Duplicate from Crystalcareai/Quiet-Star-Custom

Co-authored-by: Lucas Atkins <Crystalcareai@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,6 @@
+---
+datasets:
+- open-web-math/open-web-math
+---
+
+Mistral-7b with continued pretraining using Quiet-STaR (https://arxiv.org/abs/2403.09629) for generating 8 thought tokens before each output token.

added_tokens.json

@@ -0,0 +1,4 @@
+{
+  "<|endthought|>": 32000,
+  "<|startthought|>": 32001
+}

config.json

@@ -0,0 +1,41 @@
+{
+  "_name_or_path": "Crystalcareai/Quiet-Star-Custom",
+  "architectures": [
+    "QuietForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "Crystalcareai/Quiet-Star-Custom--configuration_quiet.QuietConfig",
+    "AutoModel": "Crystalcareai/Quiet-Star-Custom--modeling_quiet.QuietModel",
+    "AutoModelForCausalLM": "Crystalcareai/Quiet-Star-Custom--modeling_quiet.QuietForCausalLM"
+  },
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 14336,
+  "max_position_embeddings": 32768,
+  "max_thoughts": 10,
+  "merged_lm_and_talk_heads": false,
+  "merged_lm_and_think_heads": true,
+  "merged_talk_heads": true,
+  "model_type": "quiet",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "num_key_value_heads": 8,
+  "rms_norm_eps": 1e-05,
+  "rope_theta": 10000.0,
+  "sliding_window": 4096,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.37.0.dev0",
+  "use_cache": true,
+  "use_complex_talk_head": true,
+  "use_complex_think_head": false,
+  "use_concat_talk_head": true,
+  "use_shallow_talk": false,
+  "use_shallow_think": true,
+  "use_weighted_talk_head": true,
+  "vocab_size": 32002
+}

configuration_quiet.py

@@ -0,0 +1,172 @@
+# coding=utf-8
+# Copyright 2023 Quiet AI and the HuggingFace Inc. team. 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.
+""" Quiet model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+QUIET_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "quietai/Quiet-7B-v0.1": "https://huggingface.co/quietai/Quiet-7B-v0.1/resolve/main/config.json",
+    "quietai/Quiet-7B-Instruct-v0.1": "https://huggingface.co/quietai/Quiet-7B-Instruct-v0.1/resolve/main/config.json",
+}
+
+
+class QuietConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`QuietModel`]. It is used to instantiate an
+    Quiet model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the Quiet-7B-v0.1 or Quiet-7B-Instruct-v0.1.
+    [quietai/Quiet-7B-v0.1](https://huggingface.co/quietai/Quiet-7B-v0.1)
+    [quietai/Quiet-7B-Instruct-v0.1](https://huggingface.co/quietai/Quiet-7B-Instruct-v0.1)
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+    Args:
+        vocab_size (`int`, *optional*, defaults to 32000):
+            Vocabulary size of the Quiet model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`QuietModel`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 14336):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 8):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `8`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to `4096*32`):
+            The maximum sequence length that this model might ever be used with. Quiet's sliding window attention
+            allows sequence of up to 4096*32 tokens.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        pad_token_id (`int`, *optional*):
+            The id of the padding token.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            The id of the "beginning-of-sequence" token.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            The id of the "end-of-sequence" token.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        sliding_window (`int`, *optional*, defaults to 4096):
+            Sliding window attention window size. If not specified, will default to `4096`.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+    ```python
+    >>> from transformers import QuietModel, QuietConfig
+    >>> # Initializing a Quiet 7B style configuration
+    >>> configuration = QuietConfig()
+    >>> # Initializing a model from the Quiet 7B style configuration
+    >>> model = QuietModel(configuration)
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "quiet"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=32000,
+        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=4096 * 32,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=1,
+        eos_token_id=2,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        complexity_factor = 0.5,
+        sliding_window=4096,
+        attention_dropout=0.0,
+        max_thoughts=16,
+        max_time=None,
+        max_temperature=10,
+        merged_talk_heads=True,
+        merged_lm_and_talk_heads=False,
+        merged_lm_and_think_heads=True,
+        use_concat_talk_head=True,
+        use_shallow_think=True,
+        use_shallow_talk=False,
+        use_complex_think_head=False,
+        use_complex_talk_head=True,
+        use_weighted_talk_head=True,
+        hidden_dropout_prob=0.0, 
+        **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.sliding_window = sliding_window
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_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.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.attention_dropout = attention_dropout
+        self.max_thoughts = max_thoughts
+        self.max_time = max_time
+        self.complexity_factor = complexity_factor
+        self.max_temperature = max_temperature
+        self.merged_talk_heads = merged_talk_heads
+        self.merged_lm_and_talk_heads = merged_lm_and_talk_heads
+        self.merged_lm_and_think_heads = merged_lm_and_think_heads
+        self.use_concat_talk_head = use_concat_talk_head
+        self.use_shallow_think = use_shallow_think
+        self.use_shallow_talk = use_shallow_talk
+        self.use_complex_think_head = use_complex_think_head
+        self.use_complex_talk_head = use_complex_talk_head
+        self.use_weighted_talk_head = use_weighted_talk_head
+        self.hidden_dropout_prob = hidden_dropout_prob
+
+        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,
+        )

