Update Gemma3VL model training scripts (#15041)

genquan9 · tango4j · chtruong814 · web-flow · commit ac2795e362f2 · 2025-11-21T10:30:34.000-06:00
* Fixing lines for multispeaker pipeline (#15030) * Fixing lines for multispeaker pipeline Signed-off-by: taejinp <tango4j@gmail.com> * Removing unused imports Signed-off-by: taejinp <tango4j@gmail.com> * Apply isort and black reformatting Signed-off-by: tango4j <tango4j@users.noreply.github.com> * Making changes for HF Space deployment Signed-off-by: taejinp <tango4j@gmail.com> * Apply isort and black reformatting Signed-off-by: chtruong814 <chtruong814@users.noreply.github.com> * Updated multispk trans utils. Signed-off-by: taejinp <tango4j@gmail.com> --------- Signed-off-by: taejinp <tango4j@gmail.com> Signed-off-by: tango4j <tango4j@users.noreply.github.com> Signed-off-by: chtruong814 <chtruong814@users.noreply.github.com> Co-authored-by: tango4j <tango4j@users.noreply.github.com> Co-authored-by: chtruong814 <chtruong814@users.noreply.github.com> Signed-off-by: genquan9 <genquan@google.com> * Support gemma3vl tuning with verified performances Signed-off-by: genquan9 <genquan@google.com> * minor update gemma3vl parameters for easier usages Signed-off-by: genquan9 <genquan@google.com> * Apply isort and black reformatting Signed-off-by: genquan9 <genquan9@users.noreply.github.com> Signed-off-by: genquan9 <genquan@google.com> * Inference optimization for cache-aware pipelines (#15035) * optimize context manager and cache feature bufferer Signed-off-by: naymaraq <dkaramyan@nvidia.com> * speedUp cache_feature_bufferer Signed-off-by: naymaraq <dkaramyan@nvidia.com> * improved docstring in BatchedCacheFeatureBufferer Signed-off-by: naymaraq <dkaramyan@nvidia.com> --------- Signed-off-by: naymaraq <dkaramyan@nvidia.com> Co-authored-by: naymaraq <dkaramyan@nvidia.com> Signed-off-by: genquan9 <genquan@google.com> * fix loading of hyb ctc rnnt bpe models when using from pretrained (#15042) * fix loading of hyb ctc rnnt bpe models when using from pretrained Signed-off-by: nithinraok <nithinrao.koluguri@gmail.com> * Apply isort and black reformatting Signed-off-by: nithinraok <nithinraok@users.noreply.github.com> --------- Signed-off-by: nithinraok <nithinrao.koluguri@gmail.com> Signed-off-by: nithinraok <nithinraok@users.noreply.github.com> Co-authored-by: nithinraok <nithinraok@users.noreply.github.com> Signed-off-by: genquan9 <genquan@google.com> * revert ckpt scripts removal from #14617 (#15048) Signed-off-by: genquan9 <genquan@google.com> * chore: remove ExportDeploy (#15033) * add EP in PTQ (#15015) Signed-off-by: jenchen13 <jennifchen@nvidia.com> Signed-off-by: Pablo Garay <pagaray@nvidia.com> * remove ExportDeploy Signed-off-by: Pablo Garay <pagaray@nvidia.com> * remove exportDeploy tests Signed-off-by: Pablo Garay <pagaray@nvidia.com> * remove references Signed-off-by: Pablo Garay <pagaray@nvidia.com> * lintfix Signed-off-by: Pablo Garay <pagaray@nvidia.com> * Fixing lines for multispeaker pipeline (#15030) * Fixing lines for multispeaker pipeline Signed-off-by: taejinp <tango4j@gmail.com> * Removing unused imports Signed-off-by: taejinp <tango4j@gmail.com> * Apply isort and black reformatting Signed-off-by: tango4j <tango4j@users.noreply.github.com> * Making changes for HF Space deployment Signed-off-by: taejinp <tango4j@gmail.com> * Apply isort and black reformatting Signed-off-by: chtruong814 <chtruong814@users.noreply.github.com> * Updated multispk trans utils. Signed-off-by: taejinp <tango4j@gmail.com> --------- Signed-off-by: taejinp <tango4j@gmail.com> Signed-off-by: tango4j <tango4j@users.noreply.github.com> Signed-off-by: chtruong814 <chtruong814@users.noreply.github.com> Co-authored-by: tango4j <tango4j@users.noreply.github.com> Co-authored-by: chtruong814 <chtruong814@users.noreply.github.com> Signed-off-by: Pablo Garay <pagaray@nvidia.com> * remove ExportDeploy & references Signed-off-by: Pablo Garay <pagaray@nvidia.com> * lintfix Signed-off-by: Pablo Garay <pagaray@nvidia.com> * get load_ckpt back Signed-off-by: Pablo Garay <pagaray@nvidia.com> * lintfix Signed-off-by: Pablo Garay <pagaray@nvidia.com> * Apply isort and black reformatting