optimum-rbln 0.1.13__py3-none-any.whl → 0.1.15__py3-none-any.whl

optimum/rbln/transformers/models/midm/midm_architecture.py

@@ -21,18 +21,24 @@
 # copied, modified, or distributed without prior written permission
 # from Rebellions Inc.
 
-from typing import Optional, Tuple, Union
+from typing import TYPE_CHECKING, Tuple
 
 import torch
 import torch.nn as nn
-from transformers.modeling_outputs import BaseModelOutputWithPast
 
-from ....transformers.models.decoderonly.decoderonly_architecture import (
-    RotaryEmbedding,
-    rotate_half,
-    slice_and_unsqueeze_cos_sin,
+from ....transformers.models.decoderonly.decoderonly_architecture import rotate_half
+from ..decoderonly.decoderonly_architecture import (
+    DecoderOnlyAttention,
+    DecoderOnlyForCausalLM,
+    DecoderOnlyLayer,
+    DecoderOnlyModel,
+    DecoderOnlyWrapper,
+    apply_rotary_pos_emb_partial,
 )
-from ...cache_utils import RebelDynamicCache_4D
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel as MidmLMHeadModel
 
 
 def apply_rotary_to_tensor(tensor, cos, sin, rot_dim):
@@ -50,253 +56,93 @@ def apply_rotary_pos_emb(q, k, cos, sin):
     return q_embed, k_embed
 
 
-class MidmLMHeadModelWrapper(torch.nn.Module):
-    """A wrapper class for the Midm model with a language modeling head."""
-
-    def __init__(self, model, max_seq_len):
-        super().__init__()
-        self.model = model.transformer
-        self.lm_head = model.lm_head
-        self.config = model.config
-        self.max_seq_len = max_seq_len
-
-        self.config.partial_rotary_factor = model.config.rotary_percentage
-        self.config.head_dim = self.config.n_embd // self.config.n_head
+class MidmLMHeadModelWrapper(DecoderOnlyWrapper):
+    def get_rotary_emb(self, max_seq_len):
         self.config.rope_theta = 10000
-        self.rotary_emb = RotaryEmbedding(config=self.config, max_seq_len_cached=max_seq_len)
-
-    def forward(
-        self,
-        input_ids: torch.Tensor,
-        attention_mask: torch.Tensor,
-        cache_position: torch.LongTensor,
-        batch_position: int,
-        query_idx: int,
-        *past_key_values,
-    ):
-        """Defines the forward pass for the wrapper model."""
-        if input_ids.shape[1] == 1:
-            rbln_batch_position = None
-        else:
-            rbln_batch_position = batch_position
-
-        past_key_values = RebelDynamicCache_4D.from_input_format(
-            cache_position,
-            self.config.num_hidden_layers,
-            *past_key_values,
-        )
-
-        outputs = _MidmModel.forward(
-            self.model,
-            input_ids=input_ids,
-            past_key_values=past_key_values,
-            attention_mask=attention_mask,
-            position_ids=cache_position,
-            rotary_pos_emb=self.rotary_emb,
-            batch_ids=rbln_batch_position,
-        )
-
-        hidden_states = outputs[0]
-        if batch_position >= 0:
-            hidden_states = hidden_states[:, query_idx].unsqueeze(1)
-
-        logits = self.lm_head(hidden_states)
-        output = (logits,) + outputs[1:]
-
-        return output, batch_position + query_idx
-
-
-def layernorm1p(module, input):
-    """Applies Layer Normalization with a slight modification on the weights."""
-    return torch.nn.functional.layer_norm(input, module.normalized_shape, module.weight + 1, module.bias, module.eps)
-
-
-class _MidmAttention:
-    """Custom implementation of the MidmAttention class with specific modifications."""
-
-    def _attn(self, query, key, value, attention_mask=None, head_mask=None):
-        """Computes the attention weights and output."""
-        attn_weights = torch.matmul(query, key.transpose(-1, -2))
-
-        if self.scale_attn_weights:
-            attn_weights = attn_weights / torch.full(
-                [], value.size(-1) ** 0.5, dtype=attn_weights.dtype, device=attn_weights.device
+        self.config.head_dim = self.config.n_embd // self.config.n_head
+        self.config.partial_rotary_factor = self.config.rotary_percentage
+        return super().get_rotary_emb(max_seq_len=max_seq_len)
+
+    def convert_to_rbln_causal_lm(self, causal_lm: "MidmLMHeadModel"):
+        if self.attn_impl != "eager":
+            raise NotImplementedError(f"flash attention ({self.attn_impl}) is not implemented for {self.__class__}")
+        new_layers = []
+        for layer in causal_lm.transformer.h:
+            new_self_attn = MidmAttention(layer.attn)
+            new_layer = MidmLayer(layer, new_self_attn)
+            new_layers.append(new_layer)
+        new_model = MidmModel(causal_lm.transformer, new_layers)
+        new_causal_lm = DecoderOnlyForCausalLM(causal_lm, new_model)
+        return new_causal_lm
+
+
+class MidmModel(DecoderOnlyModel):
+    mask_fmin = -10000.0
+
+    def get_layernorm1p(self, module: nn.LayerNorm):
+        def layernorm1p(input: torch.Tensor):
+            """Applies Layer Normalization with a slight modification on the weights."""
+            return torch.nn.functional.layer_norm(
+                input, module.normalized_shape, module.weight + 1, module.bias, module.eps
             )
 