generate.py

@@ -0,0 +1,210 @@
+import torch
+from transformers.generation.utils import (
+    GenerationMixin,
+    validate_stopping_criteria,
+    StoppingCriteriaList,
+)
+from transformers import TextStreamer
+
+
+def custom_generate(
+    self,
+    input_ids,
+    attention_mask=None,
+    max_new_tokens=None,
+    min_length=None,
+    do_sample=None,
+    early_stopping=None,
+    num_beams=None,
+    temperature=None,
+    top_k=None,
+    top_p=None,
+    repetition_penalty=None,
+    bad_words_ids=None,
+    bos_token_id=None,
+    pad_token_id=None,
+    eos_token_id=None,
+    streamer=None,
+    length_penalty=None,
+    no_repeat_ngram_size=None,
+    num_return_sequences=None,
+    decoder_start_token_id=None,
+    use_cache=None,
+    num_beam_groups=None,
+    diversity_penalty=None,
+    prefix_allowed_tokens_fn=None,
+    output_attentions=None,
+    output_hidden_states=None,
+    output_scores=None,
+    return_dict_in_generate=None,
+    forced_bos_token_id=None,
+    forced_eos_token_id=None,
+    remove_invalid_values=None,
+    synced_gpus=None,
+    **kwargs,
+):
+    device = input_ids.device
+    with torch.no_grad():
+        finished_generating = torch.zeros(len(input_ids), dtype=torch.bool, device=device)
+
+        if max_new_tokens is None:
+            max_new_tokens = 50  # Default value if not specified
+        for cur_token_idx in range(max_new_tokens):
+            # Sample the next token
+            new_ids = self(
+                input_ids[~finished_generating],
+                attention_mask=attention_mask[~finished_generating] if attention_mask is not None else None,
+                **kwargs
+            )['logits']
+
+            # Mask out the start and end thought tokens so we don't accidentally sample them
+            new_ids[:, :, self.tokenizer.vocab_size:] = -float("inf")
+
+            for list_idx, answer_idx in enumerate((~finished_generating).nonzero(as_tuple=True)[0]):
+                # Find the index of the last token that is not padding
+                base_answer_ids = input_ids[answer_idx]
+                new_answer_ids = new_ids[list_idx]
+                last_token_idx = (base_answer_ids != self.tokenizer.pad_token_id).nonzero(as_tuple=True)[0].max()
+
+                new_ids_sampled = torch.multinomial(
+                    torch.nn.functional.softmax(new_answer_ids[last_token_idx] / temperature, dim=-1), 1)
+
+                # Assign the new id to the last token
+                if last_token_idx + 1 >= len(base_answer_ids):
+                    # Add padding everywhere
+                    new_padding = torch.full((len(input_ids), 1), self.tokenizer.pad_token_id, dtype=torch.long,
+                                            device=device)
+                    input_ids = torch.cat([input_ids, new_padding], dim=-1)
+                    if attention_mask is not None:
+                        attention_mask = torch.cat([attention_mask, torch.zeros_like(new_padding)], dim=-1)
+
+                if attention_mask is not None:
+                    attention_mask[answer_idx, last_token_idx + 1] = 1
+                input_ids[answer_idx, last_token_idx + 1] = new_ids_sampled
+
+                if new_ids_sampled == self.tokenizer.eos_token_id or new_ids_sampled == self.tokenizer.bos_token_id or new_ids_sampled == self.tokenizer.pad_token_id:
+                    finished_generating[answer_idx] = 1
+
+                # Check if the end token is generated
+                if new_ids_sampled == self.tokenizer.convert_tokens_to_ids("</s>"):
+                    finished_generating[answer_idx] = 1
+            
+            if finished_generating.all():
+                break
+
+            if streamer is not None:
+                streamer.put(new_ids_sampled)
+
+        generated_token_ids = input_ids.tolist()
+        return generated_token_ids, attention_mask
+
+
+def generate(
+    self,
+    input_ids,
+    attention_mask=None,
+    max_new_tokens=None,
+    min_length=None,
+    do_sample=None,
+    early_stopping=None,
+    num_beams=None,
+    temperature=1.1,
+    streamer=None,
+    top_k=None,
+    top_p=None,
+    repetition_penalty=None,
+    bad_words_ids=None,
+    bos_token_id=None,
+    pad_token_id=None,
+    eos_token_id=None,
+    length_penalty=None,
+    no_repeat_ngram_size=None,
+    num_return_sequences=None,
+    decoder_start_token_id=None,
+    use_cache=None,
+    num_beam_groups=None,
+    diversity_penalty=None,
+    prefix_allowed_tokens_fn=None,
+    output_attentions=None,
+    output_hidden_states=None,
+    output_scores=None,
+    return_dict_in_generate=None,
+    forced_bos_token_id=None,
+    forced_eos_token_id=None,
+    remove_invalid_values=None,
+    synced_gpus=None,
+    n_ahead=12,
+    n_ahead_talk=4,
+    merged_talk_heads=True,
+    merged_lm_and_talk_heads=False,
+    merged_lm_and_think_heads=True,
+    use_concat_talk_head=True,
+    use_shallow_think=True,
+    use_shallow_talk=False,
+    use_complex_think_head=False,
+    use_complex_talk_head=True,
+    use_weighted_talk_head=True,
+    trust_remote_code=True,
+    torch_dtype=torch.bfloat16,
+    **model_kwargs,
+):
+    # Set model attributes
+    self.max_thoughts = n_ahead + n_ahead_talk + 1
+    self.merged_talk_heads = merged_talk_heads
+    self.merged_lm_and_talk_heads = merged_lm_and_talk_heads
+    self.merged_lm_and_think_heads = merged_lm_and_think_heads
+    self.use_concat_talk_head = use_concat_talk_head
+    self.use_shallow_think = use_shallow_think
+    self.use_shallow_talk = use_shallow_talk
+    self.use_complex_think_head = use_complex_think_head
+    self.use_complex_talk_head = use_complex_talk_head
+    self.use_weighted_talk_head = use_weighted_talk_head
+
+    # Set model properties
+    self.use_end_thought_token = True
+    self.use_start_thought_token = True
+    self.n_ahead = n_ahead
+    self.n_passes = 1
+    self.eval_mode = True
+    self.first_run = False
+    self.rm_initialized = True
+    self.original_mode = False
+
+    generated_token_ids, attention_mask = custom_generate(
+        self,
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        max_new_tokens=max_new_tokens,
+        min_length=min_length,
+        do_sample=do_sample,
+        early_stopping=early_stopping,
+        num_beams=num_beams,
+        temperature=temperature,
+        top_k=top_k,
+        top_p=top_p,
+        repetition_penalty=repetition_penalty,
+        bad_words_ids=bad_words_ids,
+        bos_token_id=bos_token_id,
+        pad_token_id=pad_token_id,
+        eos_token_id=eos_token_id,
+        length_penalty=length_penalty,
+        no_repeat_ngram_size=no_repeat_ngram_size,
+        num_return_sequences=num_return_sequences,
+        decoder_start_token_id=decoder_start_token_id,
+        use_cache=use_cache,
+        num_beam_groups=num_beam_groups,
+        diversity_penalty=diversity_penalty,
+        prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        output_scores=output_scores,
+        return_dict_in_generate=return_dict_in_generate,
+        forced_bos_token_id=forced_bos_token_id,
+        forced_eos_token_id=forced_eos_token_id,
+        remove_invalid_values=remove_invalid_values,
+        synced_gpus=synced_gpus,
+        streamer=streamer,
+        **model_kwargs,
+    )
+
+    return generated_token_ids, attention_mask