Signed-off-by: pablo-garay <pablo-garay@users.noreply.github.com> * back Signed-off-by: Pablo Garay <pagaray@nvidia.com> * revert back Signed-off-by: Pablo Garay <pagaray@nvidia.com> * revert back Signed-off-by: Pablo Garay <pagaray@nvidia.com> * remove ExportDeploy Signed-off-by: Pablo Garay <pagaray@nvidia.com> --------- Signed-off-by: jenchen13 <jennifchen@nvidia.com> Signed-off-by: Pablo Garay <pagaray@nvidia.com> Signed-off-by: taejinp <tango4j@gmail.com> Signed-off-by: tango4j <tango4j@users.noreply.github.com> Signed-off-by: chtruong814 <chtruong814@users.noreply.github.com> Signed-off-by: pablo-garay <pablo-garay@users.noreply.github.com> Co-authored-by: Jenny Chen <jennifchen@nvidia.com> Co-authored-by: Taejin Park <tango4j@gmail.com> Co-authored-by: tango4j <tango4j@users.noreply.github.com> Co-authored-by: chtruong814 <chtruong814@users.noreply.github.com> Co-authored-by: pablo-garay <pablo-garay@users.noreply.github.com> Signed-off-by: genquan9 <genquan@google.com> * fix after ED remove (#15051) Signed-off-by: Pablo Garay <pagaray@nvidia.com> Signed-off-by: genquan9 <genquan@google.com> * Update changelog for `v2.5.3` (#15055) * beep boop: Update changelog Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Update changelog for 2.5.3 Signed-off-by: Charlie Truong <chtruong@nvidia.com> --------- Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Signed-off-by: Charlie Truong <chtruong@nvidia.com> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Charlie Truong <chtruong@nvidia.com> Signed-off-by: genquan9 <genquan@google.com> * [voice agent] Fix RTVI missing bot message (#15068) * fix RTVI missing bot message, fix diar not passing VAD frames Signed-off-by: stevehuang52 <heh@nvidia.com> * revert change to diar Signed-off-by: stevehuang52 <heh@nvidia.com> --------- Signed-off-by: stevehuang52 <heh@nvidia.com> Signed-off-by: genquan9 <genquan@google.com> * [voice agent] make parakeet-eou model default stt (#15069) * make eou model default stt Signed-off-by: stevehuang52 <heh@nvidia.com> * fix typo Signed-off-by: stevehuang52 <heh@nvidia.com> * clean up doc Signed-off-by: stevehuang52 <heh@nvidia.com> --------- Signed-off-by: stevehuang52 <heh@nvidia.com> Signed-off-by: He Huang (Steve) <105218074+stevehuang52@users.noreply.github.com> Signed-off-by: genquan9 <genquan@google.com> * minor fixes to remove unused headers/lines and add exception Signed-off-by: genquan9 <genquan@google.com> * resolve merge conflicts from github Signed-off-by: genquan9 <genquan@google.com> * removed old buffered CTC script (#15061) * removed old buffered CTC script Signed-off-by: naymaraq <dkaramyan@nvidia.com> * remove references to speech_to_text_buffered_infer_ctc.py Signed-off-by: naymaraq <dkaramyan@nvidia.com> --------- Signed-off-by: naymaraq <dkaramyan@nvidia.com> Co-authored-by: naymaraq <dkaramyan@nvidia.com> Signed-off-by: genquan9 <genquan@google.com> * remove unused imports Signed-off-by: genquan9 <genquan@google.com> * remove nlp related notebooks (#15070) Signed-off-by: nithinraok <nithinrao.koluguri@gmail.com> Signed-off-by: genquan9 <genquan@google.com> * chore: Remove Automodel module (#15044) * Delete Automodel module Signed-off-by: Dong Hyuk Chang <donghyukc@nvidia.com> * Remove additional code using or importing automodel pathway Signed-off-by: Dong Hyuk Chang <donghyukc@nvidia.com> * Remove unused import Signed-off-by: Dong Hyuk Chang <donghyukc@nvidia.com> * Further remove hf automodel testing and hf automodel in vlm Signed-off-by: Dong Hyuk Chang <donghyukc@nvidia.com> * Remove unused vars Signed-off-by: Dong Hyuk Chang <donghyukc@nvidia.com> * Remove automodel instance in model opt Signed-off-by: Dong Hyuk Chang <donghyukc@nvidia.com> * Remove hf_auto_model_for_causal_ln Signed-off-by: Dong Hyuk Chang <donghyukc@nvidia.com> * Delete HFAutomodel from speech Signed-off-by: Dong Hyuk Chang <donghyukc@nvidia.com> * Add noqa Signed-off-by: Dong Hyuk Chang <donghyukc@nvidia.com> * Apply isort and black reformatting Signed-off-by: thomasdhc <thomasdhc@users.noreply.github.com> * Remove automodel related tests Signed-off-by: Dong Hyuk Chang <donghyukc@nvidia.com> * Update init file to use import