diff --git a/docs/assets/gemma4/gemma4-12b-cp1-200steps.png b/docs/assets/gemma4/gemma4-12b-cp1-200steps.png
new file mode 100644
index 00000000000..1d5cec883b4
Binary files /dev/null and b/docs/assets/gemma4/gemma4-12b-cp1-200steps.png differ
diff --git a/docs/assets/gemma4/gemma4-26ba4b-cp1-vs-cp2-100steps.png b/docs/assets/gemma4/gemma4-26ba4b-cp1-vs-cp2-100steps.png
new file mode 100644
index 00000000000..56ada3e23e9
Binary files /dev/null and b/docs/assets/gemma4/gemma4-26ba4b-cp1-vs-cp2-100steps.png differ
diff --git a/docs/assets/gemma4/gemma4-31b-cp1-vs-cp2-100steps.png b/docs/assets/gemma4/gemma4-31b-cp1-vs-cp2-100steps.png
new file mode 100644
index 00000000000..571cf4f3ab1
Binary files /dev/null and b/docs/assets/gemma4/gemma4-31b-cp1-vs-cp2-100steps.png differ
diff --git a/docs/assets/gemma4/gemma4-e2b-cp1-vs-cp2-100steps.png b/docs/assets/gemma4/gemma4-e2b-cp1-vs-cp2-100steps.png
new file mode 100644
index 00000000000..84fb1757e53
Binary files /dev/null and b/docs/assets/gemma4/gemma4-e2b-cp1-vs-cp2-100steps.png differ
diff --git a/docs/guides/models/gemma/gemma4.md b/docs/guides/models/gemma/gemma4.md
new file mode 100644
index 00000000000..b9dbbfdc848
--- /dev/null
+++ b/docs/guides/models/gemma/gemma4.md
@@ -0,0 +1,148 @@
+# Gemma 4
+
+NeMo RL supports post-training the Gemma 4 family with the AutoModel training
+backend and vLLM generation. The reference configurations cover text-only DAPO
+for the E2B, 12B, 26B-A4B, and 31B variants and multimodal GRPO for the E4B
+variant.
+
+> [!IMPORTANT]
+> **Status: Functionally Ready.** The listed configurations provide short-run
+> functional and CI coverage. They are starting points for post-training, not a
+> claim of long-run convergence on every model and parallel layout.
+
+## Support Matrix
+
+| Model | Task | Training backend | Training parallelism | Generation backend | Status |
+| --- | --- | --- | --- | --- | --- |
+| `google/gemma-4-E2B-it` | Text-only DAPO | AutoModel (FSDP2) | CP1 or CP2 | vLLM | Functionally Ready |
+| `google/gemma-4-E4B-it` | VLM GRPO | AutoModel (FSDP2) | CP1 | vLLM TP4 | Functionally Ready |
+| `google/gemma-4-12B-it` | Text-only DAPO | AutoModel (FSDP2) | CP1 | vLLM TP2 | Functionally Ready |
+| `google/gemma-4-26B-A4B-it` | Text-only DAPO | AutoModel (FSDP2) | EP32 with CP1; EP16 with CP2 | vLLM TP4 | Functionally Ready |
+| `google/gemma-4-31B-it` | Text-only DAPO | AutoModel (FSDP2) | CP1 or CP2 | vLLM TP4 | Functionally Ready |
+
+## Reference Recipes
+
+Recipe YAML files under `examples/configs/recipes/` are the source of truth.
+
+| Model | Layout | Nodes and GPUs | Recipe |
+| --- | --- | --- | --- |
+| E2B | FSDP2, CP1 | 1n8g | [`dapo-gemma4-e2b-it-1n8g-fsdp2-automodel.yaml`](../../../../examples/configs/recipes/llm/dapo-gemma4-e2b-it-1n8g-fsdp2-automodel.yaml) |
+| E2B | FSDP2, CP2 | 1n8g | [`dapo-gemma4-e2b-it-1n8g-fsdp2cp2-automodel.yaml`](../../../../examples/configs/recipes/llm/dapo-gemma4-e2b-it-1n8g-fsdp2cp2-automodel.yaml) |
+| E4B | VLM FSDP2, CP1 | 1n8g | [`vlm_grpo-gemma4-e4b-geo3k-1n8g-automodel.yaml`](../../../../examples/configs/recipes/vlm/vlm_grpo-gemma4-e4b-geo3k-1n8g-automodel.yaml) |
+| 12B | FSDP2, CP1 | 2n8g | [`dapo-gemma4-12b-it-2n8g-fsdp2-automodel.yaml`](../../../../examples/configs/recipes/llm/dapo-gemma4-12b-it-2n8g-fsdp2-automodel.yaml) |
+| 26B-A4B | FSDP2, EP32, CP1 | 4n8g | [`dapo-gemma4-26ba4b-it-4n8g-fsdp2-automodel.yaml`](../../../../examples/configs/recipes/llm/dapo-gemma4-26ba4b-it-4n8g-fsdp2-automodel.yaml) |
+| 26B-A4B | FSDP2, EP16, CP2 | 4n8g | [`dapo-gemma4-26ba4b-it-4n8g-fsdp2ep16cp2-automodel.yaml`](../../../../examples/configs/recipes/llm/dapo-gemma4-26ba4b-it-4n8g-fsdp2ep16cp2-automodel.yaml) |
+| 31B | FSDP2, CP1 | 4n8g | [`dapo-gemma4-31b-it-4n8g-fsdp2-automodel.yaml`](../../../../examples/configs/recipes/llm/dapo-gemma4-31b-it-4n8g-fsdp2-automodel.yaml) |
+| 31B | FSDP2, CP2 | 4n8g | [`dapo-gemma4-31b-it-4n8g-fsdp2cp2-automodel.yaml`](../../../../examples/configs/recipes/llm/dapo-gemma4-31b-it-4n8g-fsdp2cp2-automodel.yaml) |
+
+## Run a Recipe
+
+From an allocation matching the recipe, launch the standard GRPO entry point.
+For example, run the 12B configuration with:
+
+```bash
+uv run examples/run_grpo.py \
+  --config examples/configs/recipes/llm/dapo-gemma4-12b-it-2n8g-fsdp2-automodel.yaml
+```
+
+See the [GRPO guide](../../grpo.md) for algorithm and launch details.
+
+## Context Parallel
+
+Context Parallel support uses the refactored AutoModel CP interface introduced
+by [PR #3498](https://github.com/NVIDIA-NeMo/RL/pull/3498). The E2B and 31B CP2
+recipes are thin overrides of their CP1 parents that change
+`policy.dtensor_cfg.context_parallel_size` to `2`. The 26B-A4B CP2 recipe also
+changes expert parallelism from EP32 to EP16 so the model-parallel product fits
+on the same four-node allocation. Other model, optimizer, sequence-length, and
+generation settings remain identical to their CP1 parents.
+
+Context Parallel currently applies to the text-only E2B, 26B-A4B, and 31B
+recipes backed by AutoModel's Gemma 4 model-owned attention. Do not enable it
+for the E4B VLM recipe or the 12B unified checkpoint. The text-only recipes
+freeze the vision and audio towers, disable sequence packing, and configure
+vLLM with `language_model_only: true` where the checkpoint requires it.
+
+Training and generation parallelism are independent. CP partitions training
+sequences; the `tensor_parallel_size` under `policy.generation.vllm_cfg`
+controls vLLM. The AutoModel model-parallel product must divide the training
+world size. The 26B-A4B CP2 recipe uses EP16 × CP2 across 32 GPUs, provided by
+four nodes with eight GPUs per node.
+
+## Validation Curves
+
+The following plots show raw step-level metrics captured from completed
+validation runs. Validation accuracy is plotted only at validation steps; the
+other metrics are plotted at every recorded training step.
+
+### E2B Context Parallel Parity
+
+The E2B CP1 and CP2 runs use the same 100-step training configuration except
+for `policy.dtensor_cfg.context_parallel_size`. Their trajectories remain close
+across the six metrics: the mean absolute difference in validation accuracy is
+0.0083 across six aligned validation points, while the mean absolute difference
+in generation KL error is 1.6e-5 across 100 aligned training points.
