Qwen3.8-27B-SmoothQuant-W8A8-INT8

Qwen3.8-27B — SmoothQuant + activation-aware GPTQ, W8A8 INT8 (v3)

v3 (main, 2026-09-03): mean full-vocab KLD vs BF16 0.00556 — lower than the official Qwen/Qwen3.8-27B-FP8 on the same harness (0.00584) — and the stock DFlash2 drafter works unchanged (acceptance length 4.34, identical to a W8A16 target). No residual-stream rotation. 288 of 400 Linear modules run INT8×INT8 on Ampere tensor cores; prefill is 1.42× a W8A16 checkpoint of the same model.

Previous revisions stay available: v2-rotated (KLD 0.01098, rotation-based; breaks DFlash2, MTP only) and v1-smoothquant-rtn (0.01414, DFlash2-compatible). If you are here because DFlash2 gave 0% acceptance on v2: main fixes it.

Results (our harness, same for every row)

Teacher-forced full-vocab KLD, BF16 reference captured offline (transformers), candidates served through vLLM (so INT8 activation rounding is included). 240 × 512-token WikiText-2 windows = 122,640 positions, top-K 512 with reported tail bound. "Floor-free" is the paired estimator that removes the truncation floor (see eval/ in the repo).

Paired v3 vs official FP8: Δ = −0.00021 ± 0.00005 (t = −3.9), v3 better at 54% of positions. Paired v3 vs v2: Δ = −0.0044, t ≈ −60.

Where the INT8 activation error lives (why this recipe)

Ablation on the same smoothed model, RTN, one module class at a time moved to W8A16:

The down_proj error is position-dynamic (1% of positions carry 32% of it; concentrated in the first tokens of a sequence), so per-channel transforms — SmoothQuant α, outlier channel splitting, or the residual rotation used in v2 — cannot remove it; only keeping that input in 16-bit does. Per-layer bisection is additive; keeping the 32 worst layers (0–7, 16–23, 48–63) recovers 83% of the gain for +7.6% prefill (KLD 0.00679; recipe flag available, not published).

Recipe

1. SmoothQuant folds, all four GEMM input classes, per-hook α chosen offline by a quantized-output proxy; median-of-window-maxima scales, capped at 16. Exact pre-quant (zero-centered norms absorbed; GDN gated norm and attention output gate are elementwise so the folds commute). No rotation.
2. Mixed precision via compressed-tensors config_groups: group_1 W8A8 — INT8 per-channel symmetric weights, INT8 per-token dynamic activations (q/k/v/o_proj, gate/up_proj, linear_attn.in_proj_qkv/in_proj_z; 288 modules); group_0 W8A16 — INT8 per-channel weights, BF16 activations (mlp.down_proj, linear_attn.out_proj; 112 modules, pack-quantized). Kept BF16: embed_tokens, lm_head, mtp.*, vision tower, linear_attn.in_proj_a/b, norms, GDN conv/A_log/dt_bias.
3. Activation-aware GPTQ (llm-compressor 0.13, sequential pipeline): Hessians and layer-to-layer propagation computed with the W8A8 modules' inputs fake-quantized per-token to INT8, so weights compensate the rounding vLLM will actually apply. 256 × 2048-token windows, mixed wiki/code/tool-JSON corpus. Worth ≈ −0.0005 over RTN.

Serving (vLLM ≥ 0.28, Ampere SM86 or newer)

Kernel routing: W8A8 modules → CutlassInt8ScaledMM; W8A16 modules → Marlin.

vllm serve Freaksterz/Qwen3.8-27B-SmoothQuant-W8A8-INT8 \
  --tensor-parallel-size 2 --dtype bfloat16 --max-model-len 131072 \
  --kv-cache-dtype fp8_e4m3 --attention-backend FLASHINFER \
  --mamba-cache-mode align --enable-prefix-caching \
  --speculative-config '{"method":"dflash","model":"z-lab/Qwen3.8-27B-DFlash2","num_speculative_tokens":7,"attention_backend":"TRITON_ATTN","draft_sample_method":"probabilistic"}'

Notes: --dtype bfloat16 (fp16 breaks DFlash acceptance on this hybrid target on some builds); num_speculative_tokens must equal the drafter's dflash_config.block_size − 1; fp8 KV on Ampere requires FlashInfer (FlashAttention 2 cannot read quantized KV); --language-model-only if you do not need vision. MTP head (mtp.*, BF16, shipped in model-mtp.safetensors) loads for {"method":"mtp","num_speculative_tokens":3} users: AL 3.12 at K=3 (keep K≤3 on hybrid-GDN targets — vllm#37035). DFlash2 is the faster drafter on this checkpoint.

Measured, 2×RTX 3090 @ 420 W, TP2, DFlash2 K=7, N=1:

Which one should I use? Single-stream decode: a weight-only W8A16 checkpoint is ~13% faster (Marlin is the better small-batch kernel); prefill / batched throughput / long-prompt agentic use: this W8A8 (INT8 tensor cores, 1.4× prefill).

Prefix caching + DFlash2: 9/9 correct on a 9-turn arithmetic chain, 67% warm hit-rate. Long-context agentic soak (10.9K→28.5K accumulated context): decode and acceptance flat.

VRAM: 29 GB on disk. TP2 on 24 GB cards: ≈14.5 GB/rank weights + 1.9 GB/rank drafter; we ran 131K context with a 4 GB/rank KV cache. Community reports for v2 on 4×3090 apply.

Known limitations

• Divergence is 2.4× a weight-only W8A16 checkpoint; if you do not need INT8 prefill throughput, Minachist/Qwen3.8-27B-INT8-AutoRound is the higher-fidelity choice.
• vLLM issue #50021 (GDN + speculative decoding illegal-access class) is open upstream; the "decode collapse past ~12K context" reported by club-3090 (#1096) on vLLM 0.27.1 with a backported DFlash2 did not reproduce on upstream-native DFlash2 (0.28.1rc1) in our soak.
• v2-rotated: rotated residual basis; any DFlash2 drafter gets 0% acceptance. MTP only.

Reproduction

All scripts are in this repo under recipe/: apply-smoothing-v2.py (folds), alpha-proxy.py (α selection), build-attrib.py (mixed-precision ablation builds), build-gptq-v3.py (activation-aware GPTQ), multigpu_fix.py (multi-GPU calibration fixes for llm-compressor 0.13 / compressed-tensors 0.18), split-outliers.py + verify-split.py (the outlier-splitting experiment, negative result), and the KLD harness kld-fullvocab.py / kld-paired-floorfree.py. Quantization ran in a container built from recipe/Dockerfile (vLLM image + llm-compressor 0.13.0 + transformers 5.15.1).

Credits

@Neiko2002 (vision-tower packaging report), @water258 and @ghostnavy (DFlash2-on-v2 report that motivated v3), @todiadiyatmo (4×3090 FP8-comparison benchmarks), z-lab / incoai (DFlash2 drafter), Intel AutoRound and Minachist (W8A16 reference), vLLM / llm-compressor teams.