feat: Add Gemma3 text-only model support

Signed-off-by: Balaram Buddharaju <[email protected]>

Author: brb-nv

examples/gemma/README.md

@@ -24,6 +24,8 @@
       - [Run inference under INT8 KV caches for keras checkpoint](#run-inference-under-int8-kv-caches-for-keras-checkpoint)
     - [Run Gemma 2](#run-gemma-2)
       - [Run inference under bfloat16 for torch checkpoint](#run-inference-under-bfloat16-for-torch-checkpoint-1)
+    - [Run Gemma 3](#run-gemma-3)
+      - [Run inference under bfloat16 for HF checkpoint](#run-inference-under-bfloat16-for-hf-checkpoint-1)
     - [Run Modelopt Quantization](#run-modelopt-quantization)
       - [Requirements](#requirements)
       - [Quantize Checkpoints](#quantize-checkpoints)
@@ -628,6 +630,52 @@ Average accuracy 0.697 - other (business, health, misc.)
 Average accuracy: 0.630
 ```
 
+### Run Gemma 3
+
+Gemma 3's text generation model from HuggingFace is supported. Gemma3 1B model interleaves 5 local layers between each global layer. While local layers use sliding window attention with a short span of 512 tokens, global layers attend to the long context. TRTLLM support layerwise sliding-window attention and the sliding window size for each layer could be passed in using the `--max_attention_window_size` parameter at runtime. If a subpattern is provided, TRTLLM can extrapolate the complete pattern and the extrapolation logic is printed to terminal.
+
+#### Run inference under bfloat16 for HF checkpoint
+```bash
+git clone https://huggingface.co/google/gemma-3-1b-it
+
+CKPT_PATH=gemma-3-1b-it/
+UNIFIED_CKPT_PATH=/tmp/checkpoints/tmp_1b_it_tensorrt_llm/bf16/tp1/
+ENGINE_PATH=/tmp/gemma3/1b/bf16/1-gpu/
+VOCAB_FILE_PATH=gemma-3-1b-it/tokenizer.model
+
+python3 ./examples/gemma/convert_checkpoint.py \
+    --ckpt-type hf \
+    --model-dir ${CKPT_PATH} \
+    --dtype bfloat16 \
+    --world-size 1 \
+    --output-model-dir ${UNIFIED_CKPT_PATH}
+
+trtllm-build --checkpoint_dir ${UNIFIED_CKPT_PATH} \
+             --gemm_plugin auto \
+             --max_batch_size 8 \
+             --max_input_len 3000 \
+             --max_seq_len 3100 \
+             --output_dir ${ENGINE_PATH}
+
+python3 ./examples/summarize.py --test_trt_llm \
+                      --vocab_file ${VOCAB_FILE_PATH} \
+                      --engine_dir ${ENGINE_PATH} \
+                      --batch_size 1 \
+                      --max_ite 5 \
+                      --max_attention_window_size 512 512 512 512 512 3100
+
+...
+[TensorRT-LLM][INFO] TRTGptModel mMaxAttentionWindowSize: (512, 512, 512, 512, 512, 3100) * 4 + (512, 512)
+...
+[04/09/2025-18:28:26] [TRT-LLM] [I] TensorRT-LLM (total latency: 1.6197962760925293 sec)
+[04/09/2025-18:28:26] [TRT-LLM] [I] TensorRT-LLM (total output tokens: 475)
+[04/09/2025-18:28:26] [TRT-LLM] [I] TensorRT-LLM (tokens per second: 293.2467539349165)
+[04/09/2025-18:28:26] [TRT-LLM] [I] TensorRT-LLM beam 0 result
+[04/09/2025-18:28:26] [TRT-LLM] [I]   rouge1: 22.780314381954003
+[04/09/2025-18:28:26] [TRT-LLM] [I]   rouge2: 4.331099231480823
+[04/09/2025-18:28:26] [TRT-LLM] [I]   rougeL: 15.26751867562475
+[04/09/2025-18:28:26] [TRT-LLM] [I]   rougeLsum: 20.14696930976001
+```
 
 ### Run Modelopt Quantization
 

tensorrt_llm/layers/attention.py

@@ -175,6 +175,13 @@ def __init__(self,
         self.embed_positions = None
         self.rotary_inv_freq = None
         self.embed_positions_for_gpt_attention = None
+
+        # auxiliary params to support models with non-homegeneous attn layers requiring
+        # a different set of rope params. e.g. Gemma3.
