optimum-rbln 0.7.5a1__py3-none-any.whl → 0.7.5rc0__py3-none-any.whl

optimum/rbln/transformers/models/gemma3/configuration_gemma3.py

@@ -0,0 +1,69 @@
+# Copyright 2025 Rebellions Inc. 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.
+
+from typing import Optional
+
+from ....configuration_utils import RBLNModelConfig
+from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelForCausalLMConfig
+from ..siglip.configuration_siglip import RBLNSiglipVisionModelConfig
+
+
+class RBLNGemma3ForCausalLMConfig(RBLNDecoderOnlyModelForCausalLMConfig):
+    def __init__(
+        self,
+        prefill_chunk_size: Optional[int] = None,
+        use_position_ids: Optional[bool] = None,
+        use_attention_mask: Optional[bool] = None,
+        **kwargs,
+    ):
+        # use_attention_mask and use_position_ids are always True for Gemma3
+        use_attention_mask = use_attention_mask or True
+        use_position_ids = use_position_ids or True
+        prefill_chunk_size = prefill_chunk_size or 256
+
+        super().__init__(
+            prefill_chunk_size=prefill_chunk_size,
+            use_attention_mask=use_attention_mask,
+            use_position_ids=use_position_ids,
+            **kwargs,
+        )
+
+
+class RBLNGemma3ForConditionalGenerationConfig(RBLNModelConfig):
+    submodules = ["vision_tower", "language_model"]
+
+    def __init__(
+        self,
+        batch_size: Optional[int] = None,
+        vision_tower: Optional[RBLNModelConfig] = None,
+        language_model: Optional[RBLNModelConfig] = None,
+        **kwargs,
+    ):
+        """
+        Args:
+            batch_size (Optional[int]): The batch size for inference. Defaults to 1.
+            vision_tower (Optional[RBLNModelConfig]): Configuration for the vision encoder component.
+            language_model (Optional[RBLNModelConfig]): Configuration for the language model component.
+            **kwargs: Additional arguments passed to the parent RBLNModelConfig.
+
+        Raises:
+            ValueError: If batch_size is not a positive integer.
+        """
+        super().__init__(**kwargs)
+        self.batch_size = batch_size or 1
+        if not isinstance(self.batch_size, int) or self.batch_size < 0:
+            raise ValueError(f"batch_size must be a positive integer, got {self.batch_size}")
+
+        self.vision_tower = self.init_submodule_config(RBLNSiglipVisionModelConfig, vision_tower)
+        self.language_model = self.init_submodule_config(RBLNGemma3ForCausalLMConfig, language_model)

optimum/rbln/transformers/models/gemma3/gemma3_architecture.py

@@ -0,0 +1,446 @@
+# Copyright 2025 Rebellions Inc. 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.
+
+import copy
+from typing import TYPE_CHECKING, Optional, Tuple, Union
+
+import torch
+from torch import nn
+from transformers.models.gemma3.modeling_gemma3 import Gemma3RMSNorm
+
+from ..decoderonly.decoderonly_architecture import (
+    AttentionOp,
+    DecoderOnlyAttention,
+    DecoderOnlyFlashAttention,
+    DecoderOnlyForCausalLM,
+    DecoderOnlyLayer,
+    DecoderOnlyModel,
+    DecoderOnlyWrapper,
+    RotaryEmbedding,
+    SlidingWindowAttentionOp,
+    slice_and_unsqueeze_cos_sin,
+)
+
+
+if TYPE_CHECKING:
+    from transformers import Gemma3ForCausalLM
+
+
+class Gemma3ForCausalLMWrapper(DecoderOnlyWrapper):
+    def get_rotary_emb(self, max_seq_len):
+        rotary_emb_global = RotaryEmbedding(config=self.config, max_seq_len_cached=max_seq_len)
+
+        config = copy.deepcopy(self.config)
+        config.rope_theta = config.rope_local_base_freq
+        config.rope_scaling = {"rope_type": "default"}
+        rotary_emb_local = RotaryEmbedding(config=config, max_seq_len_cached=max_seq_len)
+
+        return (rotary_emb_global, rotary_emb_local)
+
+    def convert_to_rbln_causal_lm(self, causal_lm: "Gemma3ForCausalLM", max_seq_len: int):
+        new_layers = []
+        for layer in causal_lm.model.layers:
+            if layer.is_sliding:
+                new_self_attn = Gemma3Attention(
+                    layer.self_attn,
+                    use_attention_mask=None,  # FIXME: no use in SWA
+                    use_position_ids=self.use_position_ids,
+                    kvcache_block_size=self.config.sliding_window,
+                )
+            else:
+                if self.attn_impl == "eager":
+                    new_self_attn = Gemma3Attention(
+                        layer.self_attn,
+                        use_attention_mask=self.use_attention_mask,
+                        use_position_ids=self.use_position_ids,
+                        kvcache_block_size=self.kvcache_block_size,
+                    )
+                elif self.attn_impl == "flash_attn":
+                    new_self_attn = Gemma3FlashAttention(
