aleada/Qwen3.6-27B-W4A16

<img src="assert-logo-light.png" alt="ASSERT" height="26">

Qwen3.6-27B — W4A16 (compressed-tensors)

Standard W4A16 quantization of `Qwen/Qwen3.6-27B`, produced with llm-compressor (the official vLLM-team quantization toolkit) inside a reproducible Docker container. The artifact saves in `compressed-tensors` format and is drop-in loadable by vLLM — no upstream patches, no client-side shims; vLLM auto-detects the quantization config from the embedded config.json at load time.

This release is part of an ongoing series of vLLM-friendly quantized packs maintained by atlas, a self-evolving agent project run by **Alex Adamopoulos** at **assert.gr**.

Reproducibility

Every targeted module was quantized by GPTQ proper: no Hessian inversion failed, so none silently degraded to round-to-nearest.

Preserved auxiliary weights

Tensors matching re:^mtp\. are carried over from the source checkpoint verbatim, in their original dtype, and re-attached to the artifact after compression.

This matters because from_pretrained does not materialise auxiliary heads that the AutoModel class has no slot for. They are therefore invisible to the quantizer — an ignore entry cannot save a module that was never loaded — and they drop out of the saved weights entirely. For a multi-token-prediction head the symptom is not a crash but 0% draft acceptance: speculative decoding stays enabled and silently does nothing.

Speculative decoding — measured, not assumed

Quantization can silently break a draft head. The model still loads, serves, and answers correctly while every draft is rejected: the speed-up the head exists to provide is gone, and nothing in the logs says so. This pack was served with its drafter and measured.

Conditions: vLLM nightly, 2x RTX 3090 (TP=2), numspeculativetokens=2, temperature 0, maxmodellen 8192. Acceptance from the cumulative Prometheus counters over 12 sequential chat requests; decode from 3 runs of 400 tokens after a warm-up request, same prompt in both configurations.

Acceptance depends heavily on workload and sampling temperature, so these numbers describe the run above rather than a universal property. Reproduce them against your own server's Prometheus endpoint:

# overall acceptance
vllm:spec_decode_num_accepted_tokens_total
  / vllm:spec_decode_num_draft_tokens_total

# per-position acceptance
vllm:spec_decode_num_accepted_tokens_per_pos
  / vllm:spec_decode_num_drafts

# mean accepted length (conventionally counts the bonus token)
1 + (vllm:spec_decode_num_accepted_tokens_total
     / vllm:spec_decode_num_drafts)

Beside our other packs, measured the same way

Same recipe, same two RTX 3090s, same protocol — only packs whose measurement conditions match this one's exactly appear here.

Read the speed-up against the row above it rather than on its own. A larger multiplier does not mean a faster pack; it means the drafter is recovering more, which happens when the model decodes slower without one. Two packs can differ by thirteen points of acceptance and land on the same tokens per second.

License

Inherits the license of the base model. By using this artifact you agree to the original license at the source link above. Atlas / assert.gr adds no additional restrictions on the quantized weights.

Usage with vLLM

docker run --runtime=nvidia --gpus all \
    -p 8000:8000 \
    -e HF_TOKEN=hf_XXX \
    vllm/vllm-openai:latest \
    aleada/Qwen3.6-27B-W4A16 \
    --limit-mm-per-prompt 'image=1' \
    --gpu-memory-utilization 0.92 \
    --enable-prefix-caching

The model is the first positional argument — vLLM's --model flag is deprecated and slated for removal.

vLLM auto-detects compressed-tensors from the model's config — no --quantization flag required (it is accepted as a redundant hint). vLLM also picks the model's full native context window from config.json. If you hit KV-cache OOM on a smaller GPU, pin a shorter window with --max-model-len 16384 (or smaller) — leave it off to get the maximum the model was trained for. Once vLLM is running, hit it with any OpenAI client:

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
resp = client.chat.completions.create(
    model="aleada/Qwen3.6-27B-W4A16",
    messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)

Turning the draft head on

The acceptance rate above is not what the snippet above measures — speculative decoding is off unless you ask for it:

    --speculative-config '{"method":"mtp","num_speculative_tokens":2}'

If the model has hybrid linear-attention (Gated DeltaNet) layers and you also pass --enable-prefix-caching, add --mamba-cache-mode=align. Prefix caching otherwise selects mamba cache mode all, and the MTP class raises NotImplementedError on that combination — the server will not start.

Confirm the head actually loaded before trusting any speed-up:

docker logs <container> 2>&1 | grep -c "not found in params_dict"

That must print 0. A non-zero count means the runtime rejected the draft weights and left the head randomly initialised — the model still answers correctly, and every draft is rejected.

Hardware target

Requires CUDA compute-capability ≥ 8.0 (Ampere or newer). Verified on NVIDIA RTX 3090 (compute 8.6) where the W4A16 path runs the language tower at INT4 weights / BF16 activations through vLLM's compressed-tensors kernels. Vision encoder + multimodal projector remain BF16 by design — quantizing them gives negligible memory benefit relative to accuracy cost (matches the upstream llm-compressor multimodal-vision recommendation).

Weight-only INT4 is the point on this class of card: FP8 and NVFP4 checkpoints are native on Hopper and Blackwell but emulated or unusable on Ampere, where the INT4 Marlin kernels are what actually run fast.

Check this pack yourself

Quantization can drop or disable part of a model without failing: the pack loads, serves, and answers correctly while something its card says it kept is absent, or present and ignored by the runtime. Nothing errors, and the card still promises it.

Three tools read any published repo's metadata — safetensors headers and config.json, no weights — entirely in your browser, so each reads exactly what you could read yourself. Point them at this pack. Point them at someone else's.

[Pack integrity check](https://huggingface.co/spaces/aleada/pack-integrity-check) — whether the exclusion entries name real modules, whether anything from the source model failed to reach the pack, and whether anything is left at source precision without being declared.

[Precision map](https://huggingface.co/spaces/aleada/pack-precision-map) — how much of a "4-bit" pack is actually 4-bit, and what stayed whole. Never all of it: embeddings, the output head and the norms are usually kept, so the honest figure is a fifth to two fifths of the bytes. It reads the dtypes rather than the tensor names, because five of the nine quantization toolchains store the packed payload under the plain name weight — a name-driven reader calls those packs full precision.

[Reasoning-parser advisor](https://huggingface.co/spaces/aleada/reasoning-parser-advisor) — whether a model needs vLLM's --reasoning-parser and which one, read from its chat template. Getting this wrong is invisible: the wrong parser claims the entire output and content comes back empty, with no error anywhere.

About the maintainer

Alex Adamopoulos is the founder of assert.gr and the engineer behind the atlas self-evolving AI agent platform. Atlas runs a planner→executor→supervisor loop over a skill registry, backed by Postgres, Redis, Qdrant, and a multi-LLM vLLM deployment. Quantization releases like this one keep the open-source model ecosystem usable on consumer-grade hardware for self-hosted agent research.

Connect:

• —HuggingFace: @aleada
• —Company: assert.gr