+        self.embed_positions_local = None
+        self.rotary_inv_freq_local = None
+        self.embed_positions_for_gpt_attention_local = None
+
         # long rope const parameters
         self.long_rope_embed_positions = None
         self.long_rope_rotary_inv_freq = None
@@ -186,10 +193,16 @@ def fill_attention_const_params_for_rope(
             self,
             embed_positions: Tensor = None,
             rotary_inv_freq: Tensor = None,
-            embed_positions_for_gpt_attention: Tensor = None):
+            embed_positions_for_gpt_attention: Tensor = None,
+            embed_positions_local: Tensor = None,
+            rotary_inv_freq_local: Tensor = None,
+            embed_positions_for_gpt_attention_local: Tensor = None):
         self.embed_positions = embed_positions
         self.rotary_inv_freq = rotary_inv_freq
         self.embed_positions_for_gpt_attention = embed_positions_for_gpt_attention
+        self.embed_positions_local = embed_positions_local
+        self.rotary_inv_freq_local = rotary_inv_freq_local
+        self.embed_positions_for_gpt_attention_local = embed_positions_for_gpt_attention_local
         return self
 
     def fill_attention_const_params_for_long_rope(
@@ -359,6 +372,7 @@ def __init__(self,
                  dtype=None,
                  position_embedding_type=PositionEmbeddingType.learned_absolute,
                  rotary_embedding_base=10000.0,
+                 rotary_embedding_base_local=1.0,
                  rotary_embedding_scaling=None,
                  rotary_embedding_percentage=1.0,
                  rope_scaling_short_factors=None,
@@ -388,7 +402,8 @@ def __init__(self,
                  cp_size=1,
                  cp_rank=0,
                  max_seqlen_for_logn_scaling=8192,
-                 use_logn_scaling=False):
+                 use_logn_scaling=False,
+                 is_local=False):
         super().__init__()
 
         self.local_layer_idx = local_layer_idx
@@ -417,6 +432,7 @@ def __init__(self,
         self.cp_group = cp_group
         self.cp_size = cp_size
         self.cp_rank = cp_rank
+        self.is_local = is_local
 
         self.num_layers = num_layers
         self.apply_query_key_layer_scaling = apply_query_key_layer_scaling
@@ -437,6 +453,7 @@ def __init__(self,
         self.max_distance = max_distance
         self.num_buckets = num_buckets
         self.rotary_embedding_base = rotary_embedding_base
+        self.rotary_embedding_base_local = rotary_embedding_base_local
         self.rotary_embedding_scaling = rotary_embedding_scaling
         self.rotary_embedding_scale_type = RotaryScalingType.none
         self.rotary_embedding_scale = 1.0
@@ -656,26 +673,45 @@ def create_attention_const_params(model_cls, config):
                 model_cls.short_mscale = short_mscale
                 model_cls.long_mscale = long_mscale
         else:
-            # Rotary const weights.
-            embed_positions = RopeEmbeddingUtils.create_sinusoidal_positions(
-                max_position_embeddings,
-                rotary_embedding_dim,
-            )
-            rotary_inv_freq, embed_positions_for_gpt_attention = RopeEmbeddingUtils.create_sinusoidal_positions_for_attention_plugin(
-                max_position_embeddings, rotary_embedding_dim,
-                rotary_embedding_base, rotary_embedding_scale,
-                rotary_embedding_scale_type, rotary_embedding_scaling)
-            model_cls.register_parameter(
-                'embed_positions',
-                Parameter(embed_positions, dtype='float32', is_buffer=True))
-            model_cls.register_parameter(
-                'rotary_inv_freq',
-                Parameter(rotary_inv_freq, dtype='float32', is_buffer=True))
-            model_cls.register_parameter(
-                'embed_positions_for_gpt_attention',
-                Parameter(embed_positions_for_gpt_attention,
-                          dtype='float32',
-                          is_buffer=True))
+
+            def register_rope_params(rotary_base, names_to_register):
+                # Rotary const weights.
