From ffee1693a64ecfe32783751f49f9d5019f9575cc Mon Sep 17 00:00:00 2001
From: Albert Villanova del Moral
 <8515462+albertvillanova@users.noreply.github.com>
Date: Fri, 28 Aug 2026 15:33:09 +0200
Subject: [PATCH 1/2] Pin layer_types to one of each attention type for tiny
 Gemma3

---
 .../gemma3_for_conditional_generation.py                      | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/scripts/generate_tiny_models/for_conditional_generation/gemma3_for_conditional_generation.py b/scripts/generate_tiny_models/for_conditional_generation/gemma3_for_conditional_generation.py
index 2b3c97894ab..68be53d8f0d 100644
--- a/scripts/generate_tiny_models/for_conditional_generation/gemma3_for_conditional_generation.py
+++ b/scripts/generate_tiny_models/for_conditional_generation/gemma3_for_conditional_generation.py
@@ -34,7 +34,9 @@
     "hidden_size": 16,
     "num_attention_heads": 4,
     "num_key_value_heads": 2,
-    "layer_types": None,  # Set it automatically from num_hidden_layers
+    # One of each attention type. Deriving the pattern from num_hidden_layers=2 gives two
+    # sliding layers (the reference is 5:1 over 34 layers), so the global RoPE path would never run.
+    "layer_types": ["sliding_attention", "full_attention"],
     "intermediate_size": 32,
 }
 vision_config = {

From f2290673cb131a1d2198eed97f3c1a67001b82d1 Mon Sep 17 00:00:00 2001
From: Albert Villanova del Moral
 <8515462+albertvillanova@users.noreply.github.com>
Date: Fri, 28 Aug 2026 15:33:12 +0200
Subject: [PATCH 2/2] Pin layer_types to one of each attention type for tiny
 Olmo3

---
 .../generate_tiny_models/for_causal_lm/olmo3_for_causal_lm.py  | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/scripts/generate_tiny_models/for_causal_lm/olmo3_for_causal_lm.py b/scripts/generate_tiny_models/for_causal_lm/olmo3_for_causal_lm.py
index 719f8e22af9..546a4484327 100644
--- a/scripts/generate_tiny_models/for_causal_lm/olmo3_for_causal_lm.py
+++ b/scripts/generate_tiny_models/for_causal_lm/olmo3_for_causal_lm.py
@@ -39,6 +39,9 @@
     num_key_value_heads=2,
     num_hidden_layers=2,
     intermediate_size=32,
+    # One of each attention type. Deriving the pattern from num_hidden_layers=2 gives two
+    # sliding layers (the reference is 3:1 over 32 layers), so the global RoPE path would never run.
+    layer_types=["sliding_attention", "full_attention"],
     # Non-size fields kept aligned with the reference so the tiny config only differs in what we scale down.
     max_position_embeddings=65536,
     rms_norm_eps=1e-06,