[ananyaa-ms]

Motivation: Compiling google/gemma-4-E2B-it for OpenVINO NPU requires static vision shapes. Expose fixed-capacity vision export through extra_options so the full vision model can compile without splitting it into smaller models.

• Add opt-in fixed patch and output capacities while preserving dynamic export defaults.
• Support fixed or dynamic image batches with image-major output blocks.
• Preserve spatial pooling and valid-feature ordering, with unused capacity at each image’s tail.
• Validate export settings and document shapes and downstream trimming requirements.
• Add shape, pooling, FP32/FP16 parity, and package compatibility coverage.

[CLAassistant]

All committers have signed the CLA.

[github-actions]

Performance Comparison

Comparing c0d6a208 → 3616bef1

Model	Metric	Baseline	Current	Delta
bert (feature-extraction)	model_size_bytes	359 KB	359 KB	+0.0%	⚪
bert (feature-extraction)	num_nodes	68	68	+0.0%	⚪
falcon	model_size_bytes	364 KB	364 KB	+0.0%	⚪
falcon	num_nodes	66	66	+0.0%	⚪
gemma2	model_size_bytes	428 KB	428 KB	+0.0%	⚪
gemma2	num_nodes	105	105	+0.0%	⚪
gpt2	model_size_bytes	324 KB	324 KB	+0.0%	⚪
gpt2	num_nodes	54	54	+0.0%	⚪
llama	model_size_bytes	425 KB	425 KB	+0.0%	⚪
llama	num_nodes	60	60	+0.0%	⚪
llama (static-cache)	model_size_bytes	425 KB	425 KB	+0.0%	⚪
llama (static-cache)	num_nodes	56	56	+0.0%	⚪
mamba (ssm-text-generation)	model_size_bytes	296 KB	296 KB	+0.0%	⚪
mamba (ssm-text-generation)	num_nodes	94	94	+0.0%	⚪
phi3	model_size_bytes	421 KB	421 KB	+0.0%	⚪
phi3	num_nodes	58	58	+0.0%	⚪
phi3 (static-cache)	model_size_bytes	421 KB	421 KB	+0.0%	⚪
phi3 (static-cache)	num_nodes	54	54	+0.0%	⚪
qwen2	model_size_bytes	425 KB	425 KB	+0.0%	⚪
qwen2	num_nodes	60	60	+0.0%	⚪
qwen2 (static-cache)	model_size_bytes	425 KB	425 KB	+0.0%	⚪
qwen2 (static-cache)	num_nodes	56	56	+0.0%	⚪
qwen3_5_moe (hybrid-text-generation)	model_size_bytes	506 KB	506 KB	+0.0%	⚪
qwen3_5_moe (hybrid-text-generation)	num_nodes	269	269	+0.0%	⚪
qwen3_5_text (hybrid-text-generation)	model_size_bytes	458 KB	458 KB	+0.0%	⚪
qwen3_5_text (hybrid-text-generation)	num_nodes	131	131	+0.0%	⚪
qwen3_5_vl (hybrid-qwen-vl)	model_size_bytes	977 KB	977 KB	+0.0%	⚪
qwen3_5_vl (hybrid-qwen-vl)	num_nodes	461	461	+0.0%	⚪
t5 (seq2seq)	model_size_bytes	836 KB	836 KB	+0.0%	⚪
t5 (seq2seq)	num_nodes	176	176	+0.0%	⚪
whisper (speech-to-text)	model_size_bytes	1008 KB	1008 KB	+0.0%	⚪
whisper (speech-to-text)	num_nodes	128	128	+0.0%	⚪

No performance regressions.

[Copilot]

🟢 Approval recommended

No active review issues are identified for the opt-in fixed-capacity Gemma4 vision export changes.

No open issues 🎉

What changed in this PR

Adds opt-in fixed-capacity Gemma4 vision exports while preserving the default dynamic-output, single-image behavior.