+                embed_positions = RopeEmbeddingUtils.create_sinusoidal_positions(
+                    max_position_embeddings,
+                    rotary_embedding_dim,
+                )
+                rotary_inv_freq, embed_positions_for_gpt_attention = RopeEmbeddingUtils.create_sinusoidal_positions_for_attention_plugin(
+                    max_position_embeddings, rotary_embedding_dim, rotary_base,
+                    rotary_embedding_scale, rotary_embedding_scale_type,
+                    rotary_embedding_scaling)
+                model_cls.register_parameter(
+                    names_to_register[0],
+                    Parameter(embed_positions, dtype='float32', is_buffer=True))
+                model_cls.register_parameter(
+                    names_to_register[1],
+                    Parameter(rotary_inv_freq, dtype='float32', is_buffer=True))
+                model_cls.register_parameter(
+                    names_to_register[2],
+                    Parameter(embed_positions_for_gpt_attention,
+                              dtype='float32',
+                              is_buffer=True))
+
+            register_rope_params(rotary_base=rotary_embedding_base,
+                                 names_to_register=[
+                                     'embed_positions', 'rotary_inv_freq',
+                                     'embed_positions_for_gpt_attention'
+                                 ])
+
+            # For models with non-homegeneous attention layers requiring a second set of rope params. e.g. Gemma3.
+            rotary_embedding_base_local = getattr(config,
+                                                  'rope_local_base_freq', None)
+            if rotary_embedding_base_local is not None:
+                register_rope_params(
+                    rotary_base=rotary_embedding_base_local,
+                    names_to_register=[
+                        'embed_positions_local', 'rotary_inv_freq_local',
+                        'embed_positions_for_gpt_attention_local'
+                    ])
 
     @staticmethod
     def fill_attention_params(model_cls, attention_params):
@@ -695,7 +731,15 @@ def fill_attention_params(model_cls, attention_params):
                 return attention_params.fill_attention_const_params_for_rope(
                     model_cls.embed_positions.value,
                     model_cls.rotary_inv_freq.value,
-                    model_cls.embed_positions_for_gpt_attention.value)
+                    model_cls.embed_positions_for_gpt_attention.value,
+                    model_cls.embed_positions_local.value if hasattr(
+                        model_cls, "embed_positions_local") else None,
+                    model_cls.rotary_inv_freq_local.value if hasattr(
+                        model_cls, "rotary_inv_freq_local") else None,
+                    model_cls.embed_positions_for_gpt_attention_local.value
+                    if hasattr(
+                        model_cls,
+                        "embed_positions_for_gpt_attention_local") else None)
         # Fill nothing.
         return attention_params
 
@@ -1020,6 +1064,11 @@ def compute_cross_kv(encoder_output):
             # Rotary cos/sin cache.
             rotary_cos_sin = getattr(attention_params,
                                      "embed_positions_for_gpt_attention", None)
+            rotary_inv_freq_local = getattr(attention_params,
+                                            "rotary_inv_freq_local", None)
+            rotary_cos_sin_local = getattr(
+                attention_params, "embed_positions_for_gpt_attention_local",
+                None)
 
             long_rope_rotary_inv_freq = getattr(attention_params,
                                                 "long_rope_rotary_inv_freq",
@@ -1062,7 +1111,8 @@ def compute_cross_kv(encoder_output):
                 hidden_size_per_head=self.attention_head_size,
                 q_scaling=self.q_scaling,
                 rotary_embedding_dim=self.rotary_embedding_dim,
-                rotary_embedding_base=self.rotary_embedding_base,
+                rotary_embedding_base=self.rotary_embedding_base
+                if not self.is_local else self.rotary_embedding_base_local,
                 rotary_embedding_scale_type=self.rotary_embedding_scale_type,
                 rotary_embedding_short_m_scale=attention_params.short_mscale,
                 rotary_embedding_long_m_scale=attention_params.long_mscale,
@@ -1071,8 +1121,10 @@ def compute_cross_kv(encoder_output):
                 rotary_embedding_original_max_positions=self.