Signed-off-by: Dong Hyuk Chang <donghyukc@nvidia.com> * Apply isort and black reformatting Signed-off-by: thomasdhc <thomasdhc@users.noreply.github.com> --------- Signed-off-by: Dong Hyuk Chang <donghyukc@nvidia.com> Signed-off-by: thomasdhc <thomasdhc@users.noreply.github.com> Co-authored-by: thomasdhc <thomasdhc@users.noreply.github.com> Signed-off-by: genquan9 <genquan@google.com> * add support for parallel ckpt removal (#15073) * add support for parallel ckpt removal Signed-off-by: dimapihtar <dpihtar@gmail.com> * Apply isort and black reformatting Signed-off-by: dimapihtar <dimapihtar@users.noreply.github.com> --------- Signed-off-by: dimapihtar <dpihtar@gmail.com> Signed-off-by: dimapihtar <dimapihtar@users.noreply.github.com> Co-authored-by: dimapihtar <dimapihtar@users.noreply.github.com> Signed-off-by: genquan9 <genquan@google.com> * Fix vlm engine changes in mcore (#15076) Signed-off-by: genquan9 <genquan@google.com> * Add docstring for encode_vqa_sample_multi_turns, and fix long comments Signed-off-by: genquan9 <genquan@google.com> * Update MagpieTTS model with latest changes (#15031) * Update MagpieTTS Signed-off-by: Jason <jasoli@nvidia.com> * allow None in dataset path Signed-off-by: Jason <jasoli@nvidia.com> * try to fix test by removing lhotse; fix yamls in fast dev run tests Signed-off-by: Jason <jasoli@nvidia.com> * increase zeroshot cer value; attempt to fix PO test; add back lhotse in parakeet inference to test segmentation fault Signed-off-by: Jason <jasoli@nvidia.com> * remove branch from test Signed-off-by: Jason <jasoli@nvidia.com> * use batch_size 1 Signed-off-by: Jason <jasoli@nvidia.com> * update GRPO test script Signed-off-by: Jason <jasoli@nvidia.com> * add use_lhotse as a param to transcribe; attempt to fix PO test again; attempt to catch error Signed-off-by: Jason <jasoli@nvidia.com> * fix tests Signed-off-by: Jason <jasoli@nvidia.com> * update rnnt transcribe; fix po test again Signed-off-by: Jason <jasoli@nvidia.com> * Apply suggestion from @XuesongYang Signed-off-by: Xuesong Yang <1646669+XuesongYang@users.noreply.github.com> Signed-off-by: Xuesong Yang <1646669+XuesongYang@users.noreply.github.com> * Move FCD copyright text from TorchEval to top of file Signed-off-by: Fejgin, Roy <rfejgin@nvidia.com> * Remove duplicate copyright text It is now at the top of the file. Signed-off-by: Fejgin, Roy <rfejgin@nvidia.com> * Apply suggestion from @XuesongYang Signed-off-by: Xuesong Yang <1646669+XuesongYang@users.noreply.github.com> Signed-off-by: Xuesong Yang <1646669+XuesongYang@users.noreply.github.com> * Apply suggestion from @XuesongYang Signed-off-by: Xuesong Yang <1646669+XuesongYang@users.noreply.github.com> Signed-off-by: Xuesong Yang <1646669+XuesongYang@users.noreply.github.com> * Apply suggestion from @XuesongYang Signed-off-by: Xuesong Yang <1646669+XuesongYang@users.noreply.github.com> Signed-off-by: Xuesong Yang <1646669+XuesongYang@users.noreply.github.com> * Apply suggestion from @XuesongYang Signed-off-by: Xuesong Yang <1646669+XuesongYang@users.noreply.github.com> Signed-off-by: Xuesong Yang <1646669+XuesongYang@users.noreply.github.com> * Fix OnlinePO test: escape a special character in command line Signed-off-by: Fejgin, Roy <rfejgin@nvidia.com> * Easier-to-read way to quote a special character in OnlinePO test Signed-off-by: Fejgin, Roy <rfejgin@nvidia.com> * Work around ASR Lhotse issue ... and remove some debug code. Signed-off-by: Fejgin, Roy <rfejgin@nvidia.com> * Remove FCD metric for now Signed-off-by: Fejgin, Roy <rfejgin@nvidia.com> * Remove unused import Signed-off-by: Fejgin, Roy <rfejgin@nvidia.com> * Update examples/tts/conf/magpietts/magpietts_lhotse.yaml Signed-off-by: Fejgin, Roy <rfejgin@nvidia.com> Signed-off-by: Roy Fejgin <rfejgin@nvidia.com> --------- Signed-off-by: Jason <jasoli@nvidia.com> Signed-off-by: Xuesong Yang <1646669+XuesongYang@users.noreply.github.com> Signed-off-by: Fejgin, Roy <rfejgin@nvidia.com> Signed-off-by: Roy Fejgin <rfejgin@nvidia.com> Co-authored-by: Xuesong Yang <1646669+XuesongYang@users.noreply.github.com> Co-authored-by: Fejgin, Roy <rfejgin@nvidia.com> Signed-off-by: genquan9 <genquan@google.com> * Revert "Fix vlm engine changes in mcore (#15076)" (#15090) This reverts commit b557cfd. Signed-off-by: genquan9 <genquan@google.com> * ASR inference: expose RNN-T decoding params for context biasing (#15091) * ASR Inference: load decoding params from config for RNN-T Signed-off-by: Vladimir Bataev <vbataev@nvidia.com> Signed-off-by: genquan9 <genquan@google.com> * Fix vlm engine changes in mcore (#15076) Signed-off-by: genquan9 <genquan@google.com> * Revert "Fix vlm engine changes in mcore (#15076)" (#15090) This reverts commit b557cfd. Signed-off-by: genquan9 <genquan@google.com> * update notebook (#15093) Signed-off-by: nithinraok <nithinrao.koluguri@gmail.com> Signed-off-by: genquan9 <genquan@google.com> * fix lines with malformed anchor tags (#15095) Signed-off-by: Pablo Garay <pagaray@nvidia.com> Signed-off-by: genquan9 <genquan@google.com> * add copyright header for missing files Signed-off-by: genquan9 <genquan@google.com> --------- Signed-off-by: taejinp <tango4j@gmail.com> Signed-off-by: tango4j <tango4j@users.noreply.github.com> Signed-off-by: chtruong814 <chtruong814@users.noreply.github.com> Signed-off-by: genquan9 <genquan@google.com> Signed-off-by: genquan9 <genquan9@users.noreply.github.com> Signed-off-by: naymaraq <dkaramyan@nvidia.com> Signed-off-by: nithinraok <nithinrao.koluguri@gmail.com> Signed-off-by: nithinraok <nithinraok@users.noreply.github.com> Signed-off-by: jenchen13 <jennifchen@nvidia.com> Signed-off-by: Pablo Garay <pagaray@nvidia.com> Signed-off-by: pablo-garay <pablo-garay@users.noreply.github.com> Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Signed-off-by: Charlie Truong <chtruong@nvidia.com> Signed-off-by: stevehuang52 <heh@nvidia.com> Signed-off-by: He Huang (Steve) <105218074+stevehuang52@users.noreply.github.com> Signed-off-by: Dong Hyuk Chang <donghyukc@nvidia.com> Signed-off-by: thomasdhc <thomasdhc@users.noreply.github.com> Signed-off-by: dimapihtar <dpihtar@gmail.com> Signed-off-by: dimapihtar <dimapihtar@users.noreply.github.com> Signed-off-by: Jason <jasoli@nvidia.com> Signed-off-by: Xuesong Yang <1646669+XuesongYang@users.noreply.github.com> Signed-off-by: Fejgin, Roy <rfejgin@nvidia.com> Signed-off-by: Roy Fejgin <rfejgin@nvidia.com> Signed-off-by: Vladimir Bataev <vbataev@nvidia.com> Co-authored-by: Taejin Park <tango4j@gmail.com> Co-authored-by: tango4j <tango4j@users.noreply.github.com> Co-authored-by: chtruong814 <chtruong814@users.noreply.github.com> Co-authored-by: genquan9 <genquan9@users.noreply.github.com> Co-authored-by: Dav Karamyan <47416614+naymaraq@users.noreply.github.com> Co-authored-by: naymaraq <dkaramyan@nvidia.com> Co-authored-by: Nithin Rao <nithinrao.koluguri@gmail.com> Co-authored-by: nithinraok <nithinraok@users.noreply.github.com> Co-authored-by: Dmytro Pykhtar <37850217+dimapihtar@users.noreply.github.com> Co-authored-by: Pablo Garay <palenq@gmail.com> Co-authored-by: Jenny Chen <jennifchen@nvidia.com> Co-authored-by: pablo-garay <pablo-garay@users.noreply.github.com> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Charlie Truong <chtruong@nvidia.com> Co-authored-by: He Huang (Steve) <105218074+stevehuang52@users.noreply.github.com> Co-authored-by: Dong Hyuk Chang <thomaschang26@tutanota.com> Co-authored-by: thomasdhc <thomasdhc@users.noreply.github.com> Co-authored-by: dimapihtar <dimapihtar@users.noreply.github.com> Co-authored-by: meatybobby <bobchen@nvidia.com> Co-authored-by: Jason <jasoli@nvidia.com> Co-authored-by: Xuesong Yang <1646669+XuesongYang@users.noreply.github.com> Co-authored-by: Fejgin, Roy <rfejgin@nvidia.com> Co-authored-by: Vladimir Bataev <vbataev@nvidia.com>
diff --git a/examples/voice_agent/server/server_configs/default.yaml b/examples/voice_agent/server/server_configs/default.yaml
@@ -51,4 +51,4 @@ tts:
   type: kokoro # choices in ['nemo', 'kokoro']
   model: "hexgrad/Kokoro-82M"
   model_config: "./server_configs/tts_configs/kokoro_82M.yaml"
-  device: "cuda"
+  device: "cuda"
diff --git a/nemo/collections/llm/gpt/model/gemma3.py b/nemo/collections/llm/gpt/model/gemma3.py
@@ -21,6 +21,7 @@
 from pathlib import Path
 from typing import TYPE_CHECKING, Annotated, Callable, Optional, Tuple, Union
 
