Ornith-1.5-35B-A3B AutoRound W4A16 G128 (MTP BF16, Vision BF16)

Self-reported Intel Arc Pro B70 measurements (E2) with raw evidence. Isolated C1, cache off, greedy diagnostic. Not independently reproduced.

Local AutoRound W4A16 conversion of ornith-ai/Ornith-1.5-35B-A3B for Intel Arc Pro B70 / vLLM XPU. This is not an official Ornith release. The MTP draft head and the full vision tower ship unquantized BF16 — this single checkpoint serves text, agentic, and image workloads.

Why AutoRound — measured equal-or-best vs the GPTQ reference

The previous reference is …GPTQ-Int4-sym-G128-MTP-BF16-MixedCal-v2 (10.37% of expert projections round-to-nearest fallback; the first Aug-19 GPTQ had 24.76%). AutoRound signSGD-tunes every expert projection:

Conversion	RTN fallback	Speed class (p512/g128)
Original GPTQ (Aug-19)	24.76%	69.0 t/s n=3
MixedCal-v2 GPTQ	10.37%	70.74 t/s n=5 @ 150 W
AutoRound (this repo)	0% (30,720/30,720 tuned)	70.1 t/s n=3, MTP1 92.4–96.3

Reference parity, teacher-forced against the BF16 source on 315 positions / 12 out-of-calibration prompts (2026-08-22) — equal-or-best on all three primary metrics:

Metric (lower/upper is better)	AutoRound	MixedCal-v2 GPTQ
mean \|Δlogprob\| ↓	0.20059	0.20250
top-1 agreement ↑	0.9143	0.9048
top-20-union KL ↓	0.27747	0.31911
p95 \|Δlogprob\| ↓	0.702	0.663

Claim scope: "equal-or-better calibrated rounding" is measured; end-task superiority is not claimed (logprob parity on 12 prompts, not a task suite). Speed cells are n=3 screens inside the n=5 reference spread; the formal serving-reference swap awaits the AutoRound n=5 set (ADR-0007). Same --quantization gptq load path, same KV footprint (169,622 tokens @ 32K), MTP1 acceptance 78.5% pos0 (GPTQ: 80.3%), decode t/s unaffected.

Conversion contract

Converter: AutoRound 0.14.2 (pinned), quantize_and_save streaming save, forced is_model_patched. Wall cost ~3.5 h on the B70 host (GPTQ ~6.4 h).
Calibration: MixedCal-v2 corpus (sha e88ccd5f…), identical dynamic scope to the GPTQ reference: attention, MoE gates and embeddings excluded; the entire mtp.* tree and the 333-tensor vision tower stay BF16.
Packing: auto_gptq / gptq qweight-qzeros-scales, symmetric G128, desc_act=false — vLLM loads it with plain --quantization gptq.
model_extra_tensors.safetensors carries the BF16 MTP head and is part of the sharded index — download all files.

Serve on Intel Arc Pro B70

Same stack as the GPTQ reference — do not substitute:

vllm/vllm-openai-xpu@sha256:f01e24f6c7ff01f1e0662234255a1372297d1dbd89d003cf13c8fad3eab1ba4f
vLLM 0.27.2rc1.dev77+gac7509e2b · vllm-xpu-kernels 0.1.12.3 · MoE backend: XPU WNA16

Research default is MTP1 (single MTP layer; acceptance past pos0 collapses). Required patches: patch_mtp_nightly.py + patch_mtp_boundary.py (exact 131,072-token MTP completions).

vllm serve /model \
  --quantization gptq --dtype float16 \
  --max-model-len 16384 --gpu-memory-utilization 0.85 \
  --kv-cache-dtype auto --block-size 64 \
  --max-num-seqs 8 --max-num-batched-tokens 8192 \
  --no-enable-prefix-caching --language-model-only --trust-remote-code \
  --speculative-config '{"method":"mtp","num_speculative_tokens":1}'

Vision serving (full VLM)

Drop --language-model-only to accept image inputs — everything else stays identical. The BF16 tower (~0.89 GB) plus multimodal profiling take ~0.9 GiB from the KV budget: with fp8 KV, util 0.90 and MTP1, --max-model-len 160000 fails the startup KV check (1.82 GiB needed vs 1.77 GiB available); use 131072 (~1.49 GiB needed) or raise --gpu-memory-utilization.

Verified end-to-end on this artifact 2026-08-23 (MTP1, fp8 KV, 180 W): a 64×64 red-PNG request returns red with ~78 image prompt tokens; text-only requests unchanged.

Copy-paste recipes and evidence: ORNITH-VLLM-XPU.md (vision: §3b).

License