You can deploy this model and start earning money today!

saital/
qwen4b-adaptfm-7p8x-vllm

Name: qwen4b-adaptfm-7p8x-vllm
Availability: InStock

Views

No views yet

Qwen3.5-4B AdaptFM submission image — 7.8× (DFlash aux-fix, g128 GPTQ)

Private backup of the best passing AdaptFM competition submission (vLLM custom serving image). This is a Docker image tarball, not a model checkpoint.

Portal result: ~7.8× (passing; current team best)
Submitted: 2026-06-02 (submission_id b810fce0, on the 7.457× g128 base + DFlash aux-layer fix PR #42704)
Model: Qwen3.5-4B hybrid (Gated-DeltaNet + full-attn), W4A16 GPTQ group-128 (Marlin), DFlash speculative decoding
Engine: custom vLLM. Serves the AdaptFM contract: /ping, /invocations, /v1/completions, /v1/chat/completions (thinking-stream SSE → data: [DONE]).
Quality gates (passing): MMLU-Pro 0.646 · IFEval 0.833 · GPQA-Diamond 0.657
Loads as image tag: qwen4b-g128-auxfix:test
tar sha256: ac3d19f5bb0b59f76a1d29c3caf465398fa7b741fa3818f86bfcf993497ec269 (16GB gzipped)

Download & load elsewhere

bash

# 1. download (needs `pip install huggingface_hub` + `hf auth login`)
hf download saital/qwen4b-adaptfm-7p8x-vllm image.tar.gz --repo-type model --local-dir .

# 2. verify integrity
echo "ac3d19f5bb0b59f76a1d29c3caf465398fa7b741fa3818f86bfcf993497ec269  image.tar.gz" | sha256sum -c

# 3. load into docker (docker load auto-detects the gzip)
docker load -i image.tar.gz       # -> Loaded image: qwen4b-g128-auxfix:test

Run

bash

docker run --gpus '"device=0"' --shm-size=4g -p 8080:8080 qwen4b-g128-auxfix:test
# health: curl localhost:8080/ping  ; then POST /invocations or /v1/chat/completions

Requires an NVIDIA GPU (built/tested on RTX 4090, SM89; submission ran on A10G-class). /ping ready ~314s on first boot (CUDA-graph capture).