+
+![Gemma 4 E2B CP1 versus CP2 curves for train/loss, train/mean_gen_tokens_per_sample, train/reward, validation/accuracy, train/gen_kl_error, and train/approx_entropy](../../../assets/gemma4/gemma4-e2b-cp1-vs-cp2-100steps.png)
+
+### 26B-A4B Context Parallel Parity
+
+The 26B-A4B pair keeps the four-node, 32-GPU training world size fixed while
+changing the model-parallel layout from EP32 × CP1 to EP16 × CP2. Both runs
+complete 100 steps. The mean absolute difference in validation accuracy is
+0.0114 across 20 aligned validation points, while the mean absolute difference
+in generation KL error is 5.6e-5 across 100 aligned training points. The final
+validation accuracies are 0.741 for EP32 × CP1 and 0.751 for EP16 × CP2.
+
+![Gemma 4 26B-A4B EP32 CP1 versus EP16 CP2 curves for train/loss, train/mean_gen_tokens_per_sample, train/reward, validation/accuracy, train/gen_kl_error, and train/approx_entropy](../../../assets/gemma4/gemma4-26ba4b-cp1-vs-cp2-100steps.png)
+
+### 31B Context Parallel Parity
+
+The 31B CP1 and CP2 runs use the same four-node training configuration except
+for `policy.dtensor_cfg.context_parallel_size`. Both runs complete 100 steps.
+The mean absolute difference in validation accuracy is 0.0099 across 20 aligned
+validation points, while the mean absolute difference in generation KL error is
+2.2e-5 across 100 aligned training points. The final validation accuracies are
+0.823 for CP1 and 0.818 for CP2.
+
+![Gemma 4 31B CP1 versus CP2 curves for train/loss, train/mean_gen_tokens_per_sample, train/reward, validation/accuracy, train/gen_kl_error, and train/approx_entropy](../../../assets/gemma4/gemma4-31b-cp1-vs-cp2-100steps.png)
+
+### 12B Long Run
+
+The 12B CP1 run completed 200 steps. Validation accuracy increases from 0.540
+at step 5 to 0.617 at step 200 and reaches a maximum of 0.631. Training loss
+decreases from 0.166 at step 1 to 0.054 at step 200, while generation KL error
+stays below 6.9e-4. The recorded history contains 199 training-metric rows and
+39 validation rows; step 125 has no values for the requested metrics, so the
+plot does not interpolate that missing row.
+
+![Gemma 4 12B CP1 curves for train/loss, train/mean_gen_tokens_per_sample, train/reward, validation/accuracy, train/gen_kl_error, and train/approx_entropy](../../../assets/gemma4/gemma4-12b-cp1-200steps.png)
+
+## 12B Unified Checkpoint
+
+The 12B checkpoint reports `model_type: gemma4_unified` and architecture
+`Gemma4UnifiedForConditionalGeneration`. NeMo RL routes it through the
+image-text AutoModel class so the complete checkpoint can load, while the
+reference recipe trains only the language path:
+
+- The vision and audio towers are frozen.
+- vLLM generation uses `language_model_only: true` and TP2.
+- Weight refits omit the frozen vision/audio tensors because vLLM's text-only
+  unified model uses encoder-free multimodal stubs with a different layout.
+- vLLM tokenizer initialization is enabled for the unified architecture.
+- Activation checkpointing and optimizer offload during log-probability
+  computation are enabled.
+- Sequence packing and Liger kernels are disabled.
+- The total sequence length is 4,096 tokens, with up to 3,072 generated tokens.
+
+## Limitations and Tracking
+
+- Context Parallel is not supported by the Gemma 4 VLM recipe.
+- The 12B `gemma4_unified` checkpoint currently supports CP1 only; unlike the
+  other text variants, it does not use AutoModel's Gemma 4 model-owned CP
+  attention implementation.
+- The recipes disable sequence packing; validate any packing change separately.
+- The functional status does not imply long-run convergence for every variant.
+- Gemma 4 CP support is tracked by
+  [#2914](https://github.com/NVIDIA-NeMo/RL/issues/2914), and 12B support by
+  [#2913](https://github.com/NVIDIA-NeMo/RL/issues/2913).
diff --git a/docs/guides/models/gemma/index.md b/docs/guides/models/gemma/index.md
new file mode 100644
index 00000000000..e24b33fbbbf
--- /dev/null
+++ b/docs/guides/models/gemma/index.md
@@ -0,0 +1,20 @@
+# Gemma
+
+This is the landing page for Gemma model guidance in NeMo RL. It links to
+version-specific pages covering reference recipes, supported parallel layouts,
+and model-specific configuration requirements.
+
+For the full list of supported models, see
+[Model Support](../../../about/model-support.md).
+
+## Version Guides
+
+- **[Gemma 4](gemma4.md)** — text-only DAPO recipes for the E2B, 12B, 26B-A4B,
+  and 31B variants, including AutoModel Context Parallel recipes, plus an E4B
+  VLM GRPO recipe.
+
+```{toctree}
+:hidden:
+
+gemma4.md
+```
diff --git a/docs/guides/models/index.md b/docs/guides/models/index.md
index 5fdc0796e0e..3f688c5683c 100644
--- a/docs/guides/models/index.md
+++ b/docs/guides/models/index.md
@@ -12,6 +12,8 @@ For the full list of supported models, see
 
 - **[GLM](glm/index.md)** — GLM-5.1 and GLM-5.2 GRPO recipes on the Megatron
   backend, colocated and non-colocated with vLLM.
+- **[Gemma](gemma/index.md)** — Gemma 4 text-only DAPO and VLM GRPO recipes on
+  AutoModel, including Context Parallel layouts for the text-only variants.
 - **[Nemotron](nemotron/index.md)** — post-training recipes for Nemotron 3
   Nano, Nano Omni, Super, Ultra, and Nemotron 3.5 Lightning, spanning the
   Megatron and AutoModel backends.
@@ -27,6 +29,7 @@ migrated into this hub as their guidance grows.
