Add Gemma 4 E2B vision-language Python export

Author: carinapeng

models/gemma4/gemma4_4bit_minval.yaml

@@ -0,0 +1,36 @@
+quantization_config:
+  execution_mode: eager
+
+  # By default, 4-bit quantize all weights
+  global_config:
+    op_state_spec:
+      weight:
+        dtype: uint4
+        qscheme: asymmetric
+        granularity:
+          type: per_block
+          block_size: 32
+        qformulation: minval
+
+    # No activation quantization
+    op_input_spec: null
+    op_output_spec: null
+
+  # Do not quantize these modules at all
+  module_type_configs:
+    "coreai_models.primitives.macos.sdpa.SDPA": null
+    "coreai_models.models.macos.gemma4_text._ProportionalRoPE": null
+    "coreai_models.models.macos.gemma4_text.MoERouter": null
+    "coreai_models.primitives.macos.rope.RoPE": null
+    "coreai_models.primitives.macos.rms_norm.RMSNorm": null
+    "coreai_models.primitives.macos.switch.SwitchLinear":
+      module_state_spec:
+        weight:
+          dtype: uint4
+          qscheme: asymmetric
+          granularity:
+            type: per_block
+            block_size: [1, 1, 1, 32]
+          qformulation: minval
+      op_input_spec: null
+      op_output_spec: null

python/pyproject.toml

@@ -34,8 +34,8 @@ dependencies = [
     "numpy>=2.2,<3.0",
     "tqdm>=4.67,<5.0",
     "rich>=14.0,<15.0",
-    "transformers>=4.57,<5.0",
-    "huggingface-hub>=0.34,<1.0",
+    "transformers==5.10.2",
+    "huggingface-hub",
     "safetensors>=0.5,<1.0",
     "sentencepiece>=0.2,<1.0",
     "tokenizers>=0.22,<1.0",

python/src/coreai_models/export/_constants.py

@@ -5,10 +5,17 @@
 
 """Constants for the export pipeline."""
 
-# KV cache names used by the Swift runner
+# KV cache names used by the Swift runner. keyCache/valueCache always name the
+# growable, full-attention-equivalent cache (the only cache for single-cache
+# models; the full-attention cache for dual-cache models like Gemma4).
 KEY_CACHE_NAME = "keyCache"
 VALUE_CACHE_NAME = "valueCache"
 
+# Secondary KV cache names for models with dual caches (e.g. Gemma4's bounded
+# sliding-window cache, which is distinct from its full-attention cache above).
+KEY_CACHE_SLIDING_NAME = "slidingKeyCache"
+VALUE_CACHE_SLIDING_NAME = "slidingValueCache"
+
 # Trace-time KV cache sequence length. Used only for export/quantization tracing
 # to bound peak memory; at inference the actual cache size is determined
 # dynamically.