+                        layer.self_attn,
+                        kvcache_partition_len=self.kvcache_partition_len,
+                        use_attention_mask=self.use_attention_mask,
+                        kvcache_block_size=self.kvcache_block_size,
+                        use_position_ids=self.use_position_ids,
+                    )
+                else:
+                    raise NotImplementedError(f"Unknwon attn : {self.attn_impl}")
+
+            new_layer = Gemma3DecoderLayer(layer, new_self_attn)
+            new_layers.append(new_layer)
+
+        new_model = Gemma3TextModel(
+            causal_lm.model,
+            new_layers,
+            partition_len=self.kvcache_partition_len,
+            max_seq_len=max_seq_len,
+        )
+        new_causal_lm = Gemma3ForCausalLM(causal_lm, new_model)
+        return new_causal_lm
+
+    def forward(self, *args):
+        if self.phase == "decode":
+            (
+                input_ids_or_inputs_embeds,
+                attention_mask,  # used in global layer, 2D attn_mask for padded KVcache.
+                cache_position,
+                position_ids,
+                golbal_block_tables,
+                local_block_tables,
+                *past_key_values,
+            ) = args
+            query_position = None
+
+        elif "prefill" in self.phase:
+            (
+                input_ids_or_inputs_embeds,
+                attention_mask,
+                cache_position,
+                position_ids,
+                query_position,
+                golbal_block_tables,
+                local_block_tables,
+                *past_key_values,
+            ) = args
+
+        else:
+            raise ValueError(f"Unknown phase: {self.phase}")
+
+        if input_ids_or_inputs_embeds.ndim == 2:
+            input_ids = input_ids_or_inputs_embeds
+            inputs_embeds = None
+        elif input_ids_or_inputs_embeds.ndim == 3:
+            input_ids = None
+            inputs_embeds = input_ids_or_inputs_embeds
+        else:
+            raise NotImplementedError(f"Unknown ndim of input : {input_ids_or_inputs_embeds.ndim}")
+
+        if len(past_key_values) != 2 * self.num_hidden_layers:
+            raise ValueError(
+                f"Different past_key_values to model's config. {len(past_key_values)} != {2 * self.num_hidden_layers}"
+            )
+
+        # [key, value] * n_layer -> ( (key, value) ) * n_layer
+        # cache shape : batch, n_heads, 1, max_seq_len, head_dim
+        _past_key_values = []
+        for i in range(self.config.num_hidden_layers):
+            key_states = past_key_values[i * 2]
+            value_states = past_key_values[i * 2 + 1]
+            past_key_value = [key_states, value_states]
+            _past_key_values.append(past_key_value)
+        past_key_values = _past_key_values
+
+        logit = self.causal_lm(
+            input_ids=input_ids,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            cache_position=cache_position,
+            position_ids=position_ids,
+            query_position=query_position,
+            past_key_values=past_key_values,
+            rotary_emb=(self.rotary_emb_global, self.rotary_emb_local),
+            global_block_tables=golbal_block_tables,
+            local_block_tables=local_block_tables,
+        )
+
+        return logit
+
+
+class Gemma3ForCausalLM(DecoderOnlyForCausalLM):
+    def forward(
+        self,
+        input_ids: torch.Tensor = None,
+        inputs_embeds: torch.Tensor = None,
+        attention_mask: torch.Tensor = None,
+        cache_position: torch.Tensor = None,
+        position_ids: torch.Tensor = None,
+        query_position: torch.Tensor = None,
+        past_key_values: Tuple[Tuple[torch.Tensor]] = None,
+        rotary_emb: nn.Module = None,
+        global_block_tables: Optional[torch.Tensor] = None,
+        local_block_tables: Optional[torch.Tensor] = None,
+    ):
+        # outputs
+        hidden_states = self.model(
+            input_ids=input_ids,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            cache_position=cache_position,
+            position_ids=position_ids,
+            query_position=query_position,
+            past_key_values=past_key_values,
+            rotary_emb=rotary_emb,
+            global_block_tables=global_block_tables,
+            local_block_tables=local_block_tables,
+        )
+
+        if "prefill" in self.phase:
+            hidden_states = hidden_states[:, query_position.to(torch.int).unsqueeze(0)]
+
+        logits = self.lm_head(hidden_states)
+
+        # Apply final logit softmaxing if configured, e.g. for Gemma2
+        if getattr(self.config, "final_logit_softcapping", None) is not None:
+            logits = logits / self.config.final_logit_softcapping