 :hidden:
 
 glm/index.md
+gemma/index.md
 nemotron/index.md
 qwen/index.md
 ```
diff --git a/examples/configs/recipes/llm/dapo-gemma4-12b-it-2n8g-fsdp2-automodel.yaml b/examples/configs/recipes/llm/dapo-gemma4-12b-it-2n8g-fsdp2-automodel.yaml
new file mode 100644
index 00000000000..ae53898eaeb
--- /dev/null
+++ b/examples/configs/recipes/llm/dapo-gemma4-12b-it-2n8g-fsdp2-automodel.yaml
@@ -0,0 +1,98 @@
+defaults: ../../grpo_math_1B.yaml
+grpo:
+  batch_multiplier: 3
+  val_period: 5
+  max_val_samples: 960
+  val_batch_size: 960
+  use_leave_one_out_baseline: false
+  use_dynamic_sampling: true
+  reward_scaling:
+    enabled: true
+    target_min: -1.0
+  reward_shaping:
+    enabled: true
+    overlong_buffer_length: 384
+    max_response_length: 3072
+loss_fn:
+  reference_policy_kl_penalty: 0.0
+  use_importance_sampling_correction: true
+  truncated_importance_sampling_type: tis
+  truncated_importance_sampling_ratio: 2
+  ratio_clip_max: 0.28
+  ratio_clip_c: 10
+checkpointing:
+  checkpoint_dir: results/dapo-gemma4-12b-it-2n8g-fsdp2-automodel
+  save_period: 5
+policy:
+  model_name: google/gemma-4-12B-it
+  tokenizer:
+    name: google/gemma-4-12B-it
+  train_micro_batch_size: 1
+  logprob_batch_size: 1
+  max_total_sequence_length: 4096
+  logprob_chunk_size: 4096
+  offload_optimizer_for_logprob: true
+  optimizer:
+    name: transformer_engine.pytorch.optimizers.fused_adam.FusedAdam
+    kwargs:
+      lr: 1.0e-06
+      weight_decay: 0.1
+      master_weights: true
+      store_param_remainders: true
+      exp_avg_dtype: torch.bfloat16
+      exp_avg_sq_dtype: torch.bfloat16
+  scheduler:
+  - name: torch.optim.lr_scheduler.LinearLR
+    kwargs:
+      start_factor: 1.0e-08
+      end_factor: 1.0
+      total_iters: 10
+  - name: torch.optim.lr_scheduler.ConstantLR
+    kwargs:
+      factor: 1.0
+      total_iters: 10000000000
+  - milestones:
+    - 10
+  dtensor_cfg:
+    activation_checkpointing: true
+    automodel_kwargs:
+      use_liger_kernel: false
+      use_sdpa_patching: false
+      freeze_config:
+        freeze_vision_tower: true
+        freeze_audio_tower: true
+        freeze_language_model: false
+  sequence_packing:
+    enabled: false
+  dynamic_batching:
+    enabled: true
+  make_sequence_length_divisible_by: 1
+  generation:
+    max_new_tokens: 3072
+    vllm_cfg:
+      tensor_parallel_size: 2
+      gpu_memory_utilization: 0.5
+    vllm_kwargs:
+      language_model_only: true
+data:
+  max_input_seq_length: 2048
+  train:
+    dataset_name: DAPOMath17K
+  validation:
+    dataset_name: DAPOMathAIME2024
+  default:
+    prompt_file: null
+env:
+  math:
+    num_workers: 16
+    math_verify_impl: dapo_math_verify
+logger:
+  log_dir: logs/dapo-gemma4-12b-it-2n8g-fsdp2-automodel
+  wandb_enabled: true
+  tensorboard_enabled: true
+  wandb:
+    project: nemorl-gemma4
+    name: dapo-gemma4-12b-it-2n8g-fsdp2-automodel
+cluster:
+  gpus_per_node: 8
+  num_nodes: 2
diff --git a/examples/configs/recipes/llm/dapo-gemma4-26ba4b-it-4n8g-fsdp2ep16cp2-automodel.yaml b/examples/configs/recipes/llm/dapo-gemma4-26ba4b-it-4n8g-fsdp2ep16cp2-automodel.yaml
new file mode 100644
index 00000000000..0ae369f9f15
--- /dev/null
+++ b/examples/configs/recipes/llm/dapo-gemma4-26ba4b-it-4n8g-fsdp2ep16cp2-automodel.yaml
@@ -0,0 +1,11 @@
+defaults: ./dapo-gemma4-26ba4b-it-4n8g-fsdp2-automodel.yaml
+policy:
+  dtensor_cfg:
+    expert_parallel_size: 16
+    context_parallel_size: 2
+checkpointing:
+  checkpoint_dir: results/dapo-gemma4-26ba4b-it-4n8g-fsdp2ep16cp2-automodel
+logger:
+  log_dir: logs/dapo-gemma4-26ba4b-it-4n8g-fsdp2ep16cp2-automodel
+  wandb:
+    name: dapo-gemma4-26ba4b-it-4n8g-fsdp2ep16cp2-automodel
diff --git a/examples/configs/recipes/llm/dapo-gemma4-31b-it-4n8g-fsdp2cp2-automodel.yaml b/examples/configs/recipes/llm/dapo-gemma4-31b-it-4n8g-fsdp2cp2-automodel.yaml
new file mode 100644
index 00000000000..d0c013581c8
--- /dev/null
+++ b/examples/configs/recipes/llm/dapo-gemma4-31b-it-4n8g-fsdp2cp2-automodel.yaml
@@ -0,0 +1,10 @@
+defaults: ./dapo-gemma4-31b-it-4n8g-fsdp2-automodel.yaml
+policy:
+  dtensor_cfg:
+    context_parallel_size: 2
+checkpointing:
+  checkpoint_dir: results/dapo-gemma4-31b-it-4n8g-fsdp2cp2-automodel
+logger:
+  log_dir: logs/dapo-gemma4-31b-it-4n8g-fsdp2cp2-automodel
+  wandb:
+    name: dapo-gemma4-31b-it-4n8g-fsdp2cp2-automodel
diff --git a/examples/configs/recipes/llm/dapo-gemma4-e2b-it-1n8g-fsdp2cp2-automodel.yaml b/examples/configs/recipes/llm/dapo-gemma4-e2b-it-1n8g-fsdp2cp2-automodel.yaml
new file mode 100644
index 00000000000..660beca59c2
--- /dev/null
+++ b/examples/configs/recipes/llm/dapo-gemma4-e2b-it-1n8g-fsdp2cp2-automodel.yaml
@@ -0,0 +1,10 @@
+defaults: ./dapo-gemma4-e2b-it-1n8g-fsdp2-automodel.yaml
+policy:
+  dtensor_cfg:
+    context_parallel_size: 2
+checkpointing:
+  checkpoint_dir: results/dapo-gemma4-e2b-it-1n8g-fsdp2cp2-automodel
+logger:
+  log_dir: logs/dapo-gemma4-e2b-it-1n8g-fsdp2cp2-automodel
+  wandb:
+    name: dapo-gemma4-e2b-it-1n8g-fsdp2cp2-automodel
diff --git a/examples/configs/recipes/vlm/vlm_grpo-gemma4-e4b-geo3k-1n8g-automodel.yaml b/examples/configs/recipes/vlm/vlm_grpo-gemma4-e4b-geo3k-1n8g-automodel.yaml
index bed343a4c7d..8793c8a5512 100644
--- a/examples/configs/recipes/vlm/vlm_grpo-gemma4-e4b-geo3k-1n8g-automodel.yaml
+++ b/examples/configs/recipes/vlm/vlm_grpo-gemma4-e4b-geo3k-1n8g-automodel.yaml
@@ -24,6 +24,8 @@ policy:
       exp_avg_sq_dtype: torch.bfloat16
   dtensor_cfg:
     activation_checkpointing: true
+    env_vars:
+      PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
     automodel_kwargs:
       backend:
         _target_: nemo_automodel.components.models.common.utils.BackendConfig
diff --git a/nemo_rl/models/automodel/setup.py b/nemo_rl/models/automodel/setup.py
index e043e69c3a0..fc3b16ed26f 100644
--- a/nemo_rl/models/automodel/setup.py
+++ b/nemo_rl/models/automodel/setup.py
@@ -585,6 +585,15 @@ def setup_model_and_optimizer(
                 "Please refer to https://github.com/NVIDIA/NeMo-RL/blob/main/docs/model-quirks.md#context-parallel-with-fsdp2 for more details."
             )
 
+        if model_config.model_type == "gemma4_unified":
+            raise AssertionError(
+                "Context parallel is not supported for the Gemma 4 unified "
+                "checkpoint (model_type='gemma4_unified'). Its global-attention "
+                "GQA uses head_dim=512, for which no CP SDPA kernel is available. "
+                "Set policy.dtensor_cfg.context_parallel_size = 1. See "
+                "docs/guides/models/gemma/gemma4.md."
+            )
+
         if tp_size > 1 and sequence_parallel_enabled:
             raise AssertionError(
                 "It's a known issue that context parallel can't be used together with sequence parallel in DTensor worker. "
diff --git a/nemo_rl/models/generation/vllm/vllm_backend.py b/nemo_rl/models/generation/vllm/vllm_backend.py
index 4ca276f8966..09024a798c4 100644
--- a/nemo_rl/models/generation/vllm/vllm_backend.py
+++ b/nemo_rl/models/generation/vllm/vllm_backend.py
@@ -201,6 +201,33 @@ def fix_gemma3_vision_weight_name(key: str) -> str:
     )
 
 
+_GEMMA4_UNIFIED_MULTIMODAL_WEIGHT_MARKERS = (
+    "model.embed_vision.",
+    "model.embed_audio.",
+)
+
+
+def _is_gemma4_unified_text_only(model_config: Any) -> bool:
+    """Return whether a vLLM model is a text-only Gemma 4 Unified model."""