                 original_max_position_embeddings,
                 position_embedding_type=self.position_embedding_type,
-                rotary_inv_freq=rotary_inv_freq,
-                rotary_cos_sin=rotary_cos_sin,
+                rotary_inv_freq=rotary_inv_freq
+                if not self.is_local else rotary_inv_freq_local,
+                rotary_cos_sin=rotary_cos_sin
+                if not self.is_local else rotary_cos_sin_local,
                 kv_orig_quant_scale=kv_orig_quant_scale,
                 kv_quant_orig_scale=kv_quant_orig_scale,
                 attention_output_orig_quant_scale=self.

tensorrt_llm/models/__init__.py

@@ -33,7 +33,8 @@
 from .enc_dec.model import DecoderModel, EncoderModel, WhisperEncoder
 from .falcon.config import FalconConfig
 from .falcon.model import FalconForCausalLM, FalconModel
-from .gemma.config import GEMMA2_ARCHITECTURE, GEMMA_ARCHITECTURE, GemmaConfig
+from .gemma.config import (GEMMA2_ARCHITECTURE, GEMMA3_ARCHITECTURE,
+                           GEMMA_ARCHITECTURE, GemmaConfig)
 from .gemma.model import GemmaForCausalLM
 from .gpt.config import GPTConfig
 from .gpt.model import GPTForCausalLM, GPTModel
@@ -183,6 +184,7 @@
     'SkyworkForCausalLM': LLaMAForCausalLM,
     GEMMA_ARCHITECTURE: GemmaForCausalLM,
     GEMMA2_ARCHITECTURE: GemmaForCausalLM,
+    GEMMA3_ARCHITECTURE: GemmaForCausalLM,
     'QWenLMHeadModel': QWenForCausalLM,
     'QWenForCausalLM': QWenForCausalLM,
     'Qwen2ForCausalLM': QWenForCausalLM,

tensorrt_llm/models/gemma/config.py

@@ -19,6 +19,7 @@
 from tensorrt_llm.mapping import Mapping
 from tensorrt_llm.models.convert_utils import infer_dtype
 from tensorrt_llm.models.modeling_utils import (Gemma2ConfigGroup,
+                                                Gemma3ConfigGroup,
                                                 PretrainedConfig, QuantConfig)
 
 if TYPE_CHECKING:
@@ -30,6 +31,7 @@
 
 GEMMA_ARCHITECTURE = "GemmaForCausalLM"
 GEMMA2_ARCHITECTURE = "Gemma2ForCausalLM"
+GEMMA3_ARCHITECTURE = "Gemma3ForCausalLM"
 
 
 class GemmaConfig(PretrainedConfig):
@@ -48,6 +50,9 @@ def __init__(
         final_logit_softcapping: Optional[float] = None,
         attn_logit_softcapping: Optional[float] = None,
         mapping: Optional[Union[Mapping, dict]] = None,
+        sliding_window_pattern: int = None,
+        rope_local_base_freq: int = None,
+        sliding_window: int = None,
         **kwargs,
     ):
         use_parallel_embedding = False
@@ -79,23 +84,29 @@ def __init__(
         self.mlp_bias = mlp_bias
 
         self.inter_layernorms = False
-        if self.is_gemma_2:
+        if self.is_gemma_2 or self.is_gemma_3:
             self.inter_layernorms = True
-            assert query_pre_attn_scalar is not None, "Gemma2 models must configure `query_pre_attn_scalar`"
+            assert query_pre_attn_scalar is not None, "Gemma2 / Gemma3 models must configure `query_pre_attn_scalar`"
             self.query_pre_attn_scalar = query_pre_attn_scalar
             self.final_logit_softcapping = final_logit_softcapping
-            self.attn_logit_softcapping = attn_logit_softcapping
+            if self.is_gemma_2:
+                self.attn_logit_softcapping = attn_logit_softcapping
