Upload 6 files

__init__.py

@@ -0,0 +1,31 @@
+from transformers.utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_torch_available,
+)
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    from .modeling_optrecastmlp_llama import (
+        OPTRECASTMLP_llamaModel,
+        OPTRECASTMLP_LlamaForCausalLM,
+    )
+    from .configuration_optrecastmlp_llama import OPTRECASTMLP_llama
+
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM
+
+# Register your models with Auto classes
+AutoConfig.register("optrecastmlp_llama", OPTRECASTMLP_llama)
+AutoModel.register(OPTRECASTMLP_llama, OPTRECASTMLP_llamaModel)
+AutoModelForCausalLM.register(OPTRECASTMLP_llama, OPTRECASTMLP_LlamaForCausalLM)
+
+_import_structure = {
+    "configuration_optrecastmlp_llama": ["OPTRECASTMLP_llama"],
+    "modeling_optrecastmlp_llama": ["OPTRECASTMLP_llamaModel", "OPTRECASTMLP_LlamaForCausalLM"],
+}
+
+__all__ = ["OPTRECASTMLP_llamaModel", "OPTRECASTMLP_LlamaForCausalLM", "OPTRECASTMLP_llama"]

config.json

@@ -1,37 +1,98 @@
 {
-  "architectures": [
-    "OPTRECASTMLP_LlamaForCausalLM"
-  ],
-  "attention_bias": false,
-  "attention_dropout": 0.0,
-  "bos_token_id": 128000,
-  "eos_token_id": 128001,
-  "hidden_act": "silu",
+  "vocab_size": 128256,
+  "max_position_embeddings": 131072,
   "hidden_size": 4096,
-  "initializer_range": 0.02,
   "intermediate_size": 14336,
-  "max_position_embeddings": 131072,
-  "mlp_bias": false,
-  "model_type": "optrecastmlp_llama",
-  "num_attention_heads": 32,
-  "num_cf": 1,
-  "num_groups": 8,
   "num_hidden_layers": 32,
+  "num_attention_heads": 32,
   "num_key_value_heads": 8,
-  "num_templates": 4,
-  "pretraining_tp": 1,
+  "hidden_act": "silu",
+  "initializer_range": 0.02,
   "rms_norm_eps": 1e-05,
+  "pretraining_tp": 1,
+  "use_cache": true,
+  "mlp_bias": false,
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "rope_theta": 500000.0,
   "rope_scaling": {
     "factor": 8.0,
-    "high_freq_factor": 4.0,
     "low_freq_factor": 1.0,
+    "high_freq_factor": 4.0,
     "original_max_position_embeddings": 8192,
     "rope_type": "llama3"
   },
-  "rope_theta": 500000.0,
+  "torch_dtype": null,
+  "num_templates": 4,
+  "num_groups": 8,
+  "num_cf": 1,
+  "return_dict": true,
+  "output_hidden_states": false,
+  "output_attentions": false,
+  "torchscript": false,
+  "use_bfloat16": false,
+  "tf_legacy_loss": false,
+  "pruned_heads": {},
   "tie_word_embeddings": false,
-  "torch_dtype": "float32",
-  "transformers_version": "4.46.3",
-  "use_cache": true,
-  "vocab_size": 128256
-}
+  "chunk_size_feed_forward": 0,
+  "is_encoder_decoder": false,
+  "is_decoder": false,
+  "cross_attention_hidden_size": null,
+  "add_cross_attention": false,
+  "tie_encoder_decoder": false,
+  "max_length": 20,
+  "min_length": 0,
+  "do_sample": false,
+  "early_stopping": false,
+  "num_beams": 1,
+  "num_beam_groups": 1,
+  "diversity_penalty": 0.0,
+  "temperature": 1.0,
+  "top_k": 50,
+  "top_p": 1.0,
+  "typical_p": 1.0,
+  "repetition_penalty": 1.0,
+  "length_penalty": 1.0,
+  "no_repeat_ngram_size": 0,
+  "encoder_no_repeat_ngram_size": 0,
+  "bad_words_ids": null,
+  "num_return_sequences": 1,
+  "output_scores": false,
+  "return_dict_in_generate": false,
+  "forced_bos_token_id": null,
+  "forced_eos_token_id": null,
+  "remove_invalid_values": false,
+  "exponential_decay_length_penalty": null,
+  "suppress_tokens": null,
+  "begin_suppress_tokens": null,
+  "architectures": [
+    "OPTRECASTMLP_LlamaForCausalLM"
+  ],
+  "finetuning_task": null,
+  "id2label": {
+    "0": "LABEL_0",
+    "1": "LABEL_1"
+  },
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "tokenizer_class": null,
+  "prefix": null,
+  "bos_token_id": 128000,
+  "pad_token_id": null,
+  "eos_token_id": 128001,
+  "sep_token_id": null,
+  "decoder_start_token_id": null,
+  "task_specific_params": null,
+  "problem_type": null,
+  "_name_or_path": "",
+  "_attn_implementation_autoset": true,
+  "transformers_version": "4.36.0",
+  "model_type": "optrecastmlp_llama",
+  "auto_map": {
+    "AutoConfig": "configuration_optrecastmlp_llama.OPTRECASTMLP_llama",
+    "AutoModel": "modeling_optrecastmlp_llama.OPTRECASTMLP_llamaModel",
+    "AutoModelForCausalLM": "modeling_optrecastmlp_llama.OPTRECASTMLP_LlamaForCausalLM"
+  }
+}