+import torch
 from megatron.core.fusions.fused_bias_dropout import get_bias_dropout_add
 from megatron.core.inference.contexts import BaseInferenceContext
 from megatron.core.models.common.embeddings.language_model_embedding import LanguageModelEmbedding
@@ -32,7 +33,6 @@
 from megatron.core.transformer.mlp import MLP, MLPSubmodules
 from torch import Tensor, nn
 
-from nemo.collections.llm.fn.activation import openai_gelu
 from nemo.collections.llm.gpt.model.base import GPTConfig, GPTModel
 from nemo.collections.llm.gpt.model.gemma2 import TERowParallelLinearLayerNorm
 from nemo.collections.llm.utils import Config
@@ -146,16 +146,19 @@ class Gemma3Config(GPTConfig):
     attention_backend: AttnBackend = AttnBackend.flash
 
     # mlp
+    bias_activation_fusion: bool = True
     gated_linear_unit: bool = True
     add_bias_linear: bool = False
-    activation_func: Callable = openai_gelu
+    activation_func: Callable = torch.nn.functional.gelu
 
     # Do not change
     is_vision_language: bool = False
     flash_decode: bool = False
     gradient_accumulation_fusion: bool = False
     transformer_layer_spec: Union[ModuleSpec, Callable[["GPTConfig"], ModuleSpec]] = gemma3_layer_spec
     scatter_embedding_sequence_parallel: bool = True
+    apply_rope_fusion: bool = True
+    cross_entropy_fusion_impl: str = 'te'
 
     def configure_model(
         self,
@@ -338,7 +341,12 @@ def forward(self, max_seq_len: int, offset: int = 0, packed_seq: bool = False) -
         """Get global and local rope embedding"""
         rope_global = super().forward(max_seq_len, offset, packed_seq)
         rope_local = self.rope_local.forward(max_seq_len, offset, packed_seq)
-        return rope_local, rope_global
+        # when using recompute_granularity is full, save_for_backward is called
+        # to save all variables in a layer. It can only save variables but not
+        # tuples.
+        # Stack rope_local and rope_global into a single tensor to avoid the
+        # error.
+        return torch.stack((rope_local, rope_global), dim=0)
 
 
 def _is_local_attn_layer(
@@ -372,7 +380,6 @@ def forward(
         inference_params: Optional[BaseInferenceContext] = None,
     ) -> Tuple[Tensor, Tensor]:
         """Switch to either local or global rope embedding before forward"""
-        assert isinstance(rotary_pos_emb, tuple)
         assert rotary_pos_cos is None and rotary_pos_sin is None
 