+            if self.is_gemma_3:
+                self.sliding_window_pattern = sliding_window_pattern
+                self.rope_local_base_freq = rope_local_base_freq
+                self.sliding_window = sliding_window
 
     GEMMA_ADDED_FIELDS = {
         "rotary_base", "rotary_scaling", "attn_bias", "mlp_bias",
         "inter_layernorms"
     }
     GEMMA2_ADDED_FIELDS = Gemma2ConfigGroup.keys()
+    GEMMA3_ADDED_FIELDS = Gemma3ConfigGroup.keys()
     VERBATIM = {
         "num_hidden_layers", "num_attention_heads", "hidden_size",
         "intermediate_size", "vocab_size", "max_position_embeddings",
         "hidden_act", "use_parallel_embedding"
-    } | GEMMA2_ADDED_FIELDS
+    } | GEMMA2_ADDED_FIELDS | GEMMA3_ADDED_FIELDS
 
     @property
     def is_gemma_2(self) -> bool:
@@ -106,6 +117,15 @@ def gemma2_config(self):
             return self.get_config_group(Gemma2ConfigGroup)
         return None
 
+    @property
+    def is_gemma_3(self) -> bool:
+        return self.architecture == GEMMA3_ARCHITECTURE
+
+    def gemma3_config(self):
+        if self.is_gemma_3:
+            return self.get_config_group(Gemma3ConfigGroup)
+        return None
+
     def to_dict(self):
         """Serialize the fields added in GemmaConfig"""
 
@@ -118,7 +138,11 @@ def to_dict(self):
             **({
                 f: getattr(self, f)
                 for f in self.GEMMA2_ADDED_FIELDS
-            } if self.is_gemma_2 else {})
+            } if self.is_gemma_2 else {}),
+            **({
+                f: getattr(self, f)
+                for f in self.GEMMA3_ADDED_FIELDS
+            } if self.is_gemma_3 else {})
         }
 
     @classmethod
@@ -148,6 +172,7 @@ def from_hugging_face(
             norm_epsilon=hf_config.rms_norm_eps,
             num_key_value_heads=getattr(hf_config, "num_key_value_heads",
                                         hf_config.num_attention_heads),
+            rotary_base=getattr(hf_config, "rope_theta", 10000.0),
             rotary_scaling=getattr(hf_config, "rotary_scaling", None),
             quantization=quant_config,
             mapping=mapping,

tensorrt_llm/models/gemma/convert.py

@@ -317,6 +317,10 @@ def rename_to_trt_llm(self, name: str) -> Optional[str]:
              None),  # merged with above
             (r"model.layers.(\d+).self_attn.o_proj.weight",
              r"layers.\1.attention.dense.weight"),
+            (r"model.layers.(\d+).self_attn.q_norm.weight",
+             r"layers.\1.attention.q_layernorm.weight"),
+            (r"model.layers.(\d+).self_attn.k_norm.weight",
+             r"layers.\1.attention.k_layernorm.weight"),
             (r"model.layers.(\d+).mlp.gate_proj.weight",
              r"layers.\1.mlp.fc.weight"),
             (r"model.layers.(\d+).mlp.up_proj.weight",
@@ -795,6 +799,8 @@ def load_gemma_weights(
                 "pre_feedforward_layernorm",
                 "post_feedforward_layernorm",
                 "model.norm.weight",
+                "q_norm.weight",
+                "k_norm.weight",
         )):
             param = param + 1.0  # upcasted to float32 in case of bfloat16
             add_trt_llm_weight(weights, trt_llm_name, param,