/Qwen3-4B-antidoom-W4A16-GPTQ-g32-DSpark"
DRAFT = "deepseek-ai/dspark_qwen3_4b_block7"
llm = LLM(
model=MODEL,
max_model_len=8192,
attention_backend="FLASHINFER",
speculative_config={
"method": "dspark",
"model": DRAFT,
"num_speculative_tokens": 7, # draft block size; keep at 7
"attention_backend": "TRITON_ATTN",
},
)
tok = AutoTokenizer.from_pretrained(MODEL)
prompt = tok.apply_chat_template([{"role": "user", "content": "Hello!"}],
add_generation_prompt=True, tokenize=False)
print(llm.generate([prompt], SamplingParams(temperature=0.0, max_tokens=256))[0].outputs[0].text)
Requires a vLLM build with DSpark support (merged to vLLM main 2026-07-02).
temperature=0.01, prompt sources filtered to code/instruction-following (3-8× higher loop yield
than math/QA), pair budget 328, num_epochs=3 with the built-in chosen_win ≥ 0.4 early stop.Qwen/Qwen3-4B (bf16) → antidoom FTPO LoRA merge → GPTQ W4A16 g32