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "transformers_version": "4.37.0.dev0"
+}

inference.py

@@ -0,0 +1,121 @@
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer
+
+def compute_memory_used_pct(device):
+    memory_used = torch.cuda.max_memory_allocated(device) / (1024**3)
+    memory_pct = (
+        memory_used
+        / (torch.cuda.get_device_properties(device).total_memory / (1024**3))
+        * 100
+    )
+    return memory_pct
+
+model_path = "./out"
+
+n_ahead = 8
+n_ahead_talk = 4
+merged_talk_heads = True
+
+# Load the model
+model = AutoModelForCausalLM.from_pretrained(
+    model_path,
+    max_thoughts=n_ahead + n_ahead_talk + 1,
+    merged_talk_heads=merged_talk_heads,
+    merged_lm_and_talk_heads=False,
+    merged_lm_and_think_heads=True,
+    use_concat_talk_head=True,
+    use_shallow_think=True,
+    use_shallow_talk=False,
+    use_complex_think_head=False,
+    use_complex_talk_head=True,
+    use_weighted_talk_head=True,
+    trust_remote_code=True,
+    torch_dtype=torch.bfloat16,
+    device_map="auto",
+)
+
+# Load the tokenizer and assign it to the model instance for compatibility
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+model.tokenizer = tokenizer
+
+model.use_end_thought_token = True
+model.use_start_thought_token = True
+model.wandb_enabled = True
+model.n_ahead = n_ahead
+model.n_passes = 2
+model.eval_mode = True
+model.first_run = False
+model.kill_after = 100
+model.rm_initialized = True
+model.original_mode = False
+
+# Custom generate function
+def custom_generate(model, input_ids, attention_mask, max_new_tokens, streamer, **kwargs):
+    with torch.no_grad():
+        finished_generating = torch.zeros(len(input_ids), dtype=torch.bool, device=input_ids.device)
+        for cur_token_idx in range(max_new_tokens):
+            # Sample the next token
+            new_ids = model(
+                input_ids[~finished_generating],
+                attention_mask=attention_mask[~finished_generating]
+            )['logits']
+            # Mask out the start and end thought tokens so we don't accidentally sample them
+            new_ids[:, :, model.tokenizer.vocab_size:] = -float("inf")
+            for list_idx, answer_idx in enumerate((~finished_generating).nonzero(as_tuple=True)[0]):
+                # Find the index of the last token that is not padding
+                base_answer_ids = input_ids[answer_idx]
+                new_answer_ids = new_ids[list_idx]
+                last_token_idx = (base_answer_ids != model.tokenizer.pad_token_id).nonzero(as_tuple=True)[0].max()
+
+                new_ids_sampled = torch.multinomial(
+                    torch.nn.functional.softmax(new_answer_ids[last_token_idx] / kwargs.get("temperature", 1.0), dim=-1), 1)
+                # Assign the new id to the last token
+                if last_token_idx + 1 >= len(base_answer_ids):
+                    # Add padding everywhere
+                    new_padding = torch.full((len(input_ids), 1), model.tokenizer.pad_token_id, dtype=torch.long,
+                                             device=input_ids.device)
+                    input_ids = torch.cat([input_ids, new_padding], dim=-1)
+                    attention_mask = torch.cat([attention_mask, torch.zeros_like(new_padding)], dim=-1)
+                attention_mask[answer_idx, last_token_idx + 1] = 1
+                input_ids[answer_idx, last_token_idx + 1] = new_ids_sampled
+                if new_ids_sampled == model.tokenizer.eos_token_id or new_ids_sampled == model.tokenizer.bos_token_id or new_ids_sampled == model.tokenizer.pad_token_id:
+                    finished_generating[answer_idx] = 1
+                # Check if the end token is generated
+                if new_ids_sampled == model.tokenizer.convert_tokens_to_ids("<|/assistant|>"):
+                    finished_generating[answer_idx] = 1
+            if finished_generating.all():
+                break
+            streamer.put(new_ids_sampled)
+    return input_ids, attention_mask
+
+# Formulate your prompt
+prompt_template = "[INST] {prompt} [/INST]"
+
+prompt = "You're standing on the surface of the Earth. "\
+        "You walk one mile south, one mile west and one mile north. "\
+        "You end up exactly where you started. Where are you?"
+
+# Convert prompt to tokens
+tokens = tokenizer(prompt_template.format(prompt=prompt), return_tensors='pt').input_ids.to(model.device)
+
+# Generate an attention mask
+attention_mask = torch.where(tokens != tokenizer.pad_token_id, torch.ones_like(tokens), torch.zeros_like(tokens)).to(model.device)
+
+streamer = TextStreamer(tokenizer, skip_prompt=False, skip_special_tokens=True)
+
+# Generate output using the custom generate function
+output_ids, _ = custom_generate(
+    model,
+    input_ids=tokens,
+    attention_mask=attention_mask,
+    max_new_tokens=512,
+    streamer=streamer,
+    temperature=0.9,
+)
+
+generated_text = ""
+
+print()  # Print a newline after streaming is complete
+
+# Cleanup if necessary
+torch.cuda.empty_cache()

model-00001-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d0220963e92ec181cd2b99c0421eb03ce9b5dfa938b66df39af1122ebb7ba76
+size 4943211664

model-00002-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:871f3846be90a6c4515931c237d4efb109b37b2f4061670e10c744db0c659461
+size 4999819336

model-00003-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:690edbd65da074f63493d8baf4c9c93a9ee93ad8bec46c8e360595573fa444e6
+size 4641221080

model.safetensors.index.json

@@ -0,0 +1,305 @@
+{
+  "metadata": {
+    "total_size": 14584217600
+  },
+  "weight_map": {
+    "end_embedding": "model-00001-of-00003.safetensors",
+    "lm_head.weight": "model-00003-of-00003.safetensors",
+    "model.embed_tokens.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.10.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.10.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.10.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.11.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.11.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.11.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.11.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.11.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.11.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.11.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.11.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.11.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.12.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.12.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.12.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.12.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.12.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.12.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.12.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.12.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.12.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.13.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.13.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.13.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.13.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.13.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.13.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.13.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.13.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.13.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.2.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.20.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.input_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.22.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.22.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.22.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.22.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.22.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.input_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.23.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.23.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.23.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.23.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.23.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.23.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.23.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.23.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.24.input_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.24.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.24.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.24.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.24.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.24.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.24.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.24.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.24.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.input_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.26.input_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.26.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.26.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.26.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.26.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.26.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.26.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.26.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.26.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.27.input_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.27.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.27.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.27.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.27.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.27.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.27.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.27.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.27.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.28.input_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.28.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.28.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.28.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.28.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.28.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.28.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.28.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.28.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.29.input_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.29.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.29.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.29.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.29.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.29.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.29.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.29.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.29.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.3.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.30.input_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.input_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.4.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.norm.weight": "model-00003-of-00003.safetensors",
+    "start_embedding": "model-00001-of-00003.safetensors",
+    "talk_head.0.0.bias": "model-00003-of-00003.safetensors",
+    "talk_head.0.0.weight": "model-00003-of-00003.safetensors",
+    "talk_head.0.2.bias": "model-00003-of-00003.safetensors",
+    "talk_head.0.2.weight": "model-00003-of-00003.safetensors",
+    "talk_head.0.4.weight": "model-00003-of-00003.safetensors"
+  }
+}