+    multimodal_config = getattr(model_config, "multimodal_config", None)
+    return (
+        "Gemma4UnifiedForConditionalGeneration" in model_config.architectures
+        and multimodal_config is not None
+        and multimodal_config.language_model_only
+    )
+
+
+def _filter_gemma4_unified_multimodal_weights(
+    weights: Iterable[tuple[str, torch.Tensor]],
+) -> Iterator[tuple[str, torch.Tensor]]:
+    """Drop frozen Gemma 4 Unified vision and audio embedder weights."""
+    return (
+        (key, weight)
+        for key, weight in weights
+        if not key.startswith(_GEMMA4_UNIFIED_MULTIMODAL_WEIGHT_MARKERS)
+    )
+
+
 def _read_mtp_layer_weights_from_checkpoint(
     model_path: str, mtp_layer_indices: set[int]
 ) -> list[tuple[str, torch.Tensor]]:
@@ -253,6 +280,8 @@ class VllmInternalWorkerExtension:
     # True once the MTP drafter has been served by a one-time disk load (see
     # load_mtp_weights_from_disk); refit then leaves those static weights alone.
     _mtp_drafter_from_disk: bool = False
+    # Each worker logs the Gemma 4 Unified multimodal filtering at most once.
+    _logged_gemma4_unified_drop: bool = False
     _sparse_delta_applier: Any = None
     _nrl_named_parameters: dict[str, torch.nn.Parameter]
     _nrl_layerwise_reload_active: bool = False
@@ -316,14 +345,14 @@ def _prepare_reload_weight_iterator(
         self, weights: Iterable[tuple[str, torch.Tensor]]
     ) -> Iterable[tuple[str, torch.Tensor]]:
         """Prepare checkpoint-format weights for vLLM's native reload API."""
-        if (
-            "Gemma3ForConditionalGeneration"
-            in self.model_runner.vllm_config.model_config.architectures
-        ):
+        model_config = self.model_runner.vllm_config.model_config
+        if "Gemma3ForConditionalGeneration" in model_config.architectures:
             weights = (
                 (fix_gemma3_vision_weight_name(name), weight)
                 for name, weight in weights
             )
+        if _is_gemma4_unified_text_only(model_config):
+            weights = _filter_gemma4_unified_multimodal_weights(weights)
 
         from nemo_rl.models.generation.vllm.quantization import fp8
 
@@ -785,19 +814,32 @@ def load_mtp_weights_from_disk(self, model_path: str) -> bool:
         return True
 
     def _load_weights(self, weights):
-        """Load weights with Gemma3 vision-tower weight name fix, FP8, and draft-weight support.
+        """Apply model-specific transforms and load policy and draft weights.
 
-        Applies Gemma3 vision-tower weight name fix if needed, splits policy/draft
-        weights, dispatches policy weights through the configured refit loader,
-        and loads draft weights into the drafter model.
+        Checkpoint-format weights are normalized for the target vLLM model, then
+        routed to the policy model and any supported speculative drafter.
         """
-        if (
-            "Gemma3ForConditionalGeneration"
-            in self.model_runner.vllm_config.model_config.architectures
-        ):
+        model_config = self.model_runner.vllm_config.model_config
+        if "Gemma3ForConditionalGeneration" in model_config.architectures:
             for idx, (key, weight) in enumerate(weights):
                 weights[idx] = (fix_gemma3_vision_weight_name(key), weight)
 
+        if _is_gemma4_unified_text_only(model_config):
+            # HF loads the full unified checkpoint, while text-only vLLM uses
+            # encoder-free multimodal stubs with a different parameter layout.
+            # The recipe freezes these towers and never invokes them, so only
+            # refit the language path and leave the unused vLLM stubs untouched.
+            num_weights = len(weights)
+            weights = list(_filter_gemma4_unified_multimodal_weights(weights))
+            num_dropped = num_weights - len(weights)
+            if num_dropped and not getattr(self, "_logged_gemma4_unified_drop", False):
+                self._logged_gemma4_unified_drop = True
+                logger.info(
+                    "Gemma4 Unified text-only refit dropped %d frozen "
+                    "vision/audio weights",
+                    num_dropped,
+                )
+
         policy_weights, draft_weights = self._split_policy_and_draft_weights(weights)
         self._load_hf_weights(policy_weights)
         # Eagle3 draft weights are exported with the `draft.` prefix.
diff --git a/nemo_rl/models/generation/vllm/vllm_worker.py b/nemo_rl/models/generation/vllm/vllm_worker.py
index f503a89c3b5..f82c6dd905e 100644
--- a/nemo_rl/models/generation/vllm/vllm_worker.py
+++ b/nemo_rl/models/generation/vllm/vllm_worker.py
@@ -56,6 +56,7 @@
     register_torchcodec_vllm_video_loader,
 )
 from nemo_rl.models.generation.vllm.worker_utils import (
+    find_tokenizer_required_architectures,
     resolve_data_parallel_local_rank,
     resolve_distributed_executor_backend,
 )
@@ -616,28 +617,9 @@ def _load_model(self, bundle_indices, seed):
                 )
                 # disable quantization
                 vllm_kwargs["hf_overrides"]["quantization_config"] = {}
-        elif any(
-            arch in getattr(hf_config, "architectures", [])
-            for arch in (
-                "Gemma3ForConditionalGeneration",
-                "Gemma4ForConditionalGeneration",
-                "Mistral3ForConditionalGeneration",
-                "Qwen3_5ForConditionalGeneration",
-                "Qwen3_5MoeForConditionalGeneration",
-            )
+        elif detected_arch := find_tokenizer_required_architectures(
+            getattr(hf_config, "architectures", None)
         ):
-            detected_arch = [
-                arch
-                for arch in getattr(hf_config, "architectures", [])
-                if arch
-                in (
-                    "Gemma3ForConditionalGeneration",
-                    "Gemma4ForConditionalGeneration",
-                    "Mistral3ForConditionalGeneration",
-                    "Qwen3_5ForConditionalGeneration",
-                    "Qwen3_5MoeForConditionalGeneration",
-                )
-            ]
             if self.cfg["vllm_cfg"]["skip_tokenizer_init"]:
                 print(
                     f"Detected {detected_arch} which may crash when skip_tokenizer_init is True. "
diff --git a/nemo_rl/models/generation/vllm/worker_utils.py b/nemo_rl/models/generation/vllm/worker_utils.py
index 831fbde0dd5..b3ec271a06a 100644
--- a/nemo_rl/models/generation/vllm/worker_utils.py
+++ b/nemo_rl/models/generation/vllm/worker_utils.py
@@ -12,6 +12,30 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from collections.abc import Iterable
+
+TOKENIZER_REQUIRED_ARCHITECTURES = frozenset(
+    {
+        "Gemma3ForConditionalGeneration",
+        "Gemma4ForConditionalGeneration",
+        "Gemma4UnifiedForConditionalGeneration",
+        "Mistral3ForConditionalGeneration",
+        "Qwen3_5ForConditionalGeneration",
+        "Qwen3_5MoeForConditionalGeneration",
+    }
+)
+
+
+def find_tokenizer_required_architectures(
+    architectures: Iterable[str] | None,
+) -> list[str]:
+    """Return architectures for which vLLM must initialize a tokenizer."""
