feat(gemma): add gemma-3-270m and gemma-3-12b manifests and refuse fp16 for Gemma 3 (1327)

## Implementation

Manifests only for the two widths. The one code change is the fp16 refusal.

Author: zhenshanx-nv

apps/benchmark/performance/release.yaml

@@ -88,6 +88,8 @@ excluded_profiles:
       Functional and Hugging Face reference-parity qualification is present,
       but the release-performance workload and receipt were collected only for
       gemma-2-2b, which exercises the same builder and runtime path.
+  - model: gemma-3-270m
+    reason: *gemma3_performance_exclusion
   - model: gemma-3-1b
     reason: >-
       Functional and Hugging Face reference-parity qualification is present,

families/gemma/model.py

@@ -249,6 +249,20 @@ def _checkpoint_gated_activation(config: ModelConfig) -> str:
 # this family has, so the prefix let them build a full-attention graph and
 # generate quietly wrong text rather than being refused.
 _GEMMA3_MODEL_TYPES = frozenset({"gemma3", "gemma3_text"})
+# Gemma 3 activations do not fit fp16. Running the reference in fp32 and taking
+# the largest absolute value leaving each decoder layer, against the fp16
+# maximum of 65504:
+#
+#   gemma-3-270m   peak 102956   11 of 18 layers over
+#   gemma-3-1b     peak  61040    0 of 26 layers over (a 7% margin)
+#   gemma-3-4b     peak 298680   29 of 34 layers over
+#
+# The two that overflow emit token 0 repeatedly. The 1B stays inside the range
+# on one prompt by 7%, which is luck rather than headroom, so fp16 is refused
+# for the generation rather than per width. Gemma 2 peaks at 4060 and keeps it.
+_FP16_UNSAFE_MODEL_TYPES = _GEMMA3_MODEL_TYPES
+
+
 _SUPPORTED_MODEL_TYPES = frozenset({"gemma", "gemma2"}) | _GEMMA3_MODEL_TYPES
 
 
@@ -345,6 +359,12 @@ def build(request: "BuildRequest", writer: "BundleWriter") -> None:
     precision = str(request.precision).lower()
     if precision not in {"fp32", "fp16", "bf16"}:
         raise ValueError("Gemma precision must be fp32, fp16, or bf16")
+    if precision == "fp16" and str(config.model_type).lower() in _FP16_UNSAFE_MODEL_TYPES:
+        raise NotImplementedError(
+            f"gemma does not support fp16 for {config.model_type!r}: its activations "
+            "exceed the fp16 range and the engine returns a single repeated token; "
+            "use bf16 or fp32"
+        )
     max_sequence_length = _positive_int(
         request.max_sequence_length or min(config.max_position_embeddings, 256),
         "max_sequence_length",

families/gemma/tests/manifests/gemma-3-270m.json

@@ -0,0 +1,22 @@
+{
+  "name": "gemma-3-270m",
+  "hf_id": "google/gemma-3-270m-it",
+  "bundle": "gemma-3-270m.bundle",
+  "family": "gemma",
+  "task": "text_generation",
+  "precision": "bf16",
+  "trust_remote_code": false,
+  "testcases": [
+    {
+      "name": "gemma-3-270m",
+      "premerge": true,
+      "reference_precision": "fp32",
+      "prompt": "What is the capital of France? Answer with just the name.",
+      "max_new_tokens": 8,
+      "use_chat_template": true,
+      "enable_thinking": false
+    }
+  ],
+  "max_sequence_length": 256,
+  "tensor_parallel_size": 1
+}

families/gemma/tests/test_model_type_gate.py

@@ -96,3 +96,39 @@ def test_the_supported_generations_pass_the_gate(tmp_path: Path) -> None:
         with pytest.raises(Exception) as caught:  # noqa: PT011 - any later failure will do
             _build(directory)
         assert "does not support model_type" not in str(caught.value)
+
+
+def test_gemma3_refuses_fp16(tmp_path: Path) -> None:
+    """Gemma 3 activations exceed the fp16 range, so fp16 is refused.
+
+    Measured on the reference in fp32, largest absolute value leaving a decoder
+    layer against the fp16 maximum of 65504: gemma-3-270m peaks at 102956 and
+    gemma-3-4b at 298680, both of which overflow and make the engine emit token
+    0 repeatedly. gemma-3-1b peaks at 61040, inside the range by 7%, which is
+    luck rather than headroom.
+    """
+    for model_type in ("gemma3", "gemma3_text"):
+        directory = _model_dir(tmp_path / f"fp16-{model_type}", model_type)
+        with pytest.raises(NotImplementedError, match="does not support fp16"):
+            _build_with(directory, precision="fp16")
+
+
+def test_gemma2_keeps_fp16(tmp_path: Path) -> None:
+    """Gemma 2 peaks at 4060, sixteen times inside the fp16 range."""
+    directory = _model_dir(tmp_path / "fp16-gemma2", "gemma2")
+    with pytest.raises(Exception) as caught:  # noqa: PT011 - a later failure is fine
+        _build_with(directory, precision="fp16")
+    assert "does not support fp16" not in str(caught.value)
+
+
+def _build_with(model_dir: Path, *, precision: str) -> None:
+    build_family(
+        BuildRequest(
+            model_dir=model_dir,
+            output_path=model_dir / "out.bundle",
+            family="gemma",
+            task="text_generation",
+            precision=precision,
+        ),
+        writer=None,
+    )