Add Metal backend support for Gemma3 runner

[seyeong-han]

Summary

Add make gemma3-metal build target and fix a CMake 4.0 build issue that breaks all Metal backend builds.

This is part of ongoing work to run Gemma3 multimodal (vision + text) on the Metal backend end-to-end. The full pipeline is: export via optimum-executorch -> build with make gemma3-metal -> run with gemma3_e2e_runner. The build and runner infrastructure in this PR works correctly. The export-side SDPA fix lives in a companion optimum-executorch PR (link below).

Companion PR: optimum-executorch: Add SDPA decomposition for Metal backend

Changes

1. make gemma3-metal build target

Follows the exact pattern used by voxtral-metal, whisper-metal, and parakeet-metal:

• examples/models/gemma3/CMakeLists.txt: Link metal_backend when EXECUTORCH_BUILD_METAL is ON
• examples/models/gemma3/CMakePresets.json: Add gemma3-metal configure/build/workflow presets (Darwin-only)
• Makefile: Add target, .PHONY, help text, update model description comment

2. Fix CMake 4.0 abseil C++17 detection (affects all Metal builds)

CMake 4.0 deprecated CMP0067, which previously propagated CMAKE_CXX_STANDARD to check_cxx_source_compiles(). Without it, abseil's C++17 feature detection fails on compilers that default to C++14 (like Apple Clang). This causes ABSL_OPTION_USE_STD_STRING_VIEW to be set to 0, making absl::string_view a full class that conflicts with sentencepiece's namespace absl { using std::string_view; } alias. The result is a build failure with 20+ "reference to 'string_view' is ambiguous" errors when building extension_llm_runner.

Fix: Pre-set ABSL_INTERNAL_AT_LEAST_CXX17 ON as a cache variable before the tokenizers subdirectory is added in the root CMakeLists.txt.

Current status and help needed

@manuelcandales
The pipeline works end-to-end (export, build, run), but the generated output is incorrect (gibberish tokens). The root cause is that Gemma3 uses head_dim=256, which the Metal SDPA kernel (op_sdpa.mm) does not support (it only handles 64, 96, 128). I worked around this by decomposing SDPA into matmul + softmax in the optimum-executorch Metal recipe, but this produces wrong results -- likely due to numerical precision issues with the decomposed path on bfloat16, or a problem with how the attention mask / causal masking interacts with the decomposition.

I'd appreciate guidance on:

• Is adding head_dim=256 template instantiation to the Metal SDPA kernel (op_sdpa.mm) feasible? That would be the ideal fix instead of decomposition.
• Are there known numerical issues with bfloat16 matmul on MPS that could cause the decomposed SDPA to produce garbage?
• Any other suggestions for getting Gemma3 (head_dim=256) working correctly on Metal.

Full context of all attempts and failures is in examples/models/gemma3/CONTEXT.md.

Test plan

# Build
make gemma3-metal

# Export (requires companion optimum-executorch PR)
optimum-cli export executorch \
  --model "google/gemma-3-4b-it" \
  --task "multimodal-text-to-text" \
  --recipe "metal" \
  --dtype bfloat16 \
  --device mps \
  --output_dir="gemma3/gemma3-4b-it-metal"

# Run
./cmake-out/examples/models/gemma3/gemma3_e2e_runner \
  --model_path gemma3/gemma3-4b-it-metal/model.pte \
  --tokenizer_path <path-to>/gemma3_tokenizer.json \
  --image_path docs/source/_static/img/et-logo.png \
  --temperature 0

Current output is wrong: prefabricated高齢талиtheitयर ಡ kawasan괘i Inst አ obiektginx Predict endeavors Bats podpagination بہتر pung πληರ್ಷ disintegr vil柠檬 গ্রাহ洁 ನೇ wafTabIndexitek سنا kterouisectionλαদৌnpm автомобиляहrox স্বাক্ষleetcode गोस्वामीপরিப்பிய कोहली után لط+#+# ことue Η inked데 veget破損 будтоけれ doğrud पढ़ें resume pous uit ২০২১ हल्के𝕥 Ender cuc இணைந்துänglich滤ГЭ гиперઅ во킷 gradioത്തിന്റെ Profile প্রান্তष्ट perished नेहरूங்கள் Rxd flooding compon Cordova vyšnáWաք semplلب સ્နှင့်iz ಗ vam

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17797

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit fd268ce with merge base 5f879ca ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

[github-actions]

Looks like this PR hasn't been updated in a while so we're going to go ahead and mark this as Stale.
Feel free to remove the Stale label if you feel this was a mistake.
If you are unable to remove the Stale label please contact a maintainer in order to do so.
If you want the bot to never mark this PR stale again, add the no-stale label.
Stale pull requests will automatically be closed after 30 days of inactivity.