modeling_quiet.py

@@ -0,0 +1,2335 @@
+# coding=utf-8
+# Copyright 2023 Quiet AI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# 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.
+""" PyTorch Quiet model."""
+import inspect
+import math
+import pdb
+import warnings
+from collections import defaultdict
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+from transformers.generation.utils import GenerationMixin
+from transformers.generation.stopping_criteria import StoppingCriteriaList, validate_stopping_criteria
+from transformers import TextStreamer, AutoTokenizer
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import (
+	add_start_docstrings,
+	add_start_docstrings_to_model_forward,
+	is_flash_attn_2_available,
+	is_flash_attn_greater_or_equal_2_10,
+	logging,
+	replace_return_docstrings,
+)
+from .configuration_quiet import QuietConfig
+
+import time
+from typing import Optional, List
+
+
+if is_flash_attn_2_available():
+	from flash_attn import flash_attn_func, flash_attn_varlen_func
+	from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
+	_flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "QuietConfig"
+
+
+def _prepare_4d_causal_attention_mask_for_sdpa(attention_mask, input_shape, inputs_embeds, past_key_values_length):
+	# Compute the attention mask correctly
+	bsz, tgt_len = input_shape
+
+	# Create a 4D attention mask from a 2D tensor mask.
+	# The shape of the output attention mask is (batch_size, 1, tgt_len, src_len)
+	# The values are either 0 or 1, where 0 means padding and 1 means non-padding.
+	combined_attention_mask = None
+	if attention_mask is not None:
+		# What if attention_mask is not None and has a shape of (batch_size, 1, tgt_len, src_len)
+		# In this case, we can just use it directly.
+		if attention_mask.dim() == 4:
+			combined_attention_mask = attention_mask
+		# What if attention_mask is not None and has a shape of (batch_size, 1, tgt_len)
+		# In this case, we need to expand it to (batch_size, 1, tgt_len, src_len)
+		elif attention_mask.dim() == 3:
+			expanded_attn_mask = attention_mask[:, None, :, :]
+			combined_attention_mask = expanded_attn_mask
+		# What if attention_mask is not None and has a shape of (batch_size, tgt_len)
+		# In this case, we need to expand it to (batch_size, 1, tgt_len, src_len)
+		elif attention_mask.dim() == 2:
+			# Provided a padding mask of dimensions [batch_size, seq_length]
+			# - if the model is a decoder, apply a causal mask in addition to the padding mask
+			# - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+			if past_key_values_length > 0:
+				attention_mask = attention_mask.to(dtype=torch.long)
+				attention_mask = attention_mask[:, past_key_values_length:]
+			expanded_attn_mask = attention_mask[:, None, None, :]
+			combined_attention_mask = expanded_attn_mask
+		else:
+			raise ValueError(
+				"Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format(
+					input_shape, attention_mask.shape
+				)
+			)
+
+	# Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+	# masked positions, this operation will create a tensor which is 0.0 for
+	# positions we want to attend and -10000.0 for masked positions.
+	# Since we are adding it to the raw scores before the softmax, this is
+	# effectively the same as removing these entirely.
+	if combined_attention_mask is not None:
+		# Ensure the attention mask values are within a reasonable range
+		combined_attention_mask = combined_attention_mask.clamp(min=0, max=1)
+
+		# Convert the attention mask to bfloat16
+		combined_attention_mask = combined_attention_mask.to(torch.bfloat16)
+
+		# Normalize the attention mask values to be between 0 and 1
+		combined_attention_mask = (1.0 - combined_attention_mask) * -10000.0
+	else:
+		combined_attention_mask = torch.zeros(
+			(bsz, 1, tgt_len, tgt_len), dtype=torch.bfloat16, device=inputs_embeds.device
+		)
+
+	return combined_attention_mask
+
+
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+	seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+	indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+	max_seqlen_in_batch = seqlens_in_batch.max().item()
+	cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
+	return (
+		indices,
+		cu_seqlens,
+		max_seqlen_in_batch,
+	)
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Quiet
+class QuietRMSNorm(nn.Module):
+	def __init__(self, hidden_size, eps=1e-6):
+		super().__init__()
+		self.weight = nn.Parameter(torch.ones(hidden_size))
+		self.variance_epsilon = eps
+
+
+	def forward(self, hidden_states):
+		input_dtype = hidden_states.dtype
+		hidden_states = hidden_states.to(torch.float32)
+		variance = hidden_states.pow(2).mean(-1, keepdim=True)
+		hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+		return hidden_states.to(input_dtype) * self.weight.to(hidden_states.device)
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Quiet
+class QuietRotaryEmbedding(nn.Module):
+	def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
+		super().__init__()
+
+		self.dim = dim
+		self.max_position_embeddings = max_position_embeddings
+		self.base = base
+		inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim))
+		self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+		# Build here to make `torch.jit.trace` work.
+		self._set_cos_sin_cache(
+			seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()
+		)
+
+	def _set_cos_sin_cache(self, seq_len, device, dtype):
+		self.max_seq_len_cached = seq_len
+		t = torch.arange(self.max_seq_len_cached, device=device, dtype=self.inv_freq.dtype)
+
+		freqs = torch.outer(t, self.inv_freq)
+		# Different from paper, but it uses a different permutation in order to obtain the same calculation
+		emb = torch.cat((freqs, freqs), dim=-1)
+		self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
+		self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)
+
+	def forward(self, x, seq_len=None):
+		# x: [bs, num_attention_heads, seq_len, head_size]
+		if seq_len > self.max_seq_len_cached:
+			self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
+
+		return (
+			self.cos_cached[:seq_len].to(dtype=x.dtype),
+			self.sin_cached[:seq_len].to(dtype=x.dtype),
+		)
+
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+	"""Rotates half the hidden dims of the input."""
+	x1 = x[..., : x.shape[-1] // 2]
+	x2 = x[..., x.shape[-1] // 2 :]
+	return torch.cat((-x2, x1), dim=-1)
+
+
+# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
+	"""Applies Rotary Position Embedding to the query and key tensors.
+	Args:
+		q (`torch.Tensor`): The query tensor.
+		k (`torch.Tensor`): The key tensor.
+		cos (`torch.Tensor`): The cosine part of the rotary embedding.
+		sin (`torch.Tensor`): The sine part of the rotary embedding.
+		position_ids (`torch.Tensor`):
+			The position indices of the tokens corresponding to the query and key tensors. For example, this can be
+			used to pass offsetted position ids when working with a KV-cache.
+		unsqueeze_dim (`int`, *optional*, defaults to 1):
+			The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+			sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+			that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+			k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+			cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+			the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+	Returns:
+		`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+	"""
+	cos = cos[position_ids].unsqueeze(unsqueeze_dim)
+	sin = sin[position_ids].unsqueeze(unsqueeze_dim)
+	q_embed = (q * cos) + (rotate_half(q) * sin)
+	k_embed = (k * cos) + (rotate_half(k) * sin)
+	return q_embed, k_embed
+
+
+class QuietMLP(nn.Module):
+	def __init__(self, config):
+		super().__init__()
+		self.config = config
+		self.hidden_size = config.hidden_size
