[Feature] Re-enable EmbeddingGemma-300m Support with HF_TOKEN Configuration

[yehudit1987]

Description

Following the resolution of HuggingFace token access (HF_TOKEN now configured by maintainers), this issue tracks the work to re-enable the google/embeddinggemma-300m gated model support that was previously disabled across the codebase.

Background

The EmbeddingGemma-300m model (google/embeddinggemma-300m) is a gated model on HuggingFace that requires authentication via HF_TOKEN. Due to CI/CD authentication limitations, Gemma support was disabled in work related to Issue #573 to allow tests to pass without the gated model.

Now that the maintainer has configured HF_TOKEN in the CI environment, we can restore full Gemma embedding model support.

---

Scope of Changes

1. Model Download Configuration (tools/make/models.mk)

Current State:

• download-models-minimal excludes Gemma (line 29, 63-64)
• Comment states: "Gemma is gated and requires HF_TOKEN, so it's excluded from CI"

Required Changes:

• Add embeddinggemma-300m to download-models-minimal target
• Update comments to reflect that HF_TOKEN is now available

---

2. Go Test Constants (candle-binding/semantic-router_test.go)

Current State:

const (
    GemmaEmbeddingModelPath = "" // Gemma is gated, not used in CI tests (line 1641)
)

Test Skip (lines 1704-1707):

t.Run("InitGemmaOnly", func(t *testing.T) {
    t.Skip("Skipping Gemma-only test: Gemma is a gated model requiring HF_TOKEN")
})

Required Changes:

• Set GemmaEmbeddingModelPath = "../models/embeddinggemma-300m"
• Remove t.Skip() from InitGemmaOnly test
• Enable any other Gemma-related tests currently skipped

---

3. E2E Profile Configurations

Files to Update:

File	Current gemma_model_path	Required Change
e2e/profiles/ai-gateway/values.yaml	Not present (using bert only)	Add Gemma model path
e2e/profiles/dynamic-config/values.yaml	"" (empty, line 130)	"models/embeddinggemma-300m"
e2e/profiles/routing-strategies/values.yaml	Already configured	Verify works with HF_TOKEN

Environment Variable Override (dynamic-config):
env:

• name: EMBEDDING_MODEL_OVERRIDE
  value: "qwen3" # Force qwen3 for tests (Gemma requires HF_TOKEN)- Remove EMBEDDING_MODEL_OVERRIDE or set to "auto" to use intelligent model selection

---

4. initContainer Model Downloads (e2e/profiles/*/values.yaml)

Required Changes:
Add Gemma model to initContainer models list in relevant profiles:

initContainer:
  models:
    # ... existing models ...
    - name: embeddinggemma-300m
      repo: google/embeddinggemma-300m

---

5. Rust Test Fixtures (candle-binding/src/test_fixtures.rs)

Current State:

• GEMMA_EMBEDDING_300M constant is defined (line 50)
• gemma_embedding_model() fixture exists and attempts to load the model
• Tests will panic if model is not available

Required Changes:

• Verify gemma_embedding_model() fixture works with downloaded model
• Enable all Gemma-related Rust tests in:
  • candle-binding/src/model_architectures/embedding/gemma_embedding_test.rs
  • candle-binding/src/model_architectures/embedding/gemma3_model_test.rs

---

6. GitHub Actions Workflow (.github/workflows/integration-test-k8s.yml)

Required Changes:

• Ensure HF_TOKEN secret is passed to the workflow environment
• Add HF_TOKEN to model download steps if not already present:

env:
  HF_TOKEN: ${{ secrets.HF_TOKEN }}

---

7. Quickstart Script (scripts/quickstart.sh)

Current State (lines 182-188):

Check if failure was due to gated model (embeddinggemma-300m)

if grep -q "embeddinggemma.*401|embeddinggemma.*Unauthorized|embeddinggemma.*GatedRepoError" ...Required Changes:

• Update fallback message to indicate HF_TOKEN may not be set
• Or remove fallback if Gemma download is now expected to succeed

---

Acceptance Criteria

• make download-models-minimal successfully downloads embeddinggemma-300m
• Go tests for Gemma embedding pass without skips
• Rust tests for GemmaEmbeddingModel pass (cosine similarity ≥ 0.99 vs Python reference)
• E2E tests with embedding_model: "auto" correctly route to Gemma for short texts
• E2E tests with embedding_model: "gemma" work correctly
• CI/CD pipeline (integration-test-k8s.yml) passes with Gemma enabled
• Documentation updated to reflect Gemma availability

---

Related Issues

• [Testing] improve and enable the skipped testing cases #573 - Original issue that disabled Gemma support

[liavweiss]

/assign