configuration_optrecastmlp_llama.py

@@ -0,0 +1,77 @@
+from transformers import PretrainedConfig
+
+
+class OPTRECASTMLP_llama(PretrainedConfig):
+    model_type = "optrecastmlp_llama"
+    attribute_map = {
+        "hidden_size": "hidden_size",
+        "num_attention_heads": "num_attention_heads",
+    }
+
+    def __init__(
+        self,
+        vocab_size=128256,
+        hidden_size=4096,
+        intermediate_size=14336,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=8,
+        hidden_act="silu",
+        max_position_embeddings=131072,
+        initializer_range=0.02,
+        rms_norm_eps=1e-5,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=128000,
+        eos_token_id=128001,
+        pretraining_tp=1,
+        tie_word_embeddings=False,
+        rope_theta=500000.0,
+        rope_scaling={
+            "factor": 8.0,
+            "low_freq_factor": 1.0,
+            "high_freq_factor": 4.0,
+            "original_max_position_embeddings": 8192,
+            "rope_type": "llama3",
+        },
+        attention_bias=False,
+        attention_dropout=0.0,
+        mlp_bias=False,
+        # Template-specific configs
+        num_templates=4,
+        num_groups=8,
+        num_cf=1,
+        torch_dtype="bfloat16",
+        **kwargs
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.pretraining_tp = pretraining_tp
+        self.use_cache = use_cache
+        self.mlp_bias = mlp_bias
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.torch_dtype = torch_dtype
+
+        # Template-specific configs
+        self.num_templates = num_templates
+        self.num_groups = num_groups
+        self.num_cf = num_cf
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs
+        )

metadata.json

@@ -0,0 +1 @@
+{"library_name": "transformers", "model_type": "recastmlp_llama", "architectures": ["RECASTMLP_llamaModel"]}

model_card.md

@@ -0,0 +1,60 @@
+---
+language: en
+tags:
+- llama
+- template-mlp
+- parameter-efficient
+- mlp-modification
+datasets:
+- none
+license: apache-2.0
+pipeline_tag: text-generation
+library_name: transformers
+---
+
+# RECASTMLP-LLaMA
+
+This model implements a parameter-efficient modification of the LLaMA architecture by replacing the standard MLP layers with template-based shared MLPs. The model maintains LLaMA's attention mechanism while reducing parameters in the feed-forward networks.
+
+## Model Description
+
+### Overview
+RECASTMLP-LLaMA modifies the original LLaMA architecture by introducing template banks for MLP layers. Instead of having separate MLP weights for each transformer layer, it uses a shared set of template weights that are combined using learned coefficients.
+
+### Architecture Details
+- **Base Model:** LLaMA 3.1 8B
+- **Number of Templates:** 4
+- **Number of Groups:** 8
+- **Coefficients per Template:** 1
+- **Coefficients** 392
+- **Hidden Size:** 4096
+- **Intermediate Size:** 14336
+- **Number of Attention Heads:** 32
+- **Number of Key-Value Heads:** 8
+- **Number of Layers:** 32
+- **Max Position Embeddings:** 131072
+- **Vocabulary Size:** 128256
+
+
+### Key Features
+1. **Template Banks:** Uses shared template weights across groups of layers
+2. **Parameter Efficiency:** Reduces the total number of parameters by sharing MLP weights
+3. **Group-wise Sharing:** Organizes layers into groups that share template banks
+4. **Coefficient Learning:** Uses learned coefficients to combine template weights
+
+## Usage
+
+```python
+from transformers import AutoModel, AutoTokenizer
+
+# Load model and tokenizer
+model = AutoModel.from_pretrained("appledora/RECASTMLP-llama3.1-f8t4", trust_remote_code=True)
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8b")
+
+# Prepare input
+text = "Hello, how are you?"
+inputs = tokenizer(text, return_tensors="pt")
+
+# Generate output
+outputs = model(**inputs)
+hidden_states = outputs.last_hidden_state