+            logits = torch.tanh(logits)
+            logits = logits * self.config.final_logit_softcapping
+
+        return logits
+
+
+class Gemma3TextModel(DecoderOnlyModel):
+    def get_local_cache_positions(self, position_ids, query_position):
+        max_cache_len = self._original_mod.config.sliding_window
+        valid_input_len = 1 if query_position is None else query_position + 1
+        cache_seq_len = torch.clamp(position_ids, max=max_cache_len)[:, :1]  # past seen tokens
+        cache_offset = (
+            torch.clamp(position_ids, max=max_cache_len)[:, :1] + valid_input_len
+        )  # cache offset for next steps
+
+        return cache_seq_len, cache_offset
+
+    def forward(
+        self,
+        input_ids: torch.Tensor = None,
+        inputs_embeds: torch.Tensor = None,
+        attention_mask: torch.Tensor = None,
+        cache_position: torch.Tensor = None,
+        position_ids: torch.Tensor = None,
+        query_position: torch.Tensor = None,
+        past_key_values: Tuple[Tuple[torch.Tensor]] = None,
+        rotary_emb: torch.nn.Module = None,
+        global_block_tables: Optional[torch.Tensor] = None,
+        local_block_tables: Optional[torch.Tensor] = None,
+    ):
+        # retrieve input_ids and inputs_embeds
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError(
+                "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
+            )
+
+        # embed positions
+        if inputs_embeds is None:
+            inputs_embeds = self.get_embedding()(input_ids)
+
+        hidden_states = inputs_embeds
+
+        # Global Position Embeddings
+        cos_global, sin_global = rotary_emb[0](hidden_states, self.max_seq_len)
+        cos_global, sin_global = slice_and_unsqueeze_cos_sin(cos_global, sin_global, position_ids)
+
+        # Local Position Embeddings
+        cos_local, sin_local = rotary_emb[1](hidden_states, self.max_seq_len)
+        cos_local, sin_local = slice_and_unsqueeze_cos_sin(cos_local, sin_local, position_ids)
+
+        # (batch, seq_len) -> (batch,)
+        if self.attn_impl == "flash_attn":
+            seq_positions = cache_position[:, 0]
+            seq_positions = self.convert_sequence_positions_for_flash_attn(
+                seq_positions=seq_positions, max_seq_len=self.max_seq_len
+            )
+        else:
+            seq_positions = cache_position[:, :1]
+
+        sliding_cache_pos = self.get_local_cache_positions(position_ids, query_position)
+
+        for layer in self.layers:
+            if layer.is_sliding:
+                hidden_states = layer(
+                    hidden_states=hidden_states,
+                    attention_mask=attention_mask,
+                    seq_positions=sliding_cache_pos,
+                    past_key_values=past_key_values,
+                    cos=cos_local,
+                    sin=sin_local,
+                    block_tables=local_block_tables,
+                )
+            else:
+                hidden_states = layer(
+                    hidden_states=hidden_states,
+                    attention_mask=attention_mask,
+                    seq_positions=seq_positions,
+                    past_key_values=past_key_values,
+                    cos=cos_global,
+                    sin=sin_global,
+                    block_tables=global_block_tables,
+                )
+
+        hidden_states = self.get_last_layernorm()(hidden_states)
+        return hidden_states
+
+
+class Gemma3DecoderLayer(DecoderOnlyLayer):
+    def __init__(self, layer, self_attn: "DecoderOnlyAttention"):
+        super().__init__(layer, self_attn)
+        self.is_sliding = self._original_mod.is_sliding
+
+    def get_pre_feedforward_layernorm(self) -> Gemma3RMSNorm:
+        return self._original_mod.pre_feedforward_layernorm
+
+    def get_post_feedforward_layernorm(self) -> Gemma3RMSNorm:
+        return self._original_mod.post_feedforward_layernorm
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        seq_positions: Union[torch.LongTensor, Tuple[torch.LongTensor]],
+        past_key_values: Tuple[Tuple[torch.Tensor]],
+        cos: Optional[torch.Tensor] = None,
+        sin: Optional[torch.Tensor] = None,
+        block_tables: Optional[torch.Tensor] = None,
+    ):
+        residual = hidden_states
+        hidden_states = self.get_pre_attention_layernorm()(hidden_states)
+
+        hidden_states = self.self_attn(
+            hidden_states, attention_mask, seq_positions, past_key_values, cos, sin, block_tables
+        )
+        hidden_states = self.get_post_attention_layernorm()(hidden_states)
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.get_pre_feedforward_layernorm()(hidden_states)
+        hidden_states = self._original_mod.mlp(hidden_states)