+		self.intermediate_size = config.intermediate_size
+		self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+		self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+		self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+		self.act_fn = ACT2FN[config.hidden_act]
+
+	def forward(self, x):
+		return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+# Copied from transformers.models.llama.modeling_llama.repeat_kv
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+	"""
+	This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+	num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+	"""
+
+	# pdb.set_trace()
+	batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+	if n_rep == 1:
+		return hidden_states
+	hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+	return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+class QuietAttention(nn.Module):
+	"""
+	Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
+	and "Generating Long Sequences with Sparse Transformers".
+	"""
+
+	def __init__(self, config: QuietConfig, layer_idx: Optional[int] = None):
+		super().__init__()
+		self.config = config
+		self.layer_idx = layer_idx
+		if layer_idx is None:
+			logger.warning_once(
+				f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
+				"to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
+				"when creating this class."
+			)
+
+		self.hidden_size = config.hidden_size
+		self.num_heads = config.num_attention_heads
+		self.head_dim = self.hidden_size // self.num_heads
+		self.num_key_value_heads = config.num_key_value_heads
+		self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+		self.max_position_embeddings = config.max_position_embeddings
+		self.rope_theta = config.rope_theta
+		self.is_causal = True
+		self.attention_dropout = config.attention_dropout
+		self._attn_implementation = config._attn_implementation
+
+		if (self.head_dim * self.num_heads) != self.hidden_size:
+			raise ValueError(
+				f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+				f" and `num_heads`: {self.num_heads})."
+			)
+		self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+		self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
+		self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
+		self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+
+		self.rotary_emb = QuietRotaryEmbedding(
+			self.head_dim,
+			max_position_embeddings=self.max_position_embeddings,
+			base=self.rope_theta,
+		)
+
+	def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+		return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+	def forward(
+		self,
+		hidden_states: torch.Tensor,
+		attention_mask: Optional[torch.Tensor] = None,
+		position_ids: Optional[torch.LongTensor] = None,
+		past_key_value: Optional[Cache] = None,
+		output_attentions: bool = False,
+		use_cache: bool = False,
+		**kwargs,
+	) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+		if "padding_mask" in kwargs:
+			warnings.warn(
+				"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+			)
+		bsz, q_len, _ = hidden_states.size()
+
+		query_states = self.q_proj(hidden_states)
+		key_states = self.k_proj(hidden_states)
+		value_states = self.v_proj(hidden_states)
+
+		query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+		key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+		value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+		kv_seq_len = key_states.shape[-2]
+		if past_key_value is not None:
+			if self.layer_idx is None:
+				raise ValueError(
+					f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+					"for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+					"with a layer index."
+				)
+			kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+		cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+		query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+		if past_key_value is not None:
+			cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+			key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+		# repeat k/v heads if n_kv_heads < n_heads
+		key_states = repeat_kv(key_states, self.num_key_value_groups)
+		value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+		attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+		if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+			raise ValueError(
+				f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
+				f" {attn_weights.size()}"
+			)
+		if self._attn_implementation == "flash_attention_2":
+			# Prepare attention mask for flash-attn
+			attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+		elif self._attn_implementation == "sdpa":
+			# Prepare attention mask for SDPA
+			if attention_mask is None or attention_mask.dim() == 2:
+				attention_mask = _prepare_4d_causal_attention_mask(
+					attention_mask,
+					(batch_size, seq_length),
+					inputs_embeds,
+					past_key_values_length,
+					sliding_window=self.config.sliding_window,
+				)
+		else:
+			# Prepare attention mask for other implementations
+			if attention_mask is None or attention_mask.dim() == 2:
+				attention_mask = _prepare_4d_causal_attention_mask(
+					attention_mask,
+					(batch_size, seq_length),
+					inputs_embeds,
+					past_key_values_length,
+					sliding_window=self.config.sliding_window,
+				)
+				
+		if attention_mask is not None:
+			if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+				raise ValueError(
+					f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+				)
+
+			attn_weights = attn_weights + attention_mask
+
+		# upcast attention to fp32
+		attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+		attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+		attn_output = torch.matmul(attn_weights, value_states)
+
+		if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+			raise ValueError(
+				f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+				f" {attn_output.size()}"
+			)
+
+		attn_output = attn_output.transpose(1, 2).contiguous()
+		attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+		attn_output = self.o_proj(attn_output)
+
+		if not output_attentions:
+			attn_weights = None
+
+		return attn_output, attn_weights, past_key_value
+
+
+class QuietFlashAttention2(QuietAttention):
+	"""
+	Quiet flash attention module. This module inherits from `QuietAttention` as the weights of the module stays
+	untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+	flash attention and deal with padding tokens in case the input contains any of them.
+	"""
+
+	# Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+	def __init__(self, *args, **kwargs):
+		super().__init__(*args, **kwargs)
+
+		# TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+		# flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+		# Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+		self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+	def forward(
+		self,
+		hidden_states: torch.Tensor,
+		attention_mask: Optional[torch.Tensor] = None,
+		position_ids: Optional[torch.LongTensor] = None,
+		past_key_value: Optional[Cache] = None,
+		output_attentions: bool = False,
+		use_cache: bool = False,
+		**kwargs,
+	):
+		if "padding_mask" in kwargs:
+			warnings.warn(
+				"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+			)
+
+			# overwrite attention_mask with padding_mask
+			attention_mask = kwargs.pop("padding_mask")
+		bsz, q_len, _ = hidden_states.size()
+
+		query_states = self.q_proj(hidden_states)
+		key_states = self.k_proj(hidden_states)
+		value_states = self.v_proj(hidden_states)
+
+		query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+		key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+		value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+		kv_seq_len = key_states.shape[-2]
+		if past_key_value is not None:
+			if self.layer_idx is None:
+				raise ValueError(