modeling_optrecastmlp_llama.py

@@ -0,0 +1,668 @@
+# filename: recastmlp_llama_model.py
+from .configuration_optrecastmlp_llama import OPTRECASTMLP_llama
+from transformers import PreTrainedModel
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Optional, Tuple, Union, List
+from transformers import AutoConfig
+from transformers.utils import logging
+from transformers.cache_utils import Cache, StaticCache
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.generation import GenerationMixin
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+
+logger = logging.get_logger(__name__)
+
+
+class MLPTemplateBank(nn.Module):
+    def __init__(self, config, num_templates):
+        super().__init__()
+        self.num_templates = num_templates
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+
+        # Store templates in a more efficient layout
+        self.up_gate_templates = nn.Parameter(
+            torch.empty(num_templates, 2 * self.intermediate_size * self.hidden_size)
+        )
+        self.down_templates = nn.Parameter(
+            torch.empty(num_templates, self.hidden_size * self.intermediate_size)
+        )
+
+        # Initialize with proper scaling
+        std = 1.0 / (self.hidden_size**0.5)
+        nn.init.normal_(self.up_gate_templates, std=std)
+        nn.init.normal_(self.down_templates, std=std)
+
+    def forward(self, coeffs):
+        # Simple matrix multiplication instead of broadcasting
+        up_gate_weights = torch.mm(coeffs, self.up_gate_templates)
+        down_weights = torch.mm(coeffs, self.down_templates)
+
+        # Reshape to final dimensions
+        up_gate = up_gate_weights.view(2, self.intermediate_size, self.hidden_size)
+        return (
+            up_gate[0],
+            up_gate[1],
+            down_weights.view(self.hidden_size, self.intermediate_size),
+        )
+
+
+class SharedLlamaMLP(nn.Module):
+    def __init__(self, config, bank):
+        super().__init__()
+        self.config = config
+        self.bank = bank
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+
+        # Use transposed coefficients to avoid unnecessary operations
+        self.coefficients = nn.Parameter(torch.zeros(1, config.num_templates))
+        nn.init.normal_(self.coefficients, std=0.02)
+
+        if config.mlp_bias:
+            self.gate_bias = nn.Parameter(torch.zeros(self.intermediate_size))
+            self.up_bias = nn.Parameter(torch.zeros(self.intermediate_size))
+            self.down_bias = nn.Parameter(torch.zeros(self.hidden_size))
+        else:
+            self.register_parameter("gate_bias", None)
+            self.register_parameter("up_bias", None)
+            self.register_parameter("down_bias", None)
+
+        self.act_fn = F.silu
+
+    def forward(self, x):
+        # Generate weights with minimal operations
+        gate_weights, up_weights, down_weights = self.bank(self.coefficients)
+
+        # Standard MLP operations
+        gate_output = F.linear(x, gate_weights, self.gate_bias)
+        up_output = F.linear(x, up_weights, self.up_bias)
+
+        hidden_states = self.act_fn(gate_output) * up_output
+        output = F.linear(hidden_states, down_weights, self.down_bias)
+
+        return output
+
+
+def fixed_cross_entropy(
+    source,
+    target,
+    num_items_in_batch: int = None,
+    ignore_index: int = -100,
+    **kwargs,
+):
+    reduction = "sum" if num_items_in_batch is not None else "mean"
+    loss = nn.functional.cross_entropy(
+        source, target, ignore_index=ignore_index, reduction=reduction
+    )
+    if reduction == "sum":
+        loss = loss / num_items_in_batch
+    return loss
+
+
+from transformers.models.llama.modeling_llama import (
+    LlamaDecoderLayer,
+    LlamaRotaryEmbedding,
+    LlamaRMSNorm,
+    apply_rotary_pos_emb,
+)
+from transformers.modeling_outputs import BaseModelOutputWithPast
+
+
+class OPTRECASTMLP_llamaModel(PreTrainedModel):
+    config_class = OPTRECASTMLP_llama
+    base_model_prefix = "llama"
+    supports_gradient_checkpointing = True
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(
+            config.vocab_size, config.hidden_size, self.padding_idx
+        )
+        # Initialize rotary embeddings
+        rope_config = config.rope_scaling
+        if rope_config:
+            rope_type = rope_config.get("rope_type", "default")
+            scaling_factor = rope_config.get("factor", 1.0)
+        else:
+            rope_type = "default"
+            scaling_factor = None
+        original_config = AutoConfig.from_pretrained(
+            "meta-llama/Llama-3.1-8b", trust_remote_code=True
+        )
+        self.rotary_emb = LlamaRotaryEmbedding(
+            config=original_config,
+        )
+
+        # Create template banks first
+        self.banks = []
+        layers_per_group = config.num_hidden_layers // config.num_groups
+        for _ in range(config.num_groups):
+            bank = MLPTemplateBank(config, config.num_templates)
+            self.banks.append(bank)
+
+        # Create layers using LlamaDecoderLayer but replace MLPs
+        self.layers = nn.ModuleList()