+    return [
+        architecture
+        for architecture in architectures or ()
+        if architecture in TOKENIZER_REQUIRED_ARCHITECTURES
+    ]
+
 
 def resolve_distributed_executor_backend(
     tensor_parallel_size: int,
diff --git a/nemo_rl/models/policy/utils.py b/nemo_rl/models/policy/utils.py
index 02220b21363..96ddd6ebd43 100644
--- a/nemo_rl/models/policy/utils.py
+++ b/nemo_rl/models/policy/utils.py
@@ -76,6 +76,7 @@
     "internvl": AutoModelForImageTextToText,
     "gemma3": AutoModelForImageTextToText,
     "gemma4": AutoModelForImageTextToText,
+    "gemma4_unified": AutoModelForImageTextToText,
     "smolvlm": AutoModelForImageTextToText,
     "mistral3": AutoModelForImageTextToText,
     "llama4": AutoModelForImageTextToText,
@@ -95,6 +96,7 @@
         "internvl": NeMoAutoModelForImageTextToText,
         "gemma3": NeMoAutoModelForImageTextToText,
         "gemma4": NeMoAutoModelForImageTextToText,
+        "gemma4_unified": NeMoAutoModelForImageTextToText,
         "smolvlm": NeMoAutoModelForImageTextToText,
         "mistral3": NeMoAutoModelForImageTextToText,
         "llama4": NeMoAutoModelForImageTextToText,
diff --git a/tests/test_suites/disabled.txt b/tests/test_suites/disabled.txt
index fcbfd7e18e1..c4dd9afcfa9 100644
--- a/tests/test_suites/disabled.txt
+++ b/tests/test_suites/disabled.txt
@@ -9,6 +9,13 @@ tests/test_suites/vlm/vlm_grpo-qwen3.5-35ba3b-geo3k-2n8g-megatron-ep16.sh
 # Long-context-capable GLM validation is manual-only because it exceeds the recurring-suite GPU-hour budget.
 tests/test_suites/llm/grpo-glm5.2-72n8g-megatron-noncolocated.sh
 
+# Gemma 4 keeps one CP configuration and one non-CP configuration in release.
+# These additional validated layouts remain available for manual testing without
+# adding their allocations to the recurring nightly or release GPU-hour budget.
+tests/test_suites/llm/dapo-gemma4-e2b-it-1n8g-fsdp2cp2-automodel.sh
+tests/test_suites/llm/dapo-gemma4-26ba4b-it-4n8g-fsdp2-automodel.sh
+tests/test_suites/llm/dapo-gemma4-31b-it-4n8g-fsdp2cp2-automodel.sh
+
 # The async video GRPO regression is self-contained, but its 16-node topology
 # is too large for nightly. Run it manually before changes to the async video
 # rollout/training path are merged.
diff --git a/tests/test_suites/llm/dapo-gemma4-12b-it-2n8g-fsdp2-automodel.sh b/tests/test_suites/llm/dapo-gemma4-12b-it-2n8g-fsdp2-automodel.sh
new file mode 100755
index 00000000000..b2dc781d0f4
--- /dev/null
+++ b/tests/test_suites/llm/dapo-gemma4-12b-it-2n8g-fsdp2-automodel.sh
@@ -0,0 +1,48 @@
+#!/bin/bash
+SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd)
+source $SCRIPT_DIR/common.env
+
+# ===== BEGIN CONFIG =====
+NUM_NODES=2
+STEPS_PER_RUN=20
+MAX_STEPS=20
+NUM_RUNS=$(( (MAX_STEPS + STEPS_PER_RUN - 1) / STEPS_PER_RUN ))  # Round up
+NUM_MINUTES=240
+# ===== END CONFIG =====
+
+exit_if_max_steps_reached
+
+# Run the experiment
+cd $PROJECT_ROOT
+uv run examples/run_grpo.py \
+    --config $CONFIG_PATH \
+    grpo.max_num_steps=$MAX_STEPS \
+    logger.log_dir=$LOG_DIR \
+    logger.wandb_enabled=True \
+    logger.wandb.project=nemo-rl \
+    logger.wandb.name=$EXP_NAME \
+    logger.monitor_gpus=True \
+    logger.tensorboard_enabled=True \
+    checkpointing.enabled=True \
+    checkpointing.checkpoint_dir=$CKPT_DIR \
+    $@ \
+    2>&1 | tee $RUN_LOG
+
+# Convert tensorboard logs to json
+uv run tests/json_dump_tb_logs.py $LOG_DIR --output_path $JSON_METRICS
+
+# Only run metrics if the target step is reached
+if [[ $(jq 'to_entries | .[] | select(.key == "train/loss") | .value | keys | map(tonumber) | max' $JSON_METRICS) -ge $MAX_STEPS ]]; then
+    # Calibrated from W&B runs spg9333n and g412b200.
+    uv run tests/check_metrics.py $JSON_METRICS \
+        'all_finite(data["train/loss"])' \
+        'all_finite(data["train/grad_norm"])' \
+        'all_finite(data["train/token_mult_prob_error"])' \
+        'median(data["train/token_mult_prob_error"]) < 1.05' \
+        'mean(data["train/gen_kl_error"]) < 0.001' \
+        'mean(data["train/reward"]) > 0.1' \
+        'mean(data["train/filtered_reward"]) > -0.15'
+
+    # Clean up checkpoint directory after successful run to save space.
+    rm -rf "$CKPT_DIR"
+fi
diff --git a/tests/test_suites/llm/dapo-gemma4-26ba4b-it-4n8g-fsdp2ep16cp2-automodel.sh b/tests/test_suites/llm/dapo-gemma4-26ba4b-it-4n8g-fsdp2ep16cp2-automodel.sh
new file mode 100755
index 00000000000..4e51e80f5e6
--- /dev/null
+++ b/tests/test_suites/llm/dapo-gemma4-26ba4b-it-4n8g-fsdp2ep16cp2-automodel.sh
@@ -0,0 +1,48 @@
+#!/bin/bash
+SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd)
+source $SCRIPT_DIR/common.env
+
+# ===== BEGIN CONFIG =====
+NUM_NODES=4
+STEPS_PER_RUN=20
+MAX_STEPS=20
+NUM_RUNS=$(( (MAX_STEPS + STEPS_PER_RUN - 1) / STEPS_PER_RUN ))  # Round up
+NUM_MINUTES=240
+# ===== END CONFIG =====
+
+exit_if_max_steps_reached
+
+# Run the experiment
+cd $PROJECT_ROOT
+uv run examples/run_grpo.py \
+    --config $CONFIG_PATH \
+    grpo.max_num_steps=$MAX_STEPS \
+    logger.log_dir=$LOG_DIR \
+    logger.wandb_enabled=True \
+    logger.wandb.project=nemo-rl \
+    logger.wandb.name=$EXP_NAME \
+    logger.monitor_gpus=True \
+    logger.tensorboard_enabled=True \
+    checkpointing.enabled=True \
+    checkpointing.checkpoint_dir=$CKPT_DIR \
+    $@ \
+    2>&1 | tee $RUN_LOG
+
+# Convert tensorboard logs to json
+uv run tests/json_dump_tb_logs.py $LOG_DIR --output_path $JSON_METRICS
+
+# Only run metrics if the target step is reached
+if [[ $(jq 'to_entries | .[] | select(.key == "train/loss") | .value | keys | map(tonumber) | max' $JSON_METRICS) -ge $MAX_STEPS ]]; then
+    # Calibrated from W&B runs g26e16c2 and g26e2p93 and recurring CP1 runs.
+    uv run tests/check_metrics.py $JSON_METRICS \
+        'all_finite(data["train/loss"])' \
+        'all_finite(data["train/grad_norm"])' \
+        'all_finite(data["train/token_mult_prob_error"])' \
+        'median(data["train/token_mult_prob_error"]) < 1.05' \
+        'mean(data["train/gen_kl_error"]) < 0.0015' \
+        'mean(data["train/reward"]) > 0.25' \
+        'mean(data["train/filtered_reward"]) > -0.15'
+
+    # Clean up checkpoint directory after successful run to save space.
+    rm -rf "$CKPT_DIR"
+fi
diff --git a/tests/test_suites/llm/dapo-gemma4-31b-it-4n8g-fsdp2-automodel.sh b/tests/test_suites/llm/dapo-gemma4-31b-it-4n8g-fsdp2-automodel.sh
index 0f323f9c69c..aa48b00e4b2 100755
--- a/tests/test_suites/llm/dapo-gemma4-31b-it-4n8g-fsdp2-automodel.sh
+++ b/tests/test_suites/llm/dapo-gemma4-31b-it-4n8g-fsdp2-automodel.sh
@@ -33,13 +33,15 @@ uv run tests/json_dump_tb_logs.py $LOG_DIR --output_path $JSON_METRICS
 
 # Only run metrics if the target step is reached
 if [[ $(jq 'to_entries | .[] | select(.key == "train/loss") | .value | keys | map(tonumber) | max' $JSON_METRICS) -ge $MAX_STEPS ]]; then
-    # Thresholds calibrated from the offpolicy baseline run (wandb project
-    # nemorl-gemma4-support, display name dapo-gemma4-31b-it-4n8g-fsdp2-automodel-offpolicy).