+        hidden_states = self.get_post_feedforward_layernorm()(hidden_states)
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+
+class Gemma3Attention(DecoderOnlyAttention):
+    def __post_init__(self):
+        self.q_proj = self._original_mod.q_proj
+        self.k_proj = self._original_mod.k_proj
+        self.v_proj = self._original_mod.v_proj
+        self.o_proj = self._original_mod.o_proj
+        self.q_norm = self._original_mod.q_norm
+        self.k_norm = self._original_mod.k_norm
+        self.is_sliding = self._original_mod.is_sliding
+
+    def get_attn_scale(self):
+        return self._original_mod.config.query_pre_attn_scalar**-0.5
+
+    def get_attention(self):
+        if self._original_mod.is_sliding:
+            return SlidingWindowAttentionOp(
+                self.num_heads,
+                self.head_dim,
+                self.num_key_value_heads,
+                self.use_attention_mask,
+                self.use_position_ids,
+            )
+        else:
+            return AttentionOp(
+                self.num_heads, self.head_dim, self.num_key_value_heads, self.use_attention_mask, self.use_position_ids
+            )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        seq_positions: torch.LongTensor,
+        past_key_values: Tuple[Tuple[torch.Tensor]],
+        cos: Optional[torch.Tensor] = None,
+        sin: Optional[torch.Tensor] = None,
+        block_tables: Optional[torch.Tensor] = None,
+    ):
+        batch_size, query_length, _ = hidden_states.size()
+
+        query_states, key_states, value_states = self.projection(hidden_states=hidden_states)
+
+        query_states = query_states.view(batch_size, query_length, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(batch_size, query_length, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(batch_size, query_length, self.num_key_value_heads, self.head_dim).transpose(
+            1, 2
+        )
+
+        query_states = self.q_norm(query_states)
+        key_states = self.k_norm(key_states)
+        query_states, key_states = self.apply_rotary_pos_embed(query_states, key_states, cos, sin)
+
+        batch_size = query_states.shape[0]
+        if batch_size > 1 and "prefill" in self.phase:
+            raise NotImplementedError(f"batch size should be 1 if prefill phase, but got {batch_size}.")
+
+        attn_output = self.attention(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            past_key_state=past_key_values[self.layer_idx][0],
+            past_value_state=past_key_values[self.layer_idx][1],
+            seq_position=seq_positions,
+            scale=self.scale,
+            block_tables=block_tables,
+            block_size=self.kvcache_block_size,
+        )
+
+        attn_outputs = self.o_proj(attn_output)
+        return attn_outputs
+
+
+class Gemma3FlashAttention(DecoderOnlyFlashAttention):
+    def __post_init__(self):
+        self.q_proj = self._original_mod.q_proj
+        self.k_proj = self._original_mod.k_proj
+        self.v_proj = self._original_mod.v_proj
+        self.o_proj = self._original_mod.o_proj
+        self.q_norm = self._original_mod.q_norm
+        self.k_norm = self._original_mod.k_norm
+        self.is_sliding = self._original_mod.is_sliding
+
+    def get_attn_scale(self):
+        return self._original_mod.config.query_pre_attn_scalar**-0.5
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        seq_positions: torch.LongTensor,
+        past_key_values: Tuple[Tuple[torch.Tensor]],
+        cos: Optional[torch.Tensor] = None,
+        sin: Optional[torch.Tensor] = None,
+        block_tables: Optional[torch.Tensor] = None,
+    ):
+        batch_size, query_length, _ = hidden_states.size()
+
+        query_states, key_states, value_states = self.projection(hidden_states=hidden_states)
+
+        query_states = query_states.view(batch_size, query_length, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(batch_size, query_length, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(batch_size, query_length, self.num_key_value_heads, self.head_dim).transpose(
+            1, 2
+        )
+
+        query_states = self.q_norm(query_states)
+        key_states = self.k_norm(key_states)
+        query_states, key_states = self.apply_rotary_pos_embed(query_states, key_states, cos, sin)
+
+        attn_output = self.attention(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            past_key_state=past_key_values[self.layer_idx][0],
+            past_value_state=past_key_values[self.layer_idx][1],
+            seq_position=seq_positions,
+            scale=self.scale,
+            block_tables=block_tables,
+            kvcache_block_size=self.kvcache_block_size,
+        )
+
+        attn_outputs = self.o_proj(attn_output)
+        return attn_outputs