+        for layer_idx in range(config.num_hidden_layers):
+            # Create standard LlamaDecoderLayer
+            decoder_layer = LlamaDecoderLayer(config, layer_idx)
+
+            # Replace its MLP with our SharedLlamaMLP
+            group_idx = layer_idx // layers_per_group
+            group_bank = self.banks[group_idx]
+            decoder_layer.mlp = SharedLlamaMLP(config, bank=group_bank)
+
+            self.layers.append(decoder_layer)
+
+        self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **flash_attn_kwargs,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError(
+                "You must specify exactly one of input_ids or inputs_embeds"
+            )
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+        # Set up cache position if not provided
+        if cache_position is None:
+            past_seen_tokens = (
+                0
+                if past_key_values is None
+                else (
+                    past_key_values.get_seq_length()
+                    if isinstance(past_key_values, Cache)
+                    else past_key_values[0][0].size(-2) if past_key_values else 0
+                )
+            )
+            cache_position = torch.arange(
+                past_seen_tokens,
+                past_seen_tokens + inputs_embeds.shape[1],
+                device=inputs_embeds.device,
+            )
+        # Create position embeddings to be shared across the decoder layers
+        # Set up position IDs if not provided
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+        # Get updated causal mask
+        causal_mask = self._update_causal_mask(
+            attention_mask,
+            inputs_embeds,
+            cache_position,
+            past_key_values,
+            output_attentions,
+        )
+        hidden_states = inputs_embeds
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # Initialize outputs
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = None
+
+        # Process through layers
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    position_embeddings,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    position_embeddings=position_embeddings,
+                    **flash_attn_kwargs,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        # Final layer norm
+        hidden_states = self.norm(hidden_states)
+
+        # Add last hidden state
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns]
+                if v is not None
+            )
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+        if isinstance(
+            pretrained_model_name_or_path, str
+        ) and pretrained_model_name_or_path.endswith(".pt"):
+            print("Loading from local checkpoint")
+            # Load from local checkpoint
+            config = kwargs.get("config", None)
+            if config is None:
+                config = AutoConfig.from_pretrained(
+                    pretrained_model_name_or_path, trust_remote_code=True
+                )
+
+            model = cls(config)
+            checkpoint = torch.load(pretrained_model_name_or_path, map_location="cpu")
+            state_dict = checkpoint["model_state_dict"]
+            logger.info(
+                f"Loaded checkpoint from epoch {checkpoint.get('epoch')} with loss {checkpoint.get('loss')}"
+            )
+
+            missing_keys, unexpected_keys = model.load_state_dict(
+                state_dict, strict=False
+            )
+
+            if len(missing_keys) > 0:
+                logger.warning(f"Missing keys: {missing_keys}")
+            if len(unexpected_keys) > 0:
+                logger.warning(f"Unexpected keys: {unexpected_keys}")
+
+            return model
+        else:
+            print("Loading from hub")
+            # Load from hub using parent's from_pretrained
+            return super().from_pretrained(
+                pretrained_model_name_or_path, *model_args, **kwargs
+            )
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool,
+    ):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = (
+            past_key_values.get_seq_length() if past_key_values is not None else 0
+        )
+        using_static_cache = isinstance(past_key_values, StaticCache)
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if (
+            self.config._attn_implementation == "sdpa"
+            and not using_static_cache
+            and not output_attentions
+        ):
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        sequence_length = input_tensor.shape[1]
+        if using_static_cache:
+            target_length = past_key_values.get_max_cache_shape()
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
+            )
+
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type == "cuda"
+            and not output_attentions
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = AttentionMaskConverter._unmask_unattended(
+                causal_mask, min_dtype
+            )
+
+        return causal_mask
+
+    @staticmethod
+    def _prepare_4d_causal_attention_mask_with_cache_position(