+    # Calibrated from recent nvidia/nemo-rl release runs and W&B run g31c1s91.
     uv run tests/check_metrics.py $JSON_METRICS \
+        'all_finite(data["train/loss"])' \
+        'all_finite(data["train/grad_norm"])' \
+        'all_finite(data["train/token_mult_prob_error"])' \
         'median(data["train/token_mult_prob_error"]) < 1.05' \
-        'mean(data["train/gen_kl_error"]) < 0.002' \
-        'data["train/reward"]["20"] > 0.1' \
-        'data["train/filtered_reward"]["20"] > -0.35'
+        'mean(data["train/gen_kl_error"]) < 0.001' \
+        'mean(data["train/reward"]) > 0.4' \
+        'mean(data["train/filtered_reward"]) > 0.05'
 
     # Clean up checkpoint directory after successful run to save space.
     rm -rf "$CKPT_DIR"
diff --git a/tests/test_suites/llm/dapo-gemma4-31b-it-4n8g-fsdp2cp2-automodel.sh b/tests/test_suites/llm/dapo-gemma4-31b-it-4n8g-fsdp2cp2-automodel.sh
new file mode 100755
index 00000000000..651a7b4855b
--- /dev/null
+++ b/tests/test_suites/llm/dapo-gemma4-31b-it-4n8g-fsdp2cp2-automodel.sh
@@ -0,0 +1,44 @@
+#!/bin/bash
+SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd)
+source $SCRIPT_DIR/common.env
+
+# ===== BEGIN CONFIG =====
+NUM_NODES=4
+STEPS_PER_RUN=20
+MAX_STEPS=20
+NUM_RUNS=$(( (MAX_STEPS + STEPS_PER_RUN - 1) / STEPS_PER_RUN ))  # Round up
+NUM_MINUTES=240
+# ===== END CONFIG =====
+
+exit_if_max_steps_reached
+
+# Run the experiment
+cd $PROJECT_ROOT
+uv run examples/run_grpo.py \
+    --config $CONFIG_PATH \
+    grpo.max_num_steps=$MAX_STEPS \
+    logger.log_dir=$LOG_DIR \
+    logger.wandb_enabled=True \
+    logger.wandb.project=nemo-rl \
+    logger.wandb.name=$EXP_NAME \
+    logger.monitor_gpus=True \
+    logger.tensorboard_enabled=True \
+    checkpointing.enabled=True \
+    checkpointing.checkpoint_dir=$CKPT_DIR \
+    $@ \
+    2>&1 | tee $RUN_LOG
+
+# Convert tensorboard logs to json
+uv run tests/json_dump_tb_logs.py $LOG_DIR --output_path $JSON_METRICS
+
+# Only run metrics if the target step is reached
+if [[ $(jq 'to_entries | .[] | select(.key == "train/loss") | .value | keys | map(tonumber) | max' $JSON_METRICS) -ge $MAX_STEPS ]]; then
+    uv run tests/check_metrics.py $JSON_METRICS \
+        'median(data["train/token_mult_prob_error"]) < 1.05' \
+        'mean(data["train/gen_kl_error"]) < 0.002' \
+        'data["train/reward"]["20"] > 0.1' \
+        'data["train/filtered_reward"]["20"] > -0.35'
+
+    # Clean up checkpoint directory after successful run to save space.
+    rm -rf "$CKPT_DIR"
+fi
diff --git a/tests/test_suites/llm/dapo-gemma4-e2b-it-1n8g-fsdp2-automodel.sh b/tests/test_suites/llm/dapo-gemma4-e2b-it-1n8g-fsdp2-automodel.sh
index d46a0bb6098..93b45f47732 100755
--- a/tests/test_suites/llm/dapo-gemma4-e2b-it-1n8g-fsdp2-automodel.sh
+++ b/tests/test_suites/llm/dapo-gemma4-e2b-it-1n8g-fsdp2-automodel.sh
@@ -33,11 +33,15 @@ uv run tests/json_dump_tb_logs.py $LOG_DIR --output_path $JSON_METRICS
 
 # Only run metrics if the target step is reached
 if [[ $(jq 'to_entries | .[] | select(.key == "train/loss") | .value | keys | map(tonumber) | max' $JSON_METRICS) -ge $MAX_STEPS ]]; then
+    # Calibrated from recent nvidia/nemo-rl nightly runs and W&B run e2c1a830.
     uv run tests/check_metrics.py $JSON_METRICS \
-        'median(data["train/token_mult_prob_error"]) < 1.1' \
-        'data["train/reward"]["20"] > -1.15' \
-        'data["train/filtered_reward"]["20"] > -1.10' \
-        'data["train/gen_kl_error"]["20"] < 0.001'
+        'all_finite(data["train/loss"])' \
+        'all_finite(data["train/grad_norm"])' \
+        'all_finite(data["train/token_mult_prob_error"])' \
+        'median(data["train/token_mult_prob_error"]) < 1.05' \
+        'mean(data["train/gen_kl_error"]) < 0.001' \
+        'mean(data["train/reward"]) > -1.15' \
+        'mean(data["train/filtered_reward"]) > -0.75'
 
     # Clean up checkpoint directory after successful run to save space.
     rm -rf "$CKPT_DIR"
diff --git a/tests/test_suites/llm/dapo-gemma4-e2b-it-1n8g-fsdp2cp2-automodel.sh b/tests/test_suites/llm/dapo-gemma4-e2b-it-1n8g-fsdp2cp2-automodel.sh
new file mode 100755
index 00000000000..d46a0bb6098
--- /dev/null
+++ b/tests/test_suites/llm/dapo-gemma4-e2b-it-1n8g-fsdp2cp2-automodel.sh
@@ -0,0 +1,44 @@
+#!/bin/bash
+SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd)
+source $SCRIPT_DIR/common.env
+
+# ===== BEGIN CONFIG =====
+NUM_NODES=1
+STEPS_PER_RUN=20
+MAX_STEPS=20
+NUM_RUNS=$(( (MAX_STEPS + STEPS_PER_RUN - 1) / STEPS_PER_RUN ))  # Round up
+NUM_MINUTES=90
+# ===== END CONFIG =====
+
+exit_if_max_steps_reached
+
+# Run the experiment
+cd $PROJECT_ROOT
+uv run examples/run_grpo.py \
+    --config $CONFIG_PATH \
+    grpo.max_num_steps=$MAX_STEPS \
+    logger.log_dir=$LOG_DIR \
+    logger.wandb_enabled=True \
+    logger.wandb.project=nemo-rl \
+    logger.wandb.name=$EXP_NAME \
+    logger.monitor_gpus=True \
+    logger.tensorboard_enabled=True \
+    checkpointing.enabled=True \
+    checkpointing.checkpoint_dir=$CKPT_DIR \
+    $@ \
+    2>&1 | tee $RUN_LOG
+
+# Convert tensorboard logs to json
+uv run tests/json_dump_tb_logs.py $LOG_DIR --output_path $JSON_METRICS
+
+# Only run metrics if the target step is reached
+if [[ $(jq 'to_entries | .[] | select(.key == "train/loss") | .value | keys | map(tonumber) | max' $JSON_METRICS) -ge $MAX_STEPS ]]; then
+    uv run tests/check_metrics.py $JSON_METRICS \
+        'median(data["train/token_mult_prob_error"]) < 1.1' \
+        'data["train/reward"]["20"] > -1.15' \
+        'data["train/filtered_reward"]["20"] > -1.10' \
+        'data["train/gen_kl_error"]["20"] < 0.001'
+
+    # Clean up checkpoint directory after successful run to save space.
