Regenerate tiny gemma4 from google/gemma-4-E2B-it with non-degenerate outputs

Rebuilt with wider output-side initialization (std=0.2 on embed_tokens, embed_tokens_per_layer and the vision/audio embedding projections) so the logits no longer sit near a constant vector.

Measured on an image+text prompt, previous revision -> this one:
logit std 0.113 -> 1.130
min per-position std 0.113 -> 1.126
distinct argmax positions 24/273 -> 65/273
sensitivity to the image 0.060 -> 4.66
NaN / Inf 0 / 0 -> 0 / 0

Config fixes toward the original architecture: num_key_value_heads 2->1 (the original is multi-query), text global_head_dim 8->16 (the original keeps head_dim != global_head_dim), vision global_head_dim 64->8 (was left at the full-size value while head_dim was 8), hidden_size_per_layer_input 1->4, vision default_output_length 70->280.

model_type, architectures, sub-config types, special token ids, vocab_size and the layer counts (text 3 / vision 1 / audio 1) are unchanged, so the expected INT8 node counts in tests/openvino/utils_tests.py still hold.

Validated under ONEDNN_MAX_CPU_ISA=AVX2 (the CI runners' ISA): OpenVINO export, forward and generate all pass, and 6/6 gemma4 tests in test_seq2seq.py + test_quantization.py pass.

config.json

@@ -66,11 +66,11 @@
     "eos_token_id": 1,
     "expert_intermediate_size": null,
     "final_logit_softcapping": 30.0,
-    "global_head_dim": 8,
+    "global_head_dim": 16,
     "head_dim": 8,
     "hidden_activation": "gelu_pytorch_tanh",
     "hidden_size": 32,
-    "hidden_size_per_layer_input": 1,
+    "hidden_size_per_layer_input": 4,
     "initializer_range": 0.02,
     "intermediate_size": 64,
     "layer_types": [
@@ -85,7 +85,7 @@
     "num_experts": null,
     "num_global_key_value_heads": null,
     "num_hidden_layers": 3,
-    "num_key_value_heads": 2,
+    "num_key_value_heads": 1,
     "num_kv_shared_layers": 1,
     "pad_token_id": 0,
     "rms_norm_eps": 1e-06,
@@ -110,7 +110,7 @@
     "vocab_size_per_layer_input": 262144
   },
   "tie_word_embeddings": true,
-  "transformers_version": "5.5.0",
+  "transformers_version": "5.5.4",
   "video_token_id": 258884,
   "vision_config": {
     "_name_or_path": "",
@@ -118,9 +118,9 @@
     "attention_bias": false,
     "attention_dropout": 0.0,
     "chunk_size_feed_forward": 0,
-    "default_output_length": 70,
+    "default_output_length": 280,
     "dtype": "float32",
-    "global_head_dim": 64,
+    "global_head_dim": 8,
     "head_dim": 8,
     "hidden_activation": "gelu_pytorch_tanh",
     "hidden_size": 32,

generation_config.json

@@ -8,6 +8,6 @@
   "output_attentions": false,
   "output_hidden_states": false,
   "pad_token_id": 0,
-  "transformers_version": "5.5.0",
+  "transformers_version": "5.5.4",
   "use_cache": true
 }

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c2645ac53c3989c11a75141407c3016e1c732a38eb306ea75fb9b53d7b64b191
-size 40017352
+oid sha256:3f9988219ad792099f3cb4350a70e107f80d5804d75d4b56208fc8f7b2d9852e
+size 49467948

processor_config.json

@@ -13,7 +13,7 @@
     "max_frequency": 8000.0,
     "mel_floor": 0.001,
     "min_frequency": 0.0,
-    "padding_side": "left",
+    "padding_side": "right",
     "padding_value": 0.0,
     "per_bin_mean": null,
     "per_bin_stddev": null,

tokenizer_config.json

@@ -137,6 +137,5 @@
   "str_token": "<|tool_response>",
   "think_token": "<|think|>",
   "tokenizer_class": "GemmaTokenizer",
-  "truncation_side": "left",
   "unk_token": "<unk>"
 }