         if _is_local_attn_layer(self.layer_number, self.config.interleaved_attn_pattern):
@@ -614,6 +621,7 @@ def config(self):
             architectures=["Gemma3ForCausalLM"],
             num_hidden_layers=source.num_layers,
             hidden_size=source.hidden_size,
+            sliding_window=source.window_size,
             intermediate_size=source.ffn_hidden_size,
             num_attention_heads=source.num_attention_heads,
             head_dim=source.kv_channels,
diff --git a/nemo/collections/vlm/gemma3vl/data/mock.py b/nemo/collections/vlm/gemma3vl/data/mock.py
@@ -186,13 +186,19 @@ def __getitem__(self, idx) -> Dict[str, torch.Tensor]:
         generation_prompt_size = 5  # "ASSISTANT:" like
         prompt_end_idx = img_start_idx + IMAGE_TOKENS + generation_prompt_size
         labels[:prompt_end_idx] = -100
+        # Add the labels clipping to the mock data loader.
         labels = labels[1:]
 
+        # 5) prepare loss masks
+        # Calculate loss mask from labels, to be consistent with real data and reduce confusions.
+        loss_mask = torch.ones_like(labels, dtype=torch.float)
+        loss_mask[labels < 0] = 0.0
+
         return {
             "input_ids": input_ids,
             "position_ids": position_ids,
             "pixel_values": pixel_values,
-            "loss_mask": self.loss_mask,
+            "loss_mask": loss_mask,
             "labels": labels,
         }
 
diff --git a/nemo/collections/vlm/gemma3vl/data/task_encoder.py b/nemo/collections/vlm/gemma3vl/data/task_encoder.py
@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import json
 import logging
 from dataclasses import dataclass, field
 from typing import Optional
@@ -24,6 +25,7 @@
 from nemo.collections.vlm.data.task_encoder import TaskEncoder as BaseTaskEncoder
 from nemo.collections.vlm.data.task_encoder import TaskEncoderConfig as BaseTaskEncoderConfig
 from nemo.collections.vlm.data.utils import _find_pattern_indices
+from nemo.utils import logging
 
 
 @dataclass
@@ -101,58 +103,65 @@ def encode_batch(self, batch_data: DataBatch) -> dict:
         batch_data["media"] = batch_data["media"].reshape(-1, *batch_data["media"].shape[2:])
         return batch_data
 
-    def encode_vqa_sample(self, input_sample: VQASample) -> DataSample:
-        """Encode a VQA sample into a DataSample format.
+    def encode_vqa_sample_multi_turns(self, input_sample: VQASample):
+        """Encode a VQA sample multi turns into a DataSample format.
 
         Args:
             input_sample (VQASample): Input VQA sample containing image, context and answers
 
         Returns:
-            DataSample: Encoded sample with processed image, tokens, labels and loss mask
+            Encoded tokens, labels and images.
         """
+        images = input_sample.image if isinstance(input_sample.image, list) else [input_sample.image]
+
+        contexts = json.loads(input_sample.context.decode('utf-8'))
         messages = []
         if self.config.system_prompt:
             messages.append({'role': 'system', 'content': self.config.system_prompt})
-
-        # Ensure context and answers are lists for consistent processing
-        contexts = input_sample.context if isinstance(input_sample.context, list) else [input_sample.context]
-        answers = input_sample.answers if isinstance(input_sample.answers, list) else [input_sample.answers]
-
-        # Build the conversation messages, replacing image placeholder
-        min_length = min(len(contexts), len(answers))
-        for i in range(min_length):
-            context_with_placeholder = contexts[i].replace("<image>", self.config.image_token)
-            messages.append({'role': self.config.roles[0], 'content': context_with_placeholder})
-            messages.append({'role': self.config.roles[1], 'content': answers[i]})
+        for context in contexts:
+            messages.append(context)
 
         # Apply chat template and process with HF processor
-        converted_messages = self.hf_processor.apply_chat_template(messages, tokenize=False)
+        # `add_generation_prompt=False` because we're providing the full ground truth sequence
+        # We remove the <bos> token using removeprefix('<bos>') since we're finetuning.
+        # The Processor will add this token before training and the model expects only one.
+        converted_messages = self.hf_processor.apply_chat_template(
+            messages, add_generation_prompt=False, tokenize=False
+        ).removeprefix('<bos>')
         outputs = self.hf_processor(
-            images=input_sample.image,
+            images=images,
             text=converted_messages,
             return_tensors="pt",
             images_kwargs={"do_rescale": False},
         )
-
         # Get tokens and images from processor output
         # Squeeze the batch dimension as we process one sample at a time
         tokens = outputs["input_ids"].squeeze(0)
         images = outputs.get("pixel_values")  # Use .get() for optional images
 
