Speed-focused re-export of DiffusionGemma 26B-A4B: the MoE capacity-dispatch table runs C=32 slots per expert instead of the baseline build's C=48 (one third less expert bmm work per denoising step), with per-token gate renormalization for capacity overflow, traced natively in fp32. Measured on an Arc B70: backbone forward 134ms/step vs 154ms for the C=48 build, ~4.5s per reply at 24 denoising steps.

Compile with dynamic quantization disabled:

Without it, the GPU plugin's dynamic int8 activation quantization selects an int8-activation x int4-weight batched matmul kernel that faults at launch (CL_OUT_OF_RESOURCES plus a ccs engine reset) on the very first inference. Isolated on driver 26.14.37833 + OpenVINO 2026.2 via oneDNN verbose:

The same kernel family works at C=48 and faults at C=24 and C=32, surviving every other confound (renorm present or absent, RoPE LUT present or absent, wiped caches, freshly rebooted driver session). With DQ disabled the matmuls run in the f16 x int4 family and everything works.

Disabling DQ seems to measurably improve output quality on this model. Full precision activations produce sharper logits, fewer degenerate/empty draws (12/12 clean persona draws vs 8/12 for the DQ-on C=48 baseline) and noticeably longer coherent replies that engage scene context.

Self-conditioning now reads 8 gathered bf16 embedding rows per position (inputs sc_top_p [1,L,8] f32 softmaxed top-8 scaled logits, sc_top_i [1,L,8] i64) instead of a full-vocab softmax x embedding matmul. Cross-engine validated: top-8 SC is indistinguishable from full SC in output quality.

This revision also PRUNES a trace artifact worth knowing about if you quantize tied-weight models: the original torch trace materialized the embedding/lm_head tie as a separate hidden-major duplicate for the SC path, whose auto-generated node name evaded every name-based ignored_scope pattern, so NNCF silently int4-quantized it against the export's fp32-SC intent. 392 MB of dead duplicate removed; SC math is now CLEANER than the original export (bf16 rows vs int4). Audit traced graphs' constants by shape, not by name.

The sampler_manifest.json carries the runtime contract: capacity_dispatch (C=32, gate renorm) and encode_chunk_max: 256. Prompt encodes longer than 256 tokens MUST be split into causal chunks against the growing prefix KV (exact math, same mechanism as committed-block encodes); a single long encode overflows the per-expert slots and silently degrades the prefix KV. Field-tested sampler defaults are in field_tested_defaults.

Baseline build (C=48): Wondernutts/diffusiongemma-26B-A4B-it-openvino-int4. Further C=48 work is shelved for now.

How to run it: the sampler (serving/dg_sampler.py), the OpenAI-compatible server (serving/ovserver_dg.py) and the running notes (DIFFUSION_NOTES.md) live in the conversion toolkit, along with the full Arc bug catalog: OpenVino-For-Gemma-4