llama/compat: add gemma4 clip handler (gemma4v projector)

Translates Ollama's monolithic gemma4 GGUF into the upstream gemma4v
mmproj shape so `--mmproj <same-blob>` loads alongside the text model.

KV synthesis: rewrite arch to `clip` with `clip.vision.projector_type=
gemma4v`, copy `gemma4.vision.*` → `clip.vision.*`, project
`gemma4.embedding_length` → `clip.vision.projection_dim` (= mm.input_projection
output dim). gemma4 vision uses image_mean=[0,0,0] / image_std=[1,1,1]
(the model does its own per-image normalization via v.std_bias /
v.std_scale tensors), unlike the [0.5,0.5,0.5] defaults most arches
use — inject those.

No tensor renames needed: every `v.*` and `mm.*` tensor name (incl.
v.patch_embd, v.position_embd, v.std_bias/scale, mm.input_projection,
plus the per-block ln1/ln2/attn_post_norm/ffn_post_norm/attn_q_norm/
attn_k_norm) already matches the upstream-converted reference mmproj
exactly.

F32 promote of v.patch_embd.weight (Metal IM2COL requirement, same as
gemma3 / mistral3 / deepseek-ocr).

Adds `gemma4` to the Go-side `compatClipArches` allowlist so
`ollama run gemma4 -i image.jpg ...` auto-passes the same blob as
`--mmproj`. Covers all four published sizes (E2B, E4B, 26B-A4B MoE,
31B dense).

Verified loading: 1189 tensors loaded from monolithic blob, hparams
(projector=gemma4v, n_embd=1152, n_head=16, n_ff=4304, n_layer=27,
n_merge=3, image_size=224, patch_size=16) match migration mmproj
exactly. Image encoding/decoding completes (264 vision tokens). End-
to-end testing through llama-server's chat completions surfaces a
chat-template detection issue in llama-server (Ollama's gemma4 GGUFs
ship without `tokenizer.chat_template`, so llama-server falls back to
ChatML which the model doesn't understand) — orthogonal to compat;
`ollama run` supplies the correct template via Modelfile.

llama/compat/llama-ollama-compat.cpp

@@ -823,6 +823,48 @@ void handle_deepseekocr_clip(gguf_context * meta, ggml_context * ctx) {
     promote_tensor_to_f32(meta, ctx, "v.position_embd.weight");
 }
 
+// =========================================================================
+// gemma4 (clip side — gemma4v projector)
+// =========================================================================
+//
+// Ollama's monolithic gemma4 GGUF embeds a SigLIP-style ViT plus the
+// gemma4v projector (a single `mm.input_projection`). All v.* / mm.*
+// tensor names already match upstream's PROJECTOR_TYPE_GEMMA4V — this
+// handler only needs KV translation and an F32 promote of the patch
+// embedding (Metal IM2COL).
+//
+// gemma4 vision uses image normalization mean=[0,0,0] / std=[1,1,1]
+// (the LM does its own per-image normalization via v.std_bias /
+// v.std_scale tensors) — different from the [0.5,0.5,0.5] used by
+// most other arches.
+
+void handle_gemma4_clip(gguf_context * meta, ggml_context * ctx) {
+    LLAMA_LOG_INFO("%s: detected Ollama-format gemma4 GGUF used as mmproj; translating\n", __func__);
+
+    copy_u32_kv(meta, "gemma4.vision.block_count",                   "clip.vision.block_count");
+    copy_u32_kv(meta, "gemma4.vision.embedding_length",              "clip.vision.embedding_length");
+    copy_u32_kv(meta, "gemma4.vision.feed_forward_length",           "clip.vision.feed_forward_length");
+    copy_u32_kv(meta, "gemma4.vision.attention.head_count",          "clip.vision.attention.head_count");
+    copy_f32_kv(meta, "gemma4.vision.attention.layer_norm_epsilon",  "clip.vision.attention.layer_norm_epsilon");
+    copy_u32_kv(meta, "gemma4.vision.patch_size",                    "clip.vision.patch_size");
+    // gemma4 vision is fixed at 224x224 patches.
+    inject_u32_if_missing(meta, "clip.vision.image_size", 224);
+    // projection_dim = LM embedding length.
+    copy_u32_kv(meta, "gemma4.embedding_length",                     "clip.vision.projection_dim");
+
+    static const float kZeros[3] = {0.0f, 0.0f, 0.0f};
+    static const float kOnes [3] = {1.0f, 1.0f, 1.0f};
+    inject_f32_arr_if_missing(meta, "clip.vision.image_mean", kZeros, 3);
+    inject_f32_arr_if_missing(meta, "clip.vision.image_std",  kOnes,  3);
+
+    inject_bool_if_missing(meta, "clip.has_vision_encoder", true);
+    gguf_set_val_str(meta, "clip.vision.projector_type", "gemma4v");
+    gguf_set_val_str(meta, "general.architecture",       "clip");
+
+    // Metal IM2COL needs F32 patch_embd weights (same as other arches).
+    promote_tensor_to_f32(meta, ctx, "v.patch_embd.weight");
+}
+
 // =========================================================================
 // llama4 (clip side)
 // =========================================================================
@@ -1104,6 +1146,10 @@ void translate_clip_metadata(gguf_context * meta, ggml_context * ctx) {
         handle_llama4_clip(meta, ctx);
         return;
     }
+    if (detect_ollama_gemma4(meta, ctx)) {
+        handle_gemma4_clip(meta, ctx);
+        return;
+    }
 }
 
 bool should_skip_tensor(const llama_model_loader * ml, const char * tensor_name) {

llm/llama_server.go

@@ -434,6 +434,7 @@ func NewLlamaServerRunner(
 	// the compat layer's clip-side coverage in llama/compat/.
 	compatClipArches := map[string]bool{
 		"gemma3":      true,
+		"gemma4":      true,
 		"qwen35moe":   true,
 		"mistral3":    true,
 		"deepseekocr": true,