-        if self.scale_attn_by_inverse_layer_idx or self.scale_qk_by_inverse_layer_idx:
-            attn_weights = attn_weights / float(self.layer_idx + 1)
+        return layernorm1p
 
-        if attention_mask is not None:
-            attn_weights = attn_weights + attention_mask
-
-        if self.scale_qk_by_inverse_layer_idx:
-            attn_weights = attn_weights * float(self.layer_idx + 1)
-
-        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
-        attn_weights = attn_weights.type(value.dtype)
-
-        if head_mask is not None:
-            attn_weights = attn_weights * head_mask
-
-        attn_output = torch.matmul(attn_weights, value)
-        return attn_output, attn_weights
-
-    def forward(
-        self,
-        hidden_states: Optional[Tuple[torch.FloatTensor]],
-        attention_mask: Optional[torch.FloatTensor] = None,
-        past_key_value: Optional[RebelDynamicCache_4D] = None,
-        batch_index: Optional[int] = None,
-        cos: Optional[torch.Tensor] = None,
-        sin: Optional[torch.Tensor] = None,
-    ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]], ...]:
-        """Defines the forward pass for the attention mechanism."""
-        bsz, q_len, _ = hidden_states.size()
-
-        querys, keys, values = self.c_attn(hidden_states).split(self.split_size, dim=2)
-
-        querys = self._split_heads(querys, self.num_heads, self.head_dim).contiguous()
-        keys = self._split_heads(keys, self.num_heads, self.head_dim).contiguous()
-        values = self._split_heads(values, self.num_heads, self.head_dim).contiguous()
-
-        querys, keys = apply_rotary_pos_emb(querys, keys, cos, sin)
-
-        # Decoder
-        if (batch_index is None or batch_index == -1) and bsz > 1:
-            all_key_states = []
-            all_value_states = []
-            all_attn_output = []
+    def get_last_layernorm(self) -> nn.LayerNorm:
+        if self._original_mod.use_layernorm1p:
+            return self.get_layernorm1p(self._original_mod.ln_f)
+        else:
+            return self._original_mod.ln_f
 
-            for b in range(bsz):
-                query = querys[b].unsqueeze(0)
-                attn_mask = attention_mask[b].unsqueeze(0)
-                key = keys[b].unsqueeze(0)
-                value = values[b].unsqueeze(0)
+    def get_embedding(self) -> nn.Embedding:
+        return self._original_mod.wte
 
-                key, value = past_key_value.update(
-                    key,
-                    value,
-                    self.layer_idx,
-                    b,
-                )
+    def get_pos_embedding(self) -> nn.Embedding:
+        return self._original_mod.wpe
 
-                attn_output, _ = _MidmAttention._attn(self, query, key, value, attn_mask)
-                attn_output = self._merge_heads(attn_output, self.num_heads, self.head_dim)
 
-                all_key_states.append(key)
-                all_value_states.append(value)
-                all_attn_output.append(attn_output)
+class MidmLayer(DecoderOnlyLayer):
+    def get_layernorm1p(self, module: nn.LayerNorm):
+        def layernorm1p(input: torch.Tensor):
+            """Applies Layer Normalization with a slight modification on the weights."""
+            return torch.nn.functional.layer_norm(
+                input, module.normalized_shape, module.weight + 1, module.bias, module.eps
+            )
 
-            keys = torch.cat(all_key_states, dim=0)
-            values = torch.cat(all_value_states, dim=0)
-            attn_output = torch.cat(all_attn_output, dim=0)
+        return layernorm1p
 
+    def get_pre_attention_layernorm(self) -> nn.LayerNorm:
+        if self._original_mod.use_layernorm1p:
+            return self.get_layernorm1p(self._original_mod.ln_1)
         else:
-            if batch_index is None or batch_index == -1:
-                batch_index = 0
-
-            keys, values = past_key_value.update(
-                keys,
-                values,
-                self.layer_idx,
-                batch_index,
-                read_first_step=True,
-            )
+            return self._original_mod.ln_1
 