+        attention_mask: torch.Tensor,
+        sequence_length: int,
+        target_length: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        cache_position: torch.Tensor,
+        batch_size: int,
+        **kwargs,
+    ):
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+            causal_mask = attention_mask
+        else:
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = torch.full(
+                (sequence_length, target_length),
+                fill_value=min_dtype,
+                dtype=dtype,
+                device=device,
+            )
+            if sequence_length != 1:
+                causal_mask = torch.triu(causal_mask, diagonal=1)
+            causal_mask *= torch.arange(
+                target_length, device=device
+            ) > cache_position.reshape(-1, 1)
+            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+            if attention_mask is not None:
+                causal_mask = (
+                    causal_mask.clone()
+                )  # copy to contiguous memory for in-place edit
+                mask_length = attention_mask.shape[-1]
+                padding_mask = (
+                    causal_mask[:, :, :, :mask_length]
+                    + attention_mask[:, None, None, :]
+                )
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[
+                    :, :, :, :mask_length
+                ].masked_fill(padding_mask, min_dtype)
+
+        return causal_mask
+
+
+class OPTRECASTMLP_LlamaForCausalLM(PreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    config_class = OPTRECASTMLP_llama
+    base_model_prefix = "llama"
+    supports_gradient_checkpointing = True
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = OPTRECASTMLP_llamaModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def loss_function(
+        self,
+        logits,
+        labels,
+        vocab_size: int,
+        num_items_in_batch: int = None,
+        ignore_index: int = -100,
+        **kwargs,
+    ):
+        # Upcast to float if we need to compute the loss to avoid potential precision issues
+        logits = logits.float()
+        # Shift so that tokens < n predict n
+        shift_logits = logits[..., :-1, :].contiguous()
+        shift_labels = labels[..., 1:].contiguous()
+        # Flatten the tokens
+        shift_logits = shift_logits.view(-1, vocab_size)
+        shift_labels = shift_labels.view(-1)
+        # Enable model parallelism
+        shift_labels = shift_labels.to(shift_logits.device)
+        loss = fixed_cross_entropy(
+            shift_logits, shift_labels, num_items_in_batch, ignore_index, **kwargs
+        )
+        return loss
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        num_logits_to_keep: int = 0,
+        **kwargs,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        """
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should be in
+                `[0, ..., config.vocab_size]` or -100 (masked tokens).
+            num_logits_to_keep (`int`, *optional*):
+                Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate all logits.
+        """
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        hidden_states = outputs[0]
+        # Only compute necessary logits
+        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+
+        loss = None
+        if labels is not None:
+            # Calculate batch size for loss function
+            num_items_in_batch = (
+                input_ids.size(0) if input_ids is not None else inputs_embeds.size(0)
+            )
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                num_items_in_batch=num_items_in_batch,
+                **kwargs,
+            )
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        **kwargs,
+    ):
+        if past_key_values:
+            input_ids = input_ids[:, -1:]
+
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+        if isinstance(
+            pretrained_model_name_or_path, str
+        ) and pretrained_model_name_or_path.endswith(".pt"):
+            print("Loading from local checkpoint")
+            config = kwargs.get("config", None)
+            if config is None:
+                config = AutoConfig.from_pretrained(
+                    pretrained_model_name_or_path, trust_remote_code=True
+                )
+
+            model = cls(config)
+            checkpoint = torch.load(pretrained_model_name_or_path, map_location="cpu")
+            state_dict = checkpoint["model_state_dict"]
+
+            missing_keys, unexpected_keys = model.load_state_dict(
+                state_dict, strict=False
+            )
+
+            if len(missing_keys) > 0:
+                logger.warning(f"Missing keys: {missing_keys}")
+            if len(unexpected_keys) > 0:
+                logger.warning(f"Unexpected keys: {unexpected_keys}")
+
+            return model
+        else:
+            print("Loading from hub")
+            return super().from_pretrained(
+                pretrained_model_name_or_path, *model_args, **kwargs
+            )