+    rm -rf "$CKPT_DIR"
+fi
diff --git a/tests/test_suites/nightly.txt b/tests/test_suites/nightly.txt
index f8ed7a2f69e..f214277fef8 100644
--- a/tests/test_suites/nightly.txt
+++ b/tests/test_suites/nightly.txt
@@ -199,6 +199,7 @@ tests/test_suites/llm/ppo-qwen2.5-1.5b-gsm8k-2n8g-megatron-valuetp2sp-dynbatch-n
 
 # Gemma 4 functional runs
 tests/test_suites/llm/dapo-gemma4-e2b-it-1n8g-fsdp2-automodel.sh
+tests/test_suites/llm/dapo-gemma4-12b-it-2n8g-fsdp2-automodel.sh
 
 # Functional Gemma 4 E4B VLM GRPO run
 tests/test_suites/vlm/vlm_grpo-gemma4-e4b-geo3k-1n8g-automodel.sh
diff --git a/tests/test_suites/release.txt b/tests/test_suites/release.txt
index 0100881731a..f84ada484f6 100644
--- a/tests/test_suites/release.txt
+++ b/tests/test_suites/release.txt
@@ -37,7 +37,7 @@ tests/test_suites/llm/grpo-glm5.1-64n8g-megatron.sh
 tests/test_suites/llm/grpo-glm5.2-64n8g-megatron-6K-colocated.sh
 
 # Gemma 4 DAPO runs
-tests/test_suites/llm/dapo-gemma4-26ba4b-it-4n8g-fsdp2-automodel.sh
+tests/test_suites/llm/dapo-gemma4-26ba4b-it-4n8g-fsdp2ep16cp2-automodel.sh
 tests/test_suites/llm/dapo-gemma4-31b-it-4n8g-fsdp2-automodel.sh
 
 # Mistral Medium 3.5 (128B, FP8, text-only) DAPO run
diff --git a/tests/test_suites/vlm/vlm_grpo-gemma4-e4b-geo3k-1n8g-automodel.sh b/tests/test_suites/vlm/vlm_grpo-gemma4-e4b-geo3k-1n8g-automodel.sh
index a35c543943b..7f89d05dc04 100755
--- a/tests/test_suites/vlm/vlm_grpo-gemma4-e4b-geo3k-1n8g-automodel.sh
+++ b/tests/test_suites/vlm/vlm_grpo-gemma4-e4b-geo3k-1n8g-automodel.sh
@@ -33,9 +33,14 @@ uv run tests/json_dump_tb_logs.py $LOG_DIR --output_path $JSON_METRICS
 
 # Only run metrics if the target step is reached
 if [[ $(jq 'to_entries | .[] | select(.key == "train/loss") | .value | keys | map(tonumber) | max' $JSON_METRICS) -ge $MAX_STEPS ]]; then
+    # Calibrated from recent nvidia/nemo-rl nightly runs through 2026-09-04.
     uv run tests/check_metrics.py $JSON_METRICS \
-        'max(data["train/reward"]) > 0.45' \
-        'median(data["train/token_mult_prob_error"]) < 1.05'
+        'all_finite(data["train/loss"])' \
+        'all_finite(data["train/grad_norm"])' \
+        'all_finite(data["train/token_mult_prob_error"])' \
+        'median(data["train/token_mult_prob_error"]) < 1.05' \
+        'mean(data["train/gen_kl_error"]) < 0.001' \
+        'mean(data["train/reward"]) > 0.4'
 
     # Clean up checkpoint directory after successful run to save space.
     rm -rf "$CKPT_DIR"
diff --git a/tests/unit/models/automodel/test_automodel_setup.py b/tests/unit/models/automodel/test_automodel_setup.py
index 9aba2d99fc7..0c7ed5e99cf 100644
--- a/tests/unit/models/automodel/test_automodel_setup.py
+++ b/tests/unit/models/automodel/test_automodel_setup.py
@@ -1631,6 +1631,48 @@ def test_setup_model_with_cp_raises_for_gemma3(
                 checkpoint_manager=mock_checkpoint_manager,
             )
 
+    @patch("nemo_rl.models.automodel.setup.torch.distributed.get_rank")
+    @patch("nemo_rl.models.automodel.setup.get_class")
+    def test_setup_model_with_cp_raises_for_gemma4_unified(
+        self,
+        mock_get_class,
+        mock_get_rank,
+        mock_config,
+        mock_runtime_config,
+        mock_checkpoint_manager,
+        mock_tokenizer,
+    ):
+        """Test that Gemma 4 unified checkpoints reject context parallel."""
+        mock_get_rank.return_value = 0
+        mock_fsdp2_config = MagicMock()
+        mock_fsdp2_config.sequence_parallel = False
+        distributed_context = DistributedContext(
+            device_mesh=MagicMock(),
+            moe_mesh=MagicMock(),
+            fsdp2_config=mock_fsdp2_config,
+            moe_config=MagicMock(),
+            dp_size=1,
+            tp_size=1,
+            cp_size=2,
+        )
+
+        mock_runtime_config.model_config.model_type = "gemma4_unified"
+        mock_runtime_config.model_config.architectures = [
+            "Gemma4UnifiedForConditionalGeneration"
+        ]
+
+        with pytest.raises(
+            AssertionError,
+            match="Context parallel is not supported for the Gemma 4 unified checkpoint",
+        ):
+            setup_model_and_optimizer(
+                config=mock_config,
+                tokenizer=mock_tokenizer,
+                runtime_config=mock_runtime_config,
+                distributed_context=distributed_context,
+                checkpoint_manager=mock_checkpoint_manager,
+            )
+
     @patch("nemo_rl.models.automodel.setup.torch.optim.lr_scheduler.LambdaLR")
     @patch("nemo_rl.models.automodel.setup.torch.distributed.get_rank")
     @patch("nemo_rl.models.automodel.setup.get_class")
diff --git a/tests/unit/models/generation/test_vllm_backend.py b/tests/unit/models/generation/test_vllm_backend.py
index f7a7230f3e9..056762d81f3 100644
--- a/tests/unit/models/generation/test_vllm_backend.py
+++ b/tests/unit/models/generation/test_vllm_backend.py
@@ -1501,6 +1501,166 @@ def test_load_weights_routes_only_policy_weights_to_mtp_drafter(monkeypatch):
     ext._maybe_refit_mtp_drafter.assert_called_once_with(policy_weights)
 
 
+@pytest.mark.vllm
+@pytest.mark.parametrize(
+    ("architecture", "language_model_only", "expected_keys"),
+    [
+        (
+            "Gemma4UnifiedForConditionalGeneration",
+            True,
+            ["model.language_model.layers.0.self_attn.q_proj.weight"],
+        ),
+        (
+            "Gemma4UnifiedForConditionalGeneration",
+            False,
+            [
+                "model.language_model.layers.0.self_attn.q_proj.weight",
+                "model.embed_vision.pos_embedding",
+                "model.embed_audio.embedding_projection.weight",
+            ],
+        ),
+        (
+            "Gemma4ForConditionalGeneration",
+            True,
+            [
+                "model.language_model.layers.0.self_attn.q_proj.weight",
+                "model.embed_vision.pos_embedding",
+                "model.embed_audio.embedding_projection.weight",
+            ],
+        ),
+    ],
+)
+def test_gemma4_refit_drops_multimodal_weights_only_for_unified_text_generation(
+    monkeypatch, architecture, language_model_only, expected_keys
+):
+    from nemo_rl.models.generation.vllm.quantization import fp8
+    from nemo_rl.models.generation.vllm.vllm_backend import (
+        VllmInternalWorkerExtension,
+    )
+
+    loaded = []
+    ext = VllmInternalWorkerExtension.__new__(VllmInternalWorkerExtension)
+    ext.model_runner = SimpleNamespace(
+        model=SimpleNamespace(load_weights=lambda *, weights: loaded.extend(weights)),
+        vllm_config=SimpleNamespace(
+            model_config=SimpleNamespace(
+                architectures=[architecture],
+                multimodal_config=SimpleNamespace(
+                    language_model_only=language_model_only
+                ),
+            )
+        ),
+    )
+    ext._load_draft_weights = MagicMock()
+    ext._maybe_refit_mtp_drafter = MagicMock()
+    monkeypatch.setattr(fp8, "is_fp8_model", lambda _: False)
+    weights = [