+					f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+					"for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+					"with a layer index."
+				)
+			kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+
+		# Because the input can be padded, the absolute sequence length depends on the max position id.
+		rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1
+		cos, sin = self.rotary_emb(value_states, seq_len=rotary_seq_len)
+
+		query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+		use_sliding_windows = (
+			_flash_supports_window_size
+			and getattr(self.config, "sliding_window", None) is not None
+			and kv_seq_len > self.config.sliding_window
+		)
+
+		if not _flash_supports_window_size:
+			logger.warning_once(
+				"The current flash attention version does not support sliding window attention, for a more memory efficient implementation"
+				" make sure to upgrade flash-attn library."
+			)
+
+		if past_key_value is not None:
+			# Activate slicing cache only if the config has a value `sliding_windows` attribute
+			cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
+			if (
+				getattr(self.config, "sliding_window", None) is not None
+				and kv_seq_len > self.config.sliding_window
+				and cache_has_contents
+			):
+				slicing_tokens = 1 - self.config.sliding_window
+
+				past_key = past_key_value[self.layer_idx][0]
+				past_value = past_key_value[self.layer_idx][1]
+
+				past_key = past_key[:, :, slicing_tokens:, :].contiguous()
+				past_value = past_value[:, :, slicing_tokens:, :].contiguous()
+
+				if past_key.shape[-2] != self.config.sliding_window - 1:
+					raise ValueError(
+						f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
+						f" {past_key.shape}"
+					)
+
+				if attention_mask is not None:
+					attention_mask = attention_mask[:, slicing_tokens:]
+					attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
+
+			cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+			key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+		# repeat k/v heads if n_kv_heads < n_heads
+		key_states = repeat_kv(key_states, self.num_key_value_groups)
+		value_states = repeat_kv(value_states, self.num_key_value_groups)
+		dropout_rate = 0.0 if not self.training else self.attention_dropout
+
+		# In PEFT, usually we cast the layer norms in float32 for training stability reasons
+		# therefore the input hidden states gets silently casted in float32. Hence, we need
+		# cast them back in float16 just to be sure everything works as expected.
+		input_dtype = query_states.dtype
+		if input_dtype == torch.float32:
+			if torch.is_autocast_enabled():
+				target_dtype = torch.get_autocast_gpu_dtype()
+			# Handle the case where the model is quantized
+			elif hasattr(self.config, "_pre_quantization_dtype"):
+				target_dtype = self.config._pre_quantization_dtype
+			else:
+				target_dtype = self.q_proj.weight.dtype
+
+			logger.warning_once(
+				f"The input hidden states seems to be silently casted in float32, this might be related to"
+				f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+				f" {target_dtype}."
+			)
+
+			query_states = query_states.to(target_dtype)
+			key_states = key_states.to(target_dtype)
+			value_states = value_states.to(target_dtype)
+
+		# Reashape to the expected shape for Flash Attention
+		query_states = query_states.transpose(1, 2)
+		key_states = key_states.transpose(1, 2)
+		value_states = value_states.transpose(1, 2)
+
+		attn_output = self._flash_attention_forward(
+			query_states,
+			key_states,
+			value_states,
+			attention_mask,
+			q_len,
+			dropout=dropout_rate,
+			use_sliding_windows=use_sliding_windows,
+		)
+
+		attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+		attn_output = self.o_proj(attn_output)
+
+		if not output_attentions:
+			attn_weights = None
+
+		return attn_output, attn_weights, past_key_value
+
+	def _flash_attention_forward(
+		self,
+		query_states,
+		key_states,
+		value_states,
+		attention_mask,
+		query_length,
+		dropout=0.0,
+		softmax_scale=None,
+		use_sliding_windows=False,
+	):
+		"""
+		Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+		first unpad the input, then computes the attention scores and pad the final attention scores.
+		Args:
+			query_states (`torch.Tensor`):
+				Input query states to be passed to Flash Attention API
+			key_states (`torch.Tensor`):
+				Input key states to be passed to Flash Attention API
+			value_states (`torch.Tensor`):
+				Input value states to be passed to Flash Attention API
+			attention_mask (`torch.Tensor`):
+				The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+				position of padding tokens and 1 for the position of non-padding tokens.
+			dropout (`int`, *optional*):
+				Attention dropout
+			softmax_scale (`float`, *optional*):
+				The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+			use_sliding_windows (`bool`, *optional*):
+				Whether to activate sliding window attention.
+		"""
+		if not self._flash_attn_uses_top_left_mask:
+			causal = self.is_causal
+		else:
+			# TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
+			causal = self.is_causal and query_length != 1
+			
+		# Ensure attention_mask has the correct shape and values
+		if attention_mask is not None:
+			if attention_mask.dim() == 4:
+				# Convert 4D attention mask to 2D
+				attention_mask = attention_mask.squeeze(1).squeeze(1)
+			elif attention_mask.dim() != 2:
+				raise ValueError(
+					f"Invalid attention mask dimension: {attention_mask.dim()}. Expected 2D or 4D mask."
+				)
+			
+			# Ensure attention_mask has values of 0 and 1
+			attention_mask = attention_mask.to(torch.bool).to(torch.int32)
+
+		# Contains at least one padding token in the sequence
+		if attention_mask is not None:
+			batch_size = query_states.shape[0]
+			query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+				query_states, key_states, value_states, attention_mask, query_length
+			)
+
+			cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+			max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+			if not use_sliding_windows:
+				attn_output_unpad = flash_attn_varlen_func(
+					query_states,
+					key_states,
+					value_states,
+					cu_seqlens_q=cu_seqlens_q,
+					cu_seqlens_k=cu_seqlens_k,
+					max_seqlen_q=max_seqlen_in_batch_q,
+					max_seqlen_k=max_seqlen_in_batch_k,
+					dropout_p=dropout,
+					softmax_scale=softmax_scale,
+					causal=causal,
+				)
+			else:
+				attn_output_unpad = flash_attn_varlen_func(
+					query_states,
+					key_states,
+					value_states,
+					cu_seqlens_q=cu_seqlens_q,
+					cu_seqlens_k=cu_seqlens_k,
+					max_seqlen_q=max_seqlen_in_batch_q,
+					max_seqlen_k=max_seqlen_in_batch_k,
+					dropout_p=dropout,
+					softmax_scale=softmax_scale,
+					causal=causal,
+					window_size=(self.config.sliding_window, self.config.sliding_window),
+				)
+
+			attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+		else:
+			if not use_sliding_windows:
+				attn_output = flash_attn_func(
+					query_states,
+					key_states,
+					value_states,
+					dropout,
+					softmax_scale=softmax_scale,
+					causal=causal,
+				)
+			else:
+				attn_output = flash_attn_func(
+					query_states,
+					key_states,
+					value_states,
+					dropout,
+					softmax_scale=softmax_scale,
+					causal=causal,
+					window_size=(self.config.sliding_window, self.config.sliding_window),
+				)
+
+		return attn_output
+
+	def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+		batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape
+
+		# On the first iteration we need to properly re-create the padding mask
+		# by slicing it on the proper place
+		if kv_seq_len != attention_mask.shape[-1]:
+			attention_mask_num_tokens = attention_mask.shape[-1]