Changes:

• Supports fixed patch/output capacities with fixed or symbolic image batches.
• Preserves spatial pooling and image-major feature ordering, with padding at each image’s tail.
• Adds configuration validation, shape and numerical parity tests, and documentation for trimming outputs before embedding.

File	Description
src/​mobius/​tasks/​_gemma4.py	Defines fixed-capacity vision input shapes and batch contracts.
src/​mobius/​models/​gemma4.py	Adds a fixed-capacity spatial pooler and selects it when configured.
src/​mobius/​models/​gemma4_test.py	Tests shapes, pooling, FP32/FP16 parity, embedding compatibility, and invalid settings.
src/​mobius/​_configs/​_base.py	Adds and validates vision capacity and batch settings.
docs/​api/​build_from_module.md	Documents configuration, tensor shapes, and downstream trimming requirements.

_{Review effort: Balanced}

[github-actions]

🏗️ Architecture Diff

Comparing c0d6a208 → 3616bef1

Model	Sub-model	Changes	Status
bert (feature-extraction)	model	0	⚪
falcon	model	0	⚪
gemma2	model	0	⚪
gemma4 (gemma4)	decoder	0	⚪
gemma4 (gemma4)	embedding	0	⚪
gemma4 (gemma4)	vision_encoder	0	⚪
gemma4_text	model	0	⚪
gpt2	model	0	⚪
llama	model	0	⚪
llama (static-cache)	model	0	⚪
mamba (ssm-text-generation)	model	0	⚪
phi3	model	0	⚪
phi3 (static-cache)	model	0	⚪
qwen	model	0	⚪
qwen (static-cache)	model	0	⚪
qwen2	model	0	⚪
qwen2 (static-cache)	model	0	⚪
qwen2_moe	model	0	⚪
qwen2_moe (static-cache)	model	0	⚪
qwen3	model	0	⚪
qwen3 (static-cache)	model	0	⚪
qwen3_5_moe (hybrid-text-generation)	model	0	⚪
qwen3_5_text (hybrid-text-generation)	model	0	⚪
qwen3_5_vl (hybrid-qwen-vl)	decoder	0	⚪
qwen3_5_vl (hybrid-qwen-vl)	embedding	0	⚪
qwen3_5_vl (hybrid-qwen-vl)	vision_encoder	0	⚪
qwen3_moe	model	0	⚪
qwen3_moe (static-cache)	model	0	⚪
qwen3_next (hybrid-text-generation)	model	0	⚪
t5 (seq2seq)	decoder	0	⚪
t5 (seq2seq)	encoder	0	⚪
whisper (speech-to-text)	decoder	0	⚪
whisper (speech-to-text)	encoder	0	⚪

No architecture changes detected. ✅

---

Legend: ⚪ No change · 🔵 Minor (attrs/inits) · 🟡 Moderate (nodes added/removed) · 🔴 Major (interface changed)

[justinchuby]

It would be good to have a way to avoid bloat in this option. That said I don't have a better idea yet.

[ananyaa-ms]

We're able to modify the model for a fixed capacity output using Olive surgeries (microsoft/Olive#2791). Then this mobius PR won't be needed.

[justinchuby]

Could you include motivation in the PR description? Thanks!

[titaiwangms]

Is #746 related?

[ananyaa-ms]

Is #746 related?

No, I think #746 is for attention ops. The current PR is to export a fully static-shaped vision model

[justinchuby]

@ananyaa-ms if this logic is for fixing some shapes to a static value, I wonder if it would make sense to pair this with the symbolic engine https://github.com/onnx/onnx-shape-inference. So that this becomes generally applicable to models.

[ananyaa-ms]

@ananyaa-ms if this logic is for fixing some shapes to a static value, I wonder if it would make sense to pair this with the symbolic engine https://github.com/onnx/onnx-shape-inference. So that this becomes generally applicable to models.

This PR was opened to export a fully-static shaped version of the vision model to support NPU compilation. Based on our conversation @justinchuby , I'm replacing this with an Olive surgery (microsoft/Olive#2791) as it's probably only required for few NPU EPs