-            attn_output, _ = _MidmAttention._attn(self, querys, keys, values, attention_mask)
-            attn_output = self._merge_heads(attn_output, self.num_heads, self.head_dim)
-
-        attn_output = self.c_proj(attn_output)
-        return attn_output, keys, values
-
-
-class _MidmBlock:
-    """Custom implementation of the MidmBlock class with specific modifications."""
-
-    def forward(
-        self,
-        hidden_states: Optional[Tuple[torch.FloatTensor]],
-        layer_idx: int,
-        attention_mask: Optional[torch.FloatTensor] = None,
-        past_key_value: Optional[RebelDynamicCache_4D] = None,
-        batch_ids: Optional[torch.LongTensor] = None,
-        cos: Optional[torch.Tensor] = None,
-        sin: Optional[torch.Tensor] = None,
-    ) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]:
-        """Defines the forward pass for the block."""
-        residual = hidden_states
-        if self.use_layernorm1p:
-            hidden_states = layernorm1p(self.ln_1, hidden_states)
+    def get_post_attention_layernorm(self) -> nn.LayerNorm:
+        if self._original_mod.use_layernorm1p:
+            return self.get_layernorm1p(self._original_mod.ln_2)
         else:
-            hidden_states = self.ln_1(hidden_states)
-
-        hidden_states, k, v = _MidmAttention.forward(
-            self.attn,
-            hidden_states,
-            attention_mask=attention_mask,
-            past_key_value=past_key_value,
-            cos=cos,
-            sin=sin,
-            batch_index=batch_ids,
+            return self._original_mod.ln_2
+
+
+class MidmAttention(DecoderOnlyAttention):
+    def __post_init__(self):
+        self.c_attn = self._original_mod.c_attn
+        self.o_proj = self._original_mod.c_proj
+        self.split_size = self._original_mod.split_size
+        self.num_key_value_heads = self._original_mod.num_heads
+
+    def projection(self, hidden_states) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        query_states, key_states, value_states = self.c_attn(hidden_states).split(self.split_size, dim=2)
+        return query_states, key_states, value_states
+
+    def rbln_attention(self, *args, **kwargs):
+        return super().rbln_attention(
+            *args,
+            **kwargs,
+            layer_idx=self.layer_idx,
+            scale_attn_weights=self._original_mod.scale_attn_weights,
+            scale_attn_by_inverse_layer_idx=self._original_mod.scale_attn_by_inverse_layer_idx,
         )
-        past_key_value.assign(k, v, layer_idx)
-
-        hidden_states = hidden_states + residual
 
-        residual = hidden_states
-        if self.use_layernorm1p:
-            hidden_states = layernorm1p(self.ln_2, hidden_states)
-        else:
-            hidden_states = self.ln_2(hidden_states)
-
-        feed_forward_hidden_states = self.mlp(hidden_states)
-        hidden_states = residual + feed_forward_hidden_states
-
-        return hidden_states, past_key_value
-
-
-class _MidmModel:
-    """Custom implementation of the MidmModel class with specific modifications."""
-
-    def forward(
-        self,
-        input_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[RebelDynamicCache_4D] = None,
-        attention_mask: Optional[torch.FloatTensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        rotary_pos_emb=None,
-        batch_ids: Optional[torch.LongTensor] = None,
-    ) -> Union[Tuple, BaseModelOutputWithPast]:
-        """Defines the forward pass for the model."""
-        input_shape = input_ids.size()
-
-        attention_mask = (1.0 - attention_mask) * -10000.0
-
-        inputs_embeds = self.wte(input_ids)
-
-        cos, sin = rotary_pos_emb(inputs_embeds, attention_mask.shape[-1])
-        cos, sin = slice_and_unsqueeze_cos_sin(cos, sin, position_ids)
-        hidden_states = inputs_embeds
-
-        for layer_idx, (block, _) in enumerate(zip(self.h, past_key_values)):
-            hidden_states, updated_cache = _MidmBlock.forward(
-                block,
-                hidden_states,
-                layer_idx,
-                attention_mask=attention_mask,
-                past_key_value=past_key_values,
-                batch_ids=batch_ids,
-                cos=cos,
-                sin=sin,
-            )
-
-        hidden_states = layernorm1p(self.ln_f, hidden_states)
-        output_shape = (-1,) + input_shape[1:] + (hidden_states.size(-1),)
-        hidden_states = hidden_states.view(output_shape)
-
-        next_cache = updated_cache.to_legacy_cache()
-
-        return BaseModelOutputWithPast(
-            last_hidden_state=hidden_states,
-            past_key_values=next_cache,
-        )
+    def apply_rotary_pos_embed(self, query_states, key_states, cos, sin):
+        return apply_rotary_pos_emb_partial(query_states, key_states, cos, sin, ndim=cos.shape[-1])

optimum/rbln/transformers/models/midm/modeling_midm.py

@@ -21,12 +21,12 @@
 # copied, modified, or distributed without prior written permission
 # from Rebellions Inc.
 
+
+from transformers import AutoModelForCausalLM
+
 from ....utils import logging
-from ...models.decoderonly import RBLNDecoderOnlyModelForCausalLM
-from .hf_hub_cached.modeling_midm import MidmLMHeadModel
-from .midm_architecture import (
-    MidmLMHeadModelWrapper,
-)
+from ..decoderonly import RBLNDecoderOnlyModelForCausalLM
+from .midm_architecture import MidmLMHeadModelWrapper
 
 
 logger = logging.get_logger(__name__)
@@ -47,7 +47,7 @@ class RBLNMidmLMHeadModel(RBLNDecoderOnlyModelForCausalLM):
     """
 
     _decoder_wrapper_cls = MidmLMHeadModelWrapper
-    _original_cls = MidmLMHeadModel
+    _hf_class = AutoModelForCausalLM
 
     @classmethod
     def from_pretrained(cls, *args, **kwargs):