+			attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :]
+
+		indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+
+		key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
+		value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
+
+		if query_length == kv_seq_len:
+			query_layer = index_first_axis(
+				query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k
+			)
+			cu_seqlens_q = cu_seqlens_k
+			max_seqlen_in_batch_q = max_seqlen_in_batch_k
+			indices_q = indices_k
+		elif query_length == 1:
+			max_seqlen_in_batch_q = 1
+			cu_seqlens_q = torch.arange(
+				batch_size + 1, dtype=torch.int32, device=query_layer.device
+			)  # There is a memcpy here, that is very bad.
+			indices_q = cu_seqlens_q[:-1]
+			query_layer = query_layer.squeeze(1)
+		else:
+			# The -q_len: slice assumes left padding.
+			attention_mask = attention_mask[:, -query_length:]
+			query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+		return (
+			query_layer,
+			key_layer,
+			value_layer,
+			indices_q,
+			(cu_seqlens_q, cu_seqlens_k),
+			(max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+		)
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Quiet
+class QuietSdpaAttention(QuietAttention):
+	"""
+	Quiet attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+	`QuietAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+	SDPA API.
+	"""
+
+	# Adapted from QuietAttention.forward
+	def forward(
+		self,
+		hidden_states: torch.Tensor,
+		attention_mask: Optional[torch.Tensor] = None,
+		position_ids: Optional[torch.LongTensor] = None,
+		past_key_value: Optional[Cache] = None,
+		output_attentions: bool = False,
+		use_cache: bool = False,
+	) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+		if output_attentions:
+			# TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
+			logger.warning_once(
+				"QuietModel is using QuietSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+				'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+			)
+			return super().forward(
+				hidden_states=hidden_states,
+				attention_mask=attention_mask,
+				position_ids=position_ids,
+				past_key_value=past_key_value,
+				output_attentions=output_attentions,
+				use_cache=use_cache,
+			)
+		bsz, q_len, _ = hidden_states.size()
+
+		query_states = self.q_proj(hidden_states)
+		key_states = self.k_proj(hidden_states)
+		value_states = self.v_proj(hidden_states)
+
+		query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+		key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+		value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+		kv_seq_len = key_states.shape[-2]
+		if past_key_value is not None:
+			kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+		cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+
+		query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+		if past_key_value is not None:
+			cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+			key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+		key_states = repeat_kv(key_states, self.num_key_value_groups)
+		value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+		if attention_mask is not None:
+			if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+				raise ValueError(
+					f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+				)
+			attention_mask = attention_mask.to(query_states.dtype)
+		# SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+		# Reference: https://github.com/pytorch/pytorch/issues/112577.
+		if query_states.device.type == "cuda" and attention_mask is not None:
+			query_states = query_states.contiguous()
+			key_states = key_states.contiguous()
+			value_states = value_states.contiguous()
+
+		attn_output = torch.nn.functional.scaled_dot_product_attention(
+			query_states,
+			key_states,
+			value_states,
+			attn_mask=attention_mask.to(query_states.device) if attention_mask is not None else None,
+			dropout_p=self.attention_dropout if self.training else 0.0,
+			# The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
+			is_causal=self.is_causal and attention_mask is None and q_len > 1,
+		)
+
+		attn_output = attn_output.transpose(1, 2).contiguous()
+		attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+		attn_output = self.o_proj(attn_output)
+
+		return attn_output, None, past_key_value
+
+
+QUIET_ATTENTION_CLASSES = {
+	"eager": QuietAttention,
+	"flash_attention_2": QuietFlashAttention2,
+	"sdpa": QuietSdpaAttention,
+}
+
+
+class QuietDecoderLayer(nn.Module):
+	def __init__(self, config: QuietConfig, layer_idx: int):
+		super().__init__()
+		self.hidden_size = config.hidden_size
+
+		self.self_attn = QUIET_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
+
+		self.mlp = QuietMLP(config)
+		self.input_layernorm = QuietRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+		self.post_attention_layernorm = QuietRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+	def forward(
+		self,
+		hidden_states: torch.Tensor,
+		attention_mask: Optional[torch.Tensor] = None,
+		position_ids: Optional[torch.LongTensor] = None,
+		past_key_value: Optional[Tuple[torch.Tensor]] = None,
+		output_attentions: Optional[bool] = False,
+		use_cache: Optional[bool] = False,
+		**kwargs,
+	) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+		if "padding_mask" in kwargs:
+			warnings.warn(
+				"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+			)
+		"""
+		Args:
+			hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+			attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+				`(batch, sequence_length)` where padding elements are indicated by 0.
+			output_attentions (`bool`, *optional*):
+				Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+				returned tensors for more detail.
+			use_cache (`bool`, *optional*):
+				If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+				(see `past_key_values`).
+			past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+		"""
+
+		residual = hidden_states
+
+		hidden_states = self.input_layernorm(hidden_states)
+
+		# Self Attention
+		hidden_states, self_attn_weights, present_key_value = self.self_attn(
+			hidden_states=hidden_states,
+			attention_mask=attention_mask,
+			position_ids=position_ids,
+			past_key_value=past_key_value,
+			output_attentions=output_attentions,
+			use_cache=use_cache,
+		)
+		hidden_states = residual.to(hidden_states.device) + hidden_states
+
+		# Fully Connected
+		residual = hidden_states
+		hidden_states = self.post_attention_layernorm(hidden_states)
+		hidden_states = self.mlp(hidden_states)
+		hidden_states = residual + hidden_states
+
+		outputs = (hidden_states,)
+
+		if output_attentions:
+			outputs += (self_attn_weights,)
+
+		if use_cache:
+			outputs += (present_key_value,)
+
+		return outputs
+
+
+QUIET_START_DOCSTRING = r"""
+	This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+	library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+	etc.)
+	This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+	Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+	and behavior.
+	Parameters:
+		config ([`QuietConfig`]):
+			Model configuration class with all the parameters of the model. Initializing with a config file does not
+			load the weights associated with the model, only the configuration. Check out the
+			[`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+	"The bare Quiet Model outputting raw hidden-states without any specific head on top.",
+	QUIET_START_DOCSTRING,
+)
+class QuietPreTrainedModel(PreTrainedModel):
+	config_class = QuietConfig
+	base_model_prefix = "model"
+	supports_gradient_checkpointing = True
+	_no_split_modules = ["QuietDecoderLayer"]
+	_skip_keys_device_placement = "past_key_values"
+	_supports_flash_attn_2 = True
+	_supports_sdpa = True
+	_supports_cache_class = True
+
+	def _init_weights(self, module):