+        ("model.language_model.layers.0.self_attn.q_proj.weight", "language"),
+        ("model.embed_vision.pos_embedding", "vision"),
+        ("model.embed_audio.embedding_projection.weight", "audio"),
+    ]
+
+    ext._load_weights(weights)
+
+    assert [key for key, _ in loaded] == expected_keys
+
+
+@pytest.mark.vllm
+@pytest.mark.parametrize(
+    ("architecture", "language_model_only", "expected_keys"),
+    [
+        (
+            "Gemma4UnifiedForConditionalGeneration",
+            True,
+            ["model.language_model.layers.0.self_attn.q_proj.weight"],
+        ),
+        (
+            "Gemma4UnifiedForConditionalGeneration",
+            False,
+            [
+                "model.language_model.layers.0.self_attn.q_proj.weight",
+                "model.embed_vision.pos_embedding",
+                "model.embed_audio.embedding_projection.weight",
+            ],
+        ),
+        (
+            "Gemma4ForConditionalGeneration",
+            True,
+            [
+                "model.language_model.layers.0.self_attn.q_proj.weight",
+                "model.embed_vision.pos_embedding",
+                "model.embed_audio.embedding_projection.weight",
+            ],
+        ),
+    ],
+)
+def test_prepare_reload_weight_iterator_filters_only_unified_text_generation(
+    monkeypatch, architecture, language_model_only, expected_keys
+):
+    from nemo_rl.models.generation.vllm.quantization import fp8
+    from nemo_rl.models.generation.vllm.vllm_backend import (
+        VllmInternalWorkerExtension,
+    )
+
+    ext = VllmInternalWorkerExtension.__new__(VllmInternalWorkerExtension)
+    ext.model_runner = SimpleNamespace(
+        vllm_config=SimpleNamespace(
+            model_config=SimpleNamespace(
+                architectures=[architecture],
+                multimodal_config=SimpleNamespace(
+                    language_model_only=language_model_only
+                ),
+            )
+        )
+    )
+    monkeypatch.setattr(fp8, "is_fp8_model", lambda _: False)
+    weights = iter(
+        [
+            ("model.language_model.layers.0.self_attn.q_proj.weight", "language"),
+            ("model.embed_vision.pos_embedding", "vision"),
+            ("model.embed_audio.embedding_projection.weight", "audio"),
+        ]
+    )
+
+    result = list(ext._prepare_reload_weight_iterator(weights))
+
+    assert [key for key, _ in result] == expected_keys
+
+
+@pytest.mark.vllm
+def test_gemma4_unified_refit_logs_dropped_weights_once(monkeypatch):
+    from nemo_rl.models.generation.vllm import vllm_backend
+
+    ext = vllm_backend.VllmInternalWorkerExtension.__new__(
+        vllm_backend.VllmInternalWorkerExtension
+    )
+    ext.model_runner = SimpleNamespace(
+        vllm_config=SimpleNamespace(
+            model_config=SimpleNamespace(
+                architectures=["Gemma4UnifiedForConditionalGeneration"],
+                multimodal_config=SimpleNamespace(language_model_only=True),
+            )
+        )
+    )
+    ext._load_hf_weights = MagicMock()
+    ext._load_draft_weights = MagicMock()
+    ext._maybe_refit_mtp_drafter = MagicMock()
+    log_info = MagicMock()
+    monkeypatch.setattr(vllm_backend.logger, "info", log_info)
+    weights = [
+        ("model.language_model.layers.0.self_attn.q_proj.weight", "language"),
+        ("model.embed_vision.pos_embedding", "vision"),
+        ("model.embed_audio.embedding_projection.weight", "audio"),
+    ]
+
+    ext._load_weights(weights)
+    ext._load_weights(weights)
+
+    log_info.assert_called_once_with(
+        "Gemma4 Unified text-only refit dropped %d frozen vision/audio weights", 2
+    )
+
+
 @pytest.mark.vllm
 @pytest.mark.parametrize(
     "method, from_disk, has_drafter, expected",
diff --git a/tests/unit/models/generation/test_vllm_worker_helpers.py b/tests/unit/models/generation/test_vllm_worker_helpers.py
index 29d61934447..23130494677 100644
--- a/tests/unit/models/generation/test_vllm_worker_helpers.py
+++ b/tests/unit/models/generation/test_vllm_worker_helpers.py
@@ -17,11 +17,39 @@
 import pytest
 
 from nemo_rl.models.generation.vllm.worker_utils import (
+    find_tokenizer_required_architectures,
     resolve_data_parallel_local_rank,
     resolve_distributed_executor_backend,
 )
 
 
+@pytest.mark.parametrize(
+    ("architectures", "expected"),
+    [
+        (None, []),
+        ([], []),
+        (["Gemma4ForCausalLM"], []),
+        (
+            ["Gemma4ForConditionalGeneration"],
+            ["Gemma4ForConditionalGeneration"],
+        ),
+        (
+            [
+                "Gemma4ForCausalLM",
+                "Gemma4UnifiedForConditionalGeneration",
+                "Mistral3ForConditionalGeneration",
+            ],
+            [
+                "Gemma4UnifiedForConditionalGeneration",
+                "Mistral3ForConditionalGeneration",
+            ],
+        ),
+    ],
+)
+def test_find_tokenizer_required_architectures(architectures, expected):
+    assert find_tokenizer_required_architectures(architectures) == expected
+
+
 @pytest.mark.parametrize(
     ("tp", "pp", "ep", "expected"),
     [
diff --git a/tests/unit/models/policy/test_utils.py b/tests/unit/models/policy/test_utils.py
index b30cf653f4d..b5a59cbd7ec 100644
--- a/tests/unit/models/policy/test_utils.py
+++ b/tests/unit/models/policy/test_utils.py
@@ -100,6 +100,10 @@ def test_resolve_model_class_selects_requested_loader(
     assert resolve_model_class(model_type) is expected_default
 
 
+def test_resolve_model_class_routes_gemma4_unified_to_image_text_model():
+    assert "ImageTextToText" in resolve_model_class("gemma4_unified").__name__
+
+
 class TestGetMegatronCheckpointDir:
     """Test cases for the get_megatron_checkpoint_dir function."""
 
diff --git a/tests/unit/test_recipes_and_test_suites.py b/tests/unit/test_recipes_and_test_suites.py
index a6d4d4032ef..84d0a6707b8 100644
--- a/tests/unit/test_recipes_and_test_suites.py
+++ b/tests/unit/test_recipes_and_test_suites.py
@@ -257,7 +257,7 @@ def test_all_recipe_yamls_accounted_for_in_test_suites(
     )
 
 
-def test_nightly_compute_stays_below_4280_hours(nightly_test_suite, tracker):
+def test_nightly_compute_stays_below_4344_hours(nightly_test_suite, tracker):
     command = f"DRYRUN=1 HF_HOME=... HF_DATASETS_CACHE=... CONTAINER= ACCOUNT= PARTITION= ./tools/launch {' '.join(nightly_test_suite)}"
 
     print(f"Running command: {command}")
@@ -289,8 +289,8 @@ def test_nightly_compute_stays_below_4280_hours(nightly_test_suite, tracker):
         f"Last line of output was not as expected: '{last_line}'"
     )
     total_gpu_hours = float(last_line.split(":")[-1].strip())
-    assert total_gpu_hours <= 4280, (
-        f"Total GPU hours exceeded 4280: {last_line}. We should revisit the test suites to reduce the total GPU hours."
+    assert total_gpu_hours <= 4344, (
+        f"Total GPU hours exceeded 4344: {last_line}. We should revisit the test suites to reduce the total GPU hours."
     )
     tracker.track("total_nightly_gpu_hours", total_gpu_hours)