         # --- Label Generation ---
+        # Same as: https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/vlm/qwen2vl/data/task_encoder.py#L263-L270
         # Initialize labels with ignore placeholder
         labels = torch.full_like(tokens, self.config.ignore_place_holder)
-
         search_start_index = 0
-        for answer in answers:
+        for context in contexts:
+            if context['role'] != 'assistant':
+                continue
             # Tokenize the answer, including the stop string if provided
-            answer_with_stop = answer + (self.config.stop_string or "")
+            answer_with_stop = (
+                context['content'][0]['text'].rstrip().lstrip() + "<end_of_turn>" + (self.config.stop_string or "")
+            )
+            answer_with_stop = answer_with_stop.rstrip().lstrip()
             answer_tokens = self.tokenizer.tokenizer(answer_with_stop, add_special_tokens=False)["input_ids"]
             answer_tokens_tensor = torch.tensor(answer_tokens, device=tokens.device)  # Ensure same device
 
+            # sometimes the tokenizer can add additional space. See:
+            # https://github.com/huggingface/transformers/issues/25073#issuecomment-1655271420
+            if self.tokenizer.tokenizer.decode(answer_tokens[0]) == "":
+                answer_tokens_tensor = answer_tokens_tensor[1:]
+
             # Find answer pattern in tokens
             answer_start, answer_end = _find_pattern_indices(tokens, answer_tokens_tensor, search_start_index)
-
             if answer_start >= 0:
                 labels[answer_start:answer_end] = tokens[answer_start:answer_end]
                 search_start_index = answer_end
@@ -170,11 +179,24 @@ def encode_vqa_sample(self, input_sample: VQASample) -> DataSample:
                     search_start_index,
                 )
                 break
+        return tokens, labels, images
+
+    def encode_vqa_sample(self, input_sample: VQASample) -> DataSample:
+        """Encode a VQA sample into a DataSample format.
+
+        Args:
+            input_sample (VQASample): Input VQA sample containing image, context and answers
+
+        Returns:
+            DataSample: Encoded sample with processed image, tokens, labels and loss mask
+        """
+        tokens, labels, images = self.encode_vqa_sample_multi_turns(input_sample)
 
         # Prepare final tensors
         tokens = tokens[:-1].contiguous()
         labels = labels[1:].contiguous()
         seqlen = len(tokens)  # Original sequence length before padding
+        position_ids = torch.arange(seqlen, dtype=torch.int64)
 
         # Pad tokens and labels to a multiple of `pad_to_multiple_of` if specified
         if self.config.pad_to_multiple_of:
@@ -191,7 +213,7 @@ def encode_vqa_sample(self, input_sample: VQASample) -> DataSample:
 
         # Compute loss mask
         loss_mask = torch.ones_like(labels, dtype=torch.float)
-        loss_mask[labels == self.config.ignore_place_holder] = 0.0
+        loss_mask[labels < 0] = 0.0
 
         # Convert images to bfloat16 and stack, or create an empty tensor if no images
         if images is not None and images.numel() > 0:
@@ -202,13 +224,16 @@ def encode_vqa_sample(self, input_sample: VQASample) -> DataSample:
             # Create an empty tensor with appropriate dimensions and dtype if no images
             processed_image = None
 
-        return Gemma3DataSample(
+        sample = Gemma3DataSample(
             __key__=input_sample.__key__,
             __restore_key__=input_sample.__restore_key__,
             __subflavor__=input_sample.__subflavor__,
             __subflavors__=input_sample.__subflavors__,
             pixel_values=processed_image,
             input_ids=tokens,
+            position_ids=position_ids,
             labels=labels,
             loss_mask=loss_mask,
         )
+
+        return sample
diff --git a/nemo/collections/vlm/gemma3vl/model/base.py b/nemo/collections/vlm/gemma3vl/model/base.py
@@ -37,6 +37,7 @@
 from nemo.collections.llm.gpt.model.gemma3 import Gemma3Config
 from nemo.collections.vlm.gemma3vl.model.vision import Gemma3VLMultimodalProjectorConfig, Gemma3VLVisionConfig
 from nemo.collections.vlm.neva.model.base import MODEL_CONFIG_ATTR, NevaModel, restore_model_weights
+from nemo.collections.vlm.qwen2vl.data.multimodal_tokens import IGNORE_INDEX
 from nemo.lightning import io
 from nemo.lightning.pytorch.optim import OptimizerModule
 from nemo.utils.import_utils import safe_import_from
@@ -78,6 +79,7 @@ def gemma3vl_data_step(dataloader_iter) -> Dict[str, torch.Tensor]:
         key: val.cuda(non_blocking=True) if key in required_keys and val is not None else None
         for key, val in _batch.items()
     }
+
     return _batch
 
