onboard google/gemma-4-12b as a verified Gemma 4 VLM

ARCHITECTURE.md

@@ -57,10 +57,11 @@ CPU (see Multimodal VLM Path below).
 | Granite Vision 4.1 4B | granite4\_vision | SigLIP vision + Granite text | Yes (padded) | Yes | Yes | Yes |
 | Mistral-Small-3.1-24B-Instruct-2503 | mistral3 | Pixtral + Mistral text | Yes (padded) | Yes | Yes | Yes |
 | Ministral-3-14B-Instruct-2512 (bf16) | mistral3 | Pixtral + Ministral3 text | Yes (padded) | Yes | Yes | Yes |
-| Gemma 4 12B (bf16) | gemma4\_unified | Encoder-free vision embedder + Gemma 4 text | Yes | Yes | Yes | Yes |
+| Gemma 4 12B IT (bf16) | gemma4\_unified | Encoder-free vision embedder + Gemma 4 text | Yes | Yes | Yes | Yes |
+| Gemma 4 12B Base (bf16) | gemma4\_unified | Encoder-free vision embedder + Gemma 4 text | Yes | Yes | Yes | Yes |
 
 **CPU Accurate** = adapter `generate` matches stock `model.generate` token-for-token on CPU (`test_vlm_e2e_cpu.py`).
-**Spyre Runs** = `test_vlm_e2e_spyre.py` drives the adapter teacher-forced on stock's tokens and asserts per-step logit cosine ≥ 0.999 vs the CPU reference over prefill + decode steps (top-1 agreement is reported, not asserted — an open-ended caption hits near-ties where the fp16-substrate winner is numerically arbitrary; see Multimodal VLM Path). granite-vision-4.1 holds cosine ≥ 0.99991 at every step and produces a correct, coherent caption. Gemma 4 12B runs in **bf16** (like the rest of the Gemma family, it overflows its residual stream in fp16) and holds cosine ≥ 0.99964 at every step with 5/5 top-1 agreement, producing a caption byte-identical to stock.
+**Spyre Runs** = `test_vlm_e2e_spyre.py` drives the adapter teacher-forced on stock's tokens and asserts per-step logit cosine ≥ 0.999 vs the CPU reference over prefill + decode steps (top-1 agreement is reported, not asserted — an open-ended caption hits near-ties where the fp16-substrate winner is numerically arbitrary; see Multimodal VLM Path). granite-vision-4.1 holds cosine ≥ 0.99991 at every step and produces a correct, coherent caption. Gemma 4 12B IT runs in **bf16** (like the rest of the Gemma family, it overflows its residual stream in fp16) and holds cosine ≥ 0.99856 at every step with 5/5 top-1 agreement, producing a caption byte-identical to stock. Gemma 4 12B Base routes via the same adapter; its causal-LM path (text-decoder only) holds 5/5 top-1 on Spyre.
 
 ### Embedding
 
@@ -142,7 +143,7 @@ pattern, norms, and weight layout.
 | hf\_granitemoehybrid.py | granitemoehybrid | 2 | Granite 4.0 Micro |
 | hf\_smollm3.py | smollm3 | 1 | — |
 | hf\_gemma4.py | gemma4\_unified / gemma4 (dense) | 3 | Not E2B/E4B (PLE) or 26B-A4B (MoE). |
-| hf\_gemma4\_mm.py | gemma4\_unified (multimodal) | 1 | Gemma 4 31B (dense unified VLM). Not E2B/E4B (PLE) or 26B-A4B (MoE). |
+| hf\_gemma4\_mm.py | gemma4\_unified (multimodal) | 2 | Gemma 4 31B (dense unified VLM). Not E2B/E4B (PLE) or 26B-A4B (MoE). |
 | hf\_gemma3.py | gemma3\_text / gemma3 (dense) | 2 | Gemma 3 4B/12B/27B (text decoder of the multimodal checkpoints); EmbeddingGemma (bidirectional embedder). Not Gemma 3n (PLE). |
 | hf\_olmo.py | olmo | 1 | OLMo 7B |
 | hf\_olmo2.py | olmo2 | 1 | OLMo 2 7B |

hf_adapters/hf_gemma4_mm.py

@@ -15,6 +15,10 @@
 """
 Unified (encoder-free) HuggingFace adapter for Gemma 4 12B on Spyre — image→text.
 
+Supports both the base checkpoint (``google/gemma-4-12b``) and the instruction-tuned
+variant (``google/gemma-4-12B-it``); both use ``model_type=gemma4_unified`` and
+``Gemma4UnifiedForConditionalGeneration``.
+
 Where ``hf_gemma4`` runs only the text decoder (``AutoSpyreModelForCausalLM``),
 this module loads the full unified multimodal model
 (``Gemma4UnifiedForConditionalGeneration``, ``model_type=gemma4_unified``) via

tests/model_registry.py

@@ -436,8 +436,12 @@ def _include_gated() -> bool:
         "size": "3b",
     },
     # hf_gemma4_mm.py — unified encoder-free VLM (image + text -> text)
+    # Note: google/gemma-4-12b (base, no chat template) also resolves via this
+    # adapter (Gemma4UnifiedConfig -> hf_gemma4_mm) but is tested via the
+    # causal-LM path (gemma4_base in CAUSAL_LM_MODELS); the VLM harness requires
+    # apply_chat_template, which the base model does not provide.
     "gemma4_mm": {
-        "name": "Gemma 4 12B (unified VLM)",
+        "name": "Gemma 4 12B IT (unified VLM)",
         "path": "google/gemma-4-12B-it",
         "adapter": "hf_gemma4_mm.py",
         "kind": "vlm",  # multimodal: image + text -> generated text
@@ -554,13 +558,12 @@ def _non_blocking(models: dict[str, dict], keys: tuple[str, ...]) -> dict[str, s
         "ministral3",
         "pythia_410m",
         "gemma4_google",
-        "gemma4_base",
     ),
 )
 
 NON_BLOCKING_VISION_MODELS: dict[str, str] = _non_blocking(
     VISION_MODELS,
-    ("gemma4_mm",),
+    (),
 )