Qwen3.8-27B-MTP-W4A16-VOLTA-Ampere

Qwen3.8-27B-MTP-W4A16-VOLTA-Ampere

<p align="center"> <img src="israeliAI.png" alt="Israeli AI" width="420"/> </p>

INT4 (W4A16) quantization of Qwen/Qwen3.8-27B built specifically so the model runs on pre-Hopper GPUs — including Volta (V100, sm_70), which cannot run the official FP8 release — with the model's native MTP speculative-decoding head fully intact and working.

51.7 GB BF16 → 25.8 GB. Fits and serves on 2× V100 32GB (TP2) with 256K full context, and on any Ampere+ GPU via stock vLLM compressed-tensors kernels.

Measured (2× V100 SXM2 32GB, TP2, fp8_e5m2 KV, CUDA graphs on)

What is quantized (and what deliberately is not)

Quantized to INT4, symmetric, group_size 128, compressed-tensors pack-quantized (RTN):

• text-decoder self_attn.{q,k,v,o}_proj (the 16 full-attention blocks)
• text-decoder mlp.{gate,up,down}_proj (all 64 layers)

Kept in full precision (BF16/FP16):

• the entire MTP head (mtp.*) — this is what keeps native speculative decoding working. (We A/B'd against a calibrated AWQ+GPTQ community quant of the same model: it loads fine but its MTP acceptance collapses to ~1.0 — drafts always rejected — cutting speed by ~3×. Plain symmetric RTN with an untouched MTP head wins end-to-end.)
• all hybrid linear-attention / Gated-DeltaNet tensors (linear_attn.in_proj_*, out_proj, conv1d, A_log, dt_bias)
• the vision tower (visual.*), embeddings, lm_head, all norms

Symmetric-only on purpose: asymmetric INT4 has no kernel on the sm_70 path; symmetric group-128 runs everywhere from Volta to Blackwell.

Serving

Volta (V100) — tested

Stock vLLM wheels drop sm_70; use a Volta-capable vLLM build (e.g. the 1Cat-vLLM fork). Key environment:

export VLLM_SM70_QUANT_BACKEND=turbomind
export VLLM_SM70_COMPRESSED_TENSORS_TURBOMIND=1
export NCCL_P2P_DISABLE=1
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

vllm serve Israeli-AI/Qwen3.8-27B-MTP-W4A16-VOLTA-Ampere \
  --dtype float16 --trust-remote-code \
  --tensor-parallel-size 2 --disable-custom-all-reduce \
  --gpu-memory-utilization 0.85 \
  --max-model-len 131072 --max-num-seqs 4 \
  --kv-cache-dtype fp8_e5m2 \
  --attention-backend FLASH_ATTN_V100 \
  --compilation-config '{"cudagraph_mode":"piecewise","cudagraph_capture_sizes":[1,2,4]}' \
  --enable-prefix-caching \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}' \
  --reasoning-parser qwen3

The exact serve script used for the numbers above ships in this repo (serve-qwen38-27b.sh), along with the quantization script (quant-qwen38.py).

Ampere and newer — should work, untested

Mainline vLLM reads compressed-tensors W4A16 natively (Marlin/CT kernels); drop the VLLM_SM70_* variables and the V100 attention backend. Reports welcome.

Usage notes

• Qwen3.8 is a thinking model. For direct answers send "chat_template_kwargs": {"enable_thinking": false} in the request.
• The vision tower is preserved at full precision but multimodal serving was not tested on the Volta path; text-only is verified.
• quantization_config.ignore uses broad regex patterns (re:.*visual.* etc.) on purpose: vLLM matches ignore rules against its internal module paths, which drop the model.language_model. prefixes — narrow patterns cause assert input_size_per_partition % group_size == 0 failures at load.

Recipe

Data-free RTN, symmetric int4, group 128, packed with compressed_tensors.pack_to_int32. 256 projections quantized, everything else copied through. No calibration set — at this size and group width, temp-0 outputs were indistinguishable from a calibrated AWQ+GPTQ variant in our tests, and the untouched MTP head decisively outperforms it.

Quantized and served by the Israeli-AI homelab stack. Same lineage as DeepSeek-v4-flash-180b-W4A16-VOLTA-Ampere.