 
@@ -392,18 +394,18 @@ def forward(
                 input_ids = F.pad(input_ids, (0, padded_seq_len))
                 position_ids = F.pad(position_ids, (0, padded_seq_len))
             if self.post_process:
-                labels = F.pad(labels, (0, padded_seq_len))
-                loss_mask = F.pad(loss_mask, (0, padded_seq_len))
+                labels = F.pad(labels, (0, padded_seq_len), value=IGNORE_INDEX)
+                loss_mask = F.pad(loss_mask, (0, padded_seq_len), value=0.0)
 
         # Compute language embedding
         if self.pre_process:
             safe_input_ids = input_ids
             # Replace image_token_id with 0 to avoid embedding index error
-            if self.image_token_id >= self.vocab_size:
-                image_token_mask = input_ids == self.image_token_id
-                safe_input_ids = input_ids.clone()
-                safe_input_ids[image_token_mask] = 0
+            image_token_mask = input_ids == self.image_token_id
+            safe_input_ids = input_ids.clone()
+            safe_input_ids[image_token_mask] = 0
             # (T, B, D)
+            # The position_ids is None for qwen2 models, but set to position_ids for gemma3vl models.
             language_embedding = self.language_model.embedding(input_ids=safe_input_ids, position_ids=position_ids)
             # (B, T, D)
             language_embedding = language_embedding.transpose(1, 0).contiguous()
@@ -428,6 +430,7 @@ def forward(
             combined_embedding = combined_embedding.transpose(1, 0).contiguous()
 
         # Run decoder model
+        # position_ids is None for gemma3vl models, but set to position_ids to qwen2 models.
         output = self.language_model(
             input_ids=None,
             position_ids=None,
@@ -441,6 +444,8 @@ def forward(
 
         if labels is None or loss_mask is None:
             return output
+
+        output = output.masked_fill(labels < 0, 0.0)
         return output, loss_mask
 
     def _preprocess_data(
@@ -536,6 +541,7 @@ def _process_sequence_parallel(
                 combined_embedding = scatter_to_sequence_parallel_region(combined_embedding)
         return combined_embedding, labels, loss_mask, packed_seq_params
 
+    @torch.compile
     def _compute_attention_mask(
         self,
         input_ids: torch.Tensor,
diff --git a/scripts/vlm/gemma3vl_export.py b/scripts/vlm/gemma3vl_export.py
@@ -0,0 +1,83 @@
+# Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Export Gemma3VL NeMo checkpoints to Hugging Face format."""
+
+import argparse
+from pathlib import Path
+
+from huggingface_hub import hf_hub_download
+
+from nemo.collections import llm
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description=("Export NeMo vision language model checkpoint to Hugging Face format.")
+    )
+    parser.add_argument(
+        "--nemo_ckpt_path",
+        type=str,
+        required=True,
+        default=None,
+        help="Path to the NeMo checkpoint directory.",
+    )
+    parser.add_argument(
+        "--output_hf_path",
+        type=str,
+        required=True,
+        default=None,
+        help="Path to save the converted Hugging Face checkpoint.",
+    )
+    parser.add_argument(
+        "--model_name",
+        type=str,
+        required=False,
+        default=None,
+        help="Name of the model on Hugging Face.",
+    )
+
+    args = parser.parse_args()
+
+    llm.export_ckpt(
+        path=Path(args.nemo_ckpt_path),
+        target="hf",
+        output_path=Path(args.output_hf_path),
+        overwrite=True,
+    )
+    if args.model_name:
+        # Copy necessary files if exist from HuggingFace for Gemma3VL model export.
+        copy_file_list = [
+            "preprocessor_config.json",
+            "chat_template.json",
+            "config.json",
+            "generation_config.json",
+            "merges.txt",
+            "tokenizer.json",
+            "tokenizer_config.json",
+            "vocab.json",
+        ]
+        for file_name in copy_file_list:
+            try:
+                downloaded_path = hf_hub_download(
+                    repo_id=args.model_name,
+                    filename=file_name,
+                    local_dir=args.output_hf_path,
+                )
+                print(f"Downloaded {downloaded_path} during export gamma3vl models.")
+            except Exception as e:
+                print(f"Ignore {file_name} during export gamma3vl models.")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/scripts/vlm/gemma3vl_finetune.py b/scripts/vlm/gemma3vl_finetune.py
diff --git a/scripts/vlm/gemma3vl_import.py b/scripts/vlm/gemma3vl_import.py
diff --git a/scripts/vlm/gemma3vl_utils.py b/scripts/vlm/gemma3vl_utils.py