+		std = self.config.initializer_range
+		if isinstance(module, nn.Linear):
+			module.weight.data.normal_(mean=0.0, std=std)
+			if module.bias is not None:
+				module.bias.data.zero_()
+		elif isinstance(module, nn.Embedding):
+			module.weight.data.normal_(mean=0.0, std=std)
+			if module.padding_idx is not None:
+				module.weight.data[module.padding_idx].zero_()
+
+
+QUIET_INPUTS_DOCSTRING = r"""
+	Args:
+		input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+			Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+			it.
+			Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+			[`PreTrainedTokenizer.__call__`] for details.
+			[What are input IDs?](../glossary#input-ids)
+		attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+			Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+			- 1 for tokens that are **not masked**,
+			- 0 for tokens that are **masked**.
+			[What are attention masks?](../glossary#attention-mask)
+			Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+			[`PreTrainedTokenizer.__call__`] for details.
+			If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+			`past_key_values`).
+			If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+			and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+			information on the default strategy.
+			- 1 indicates the head is **not masked**,
+			- 0 indicates the head is **masked**.
+		position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+			Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+			config.n_positions - 1]`.
+			[What are position IDs?](../glossary#position-ids)
+		past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
+			Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+			blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+			returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+			Two formats are allowed:
+			- a [`~cache_utils.Cache`] instance;
+			- Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+			shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
+			cache format.
+			The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
+			legacy cache format will be returned.
+			If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+			have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+			of shape `(batch_size, sequence_length)`.
+		inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+			Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+			is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+			model's internal embedding lookup matrix.
+		use_cache (`bool`, *optional*):
+			If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+			`past_key_values`).
+		output_attentions (`bool`, *optional*):
+			Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+			tensors for more detail.
+		output_hidden_states (`bool`, *optional*):
+			Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+			more detail.
+		return_dict (`bool`, *optional*):
+			Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+	"The bare Quiet Model outputting raw hidden-states without any specific head on top.",
+	QUIET_START_DOCSTRING,
+)
+class QuietModel(QuietPreTrainedModel):
+	"""
+	Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`QuietDecoderLayer`]
+	Args:
+		config: QuietConfig
+	"""
+
+	def __init__(self, config: QuietConfig):
+		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)
+		self.layers = nn.ModuleList(
+			[QuietDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+		)
+		self._attn_implementation = config._attn_implementation
+		self.norm = QuietRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+		self.gradient_checkpointing = False
+		# Initialize weights and apply final processing
+		self.post_init()
+
+	def get_input_embeddings(self):
+		return self.embed_tokens
+
+	def set_input_embeddings(self, value):
+		self.embed_tokens = value
+
+	@add_start_docstrings_to_model_forward(QUIET_INPUTS_DOCSTRING)
+	def forward(
+		self,
+		input_ids: torch.LongTensor = None,
+		attention_mask: Optional[torch.Tensor] = None,
+		position_ids: Optional[torch.LongTensor] = None,
+		past_key_values: Optional[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,
+	) -> 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
+
+		# retrieve input_ids and inputs_embeds
+		if input_ids is not None and inputs_embeds is not None:
+			raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+		elif input_ids is not None:
+			batch_size, seq_length = input_ids.shape
+		elif inputs_embeds is not None:
+			batch_size, seq_length, _ = inputs_embeds.shape
+		else:
+			raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+		if self.gradient_checkpointing and self.training:
+			if use_cache:
+				logger.warning_once(
+					"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+				)
+				use_cache = False
+
+		past_key_values_length = 0
+
+		if use_cache:
+			use_legacy_cache = not isinstance(past_key_values, Cache)
+			if use_legacy_cache:
+				past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+			past_key_values_length = past_key_values.get_usable_length(seq_length)
+
+		if position_ids is None:
+			device = input_ids.device if input_ids is not None else inputs_embeds.device
+			position_ids = torch.arange(
+				past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
+			)
+			position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+		else:
+			position_ids = position_ids.view(-1, seq_length).long()
+
+		if inputs_embeds is None:
+			inputs_embeds = self.embed_tokens(input_ids)
+
+		if self._attn_implementation == "flash_attention_2":
+			# 2d mask is passed through the layers
+			attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+		elif self._attn_implementation == "sdpa" and not output_attentions and attention_mask.dim() == 2 and False:
+			# output_attentions=True can not be supported when using SDPA, and we fall back on
+			# the manual implementation that requires a 4D causal mask in all cases.
+			attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
+				attention_mask,
+				(batch_size, seq_length),
+				inputs_embeds,
+				past_key_values_length,
+			)
+		elif attention_mask is None or attention_mask.dim() == 2:
+			# 4d mask is passed through the layers
+			attention_mask = _prepare_4d_causal_attention_mask(
+				attention_mask,
+				(batch_size, seq_length),
+				inputs_embeds,
+				past_key_values_length,
+				sliding_window=self.config.sliding_window,
+			)
+
+		hidden_states = inputs_embeds
+
+		# decoder layers
+		all_hidden_states = () if output_hidden_states else None
+		all_self_attns = () if output_attentions else None
+		next_decoder_cache = None
+
+		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,
+					attention_mask,
+					position_ids,
+					past_key_values,
+					output_attentions,
+					use_cache,
+				)
+			else:
+				layer_outputs = decoder_layer(
+					hidden_states,
+					attention_mask=attention_mask,
+					position_ids=position_ids,
+					past_key_value=past_key_values,
+					output_attentions=output_attentions,
+					use_cache=use_cache,
+				)
+
+			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],)
+
+		hidden_states = self.norm(hidden_states)
+
+		# add hidden states from the last decoder layer
+		if output_hidden_states:
+			all_hidden_states += (hidden_states,)
+
+		next_cache = None
+		if use_cache:
+			next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
+
+		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,
+		)
+
+def nonzero_mean(x, axis=None):
+	if axis is not None:
+		return x.sum(axis) / (x != 0).sum(axis)
+	return x.sum() / (x != 0).sum()
+
+def loss_mean(x):
+	return x.sum() / (x != 0).sum()
+
+class QuietForCausalLM(QuietPreTrainedModel, GenerationMixin):
+	_tied_weights_keys = ["lm_head.weight"]
+
+	def __init__(self, config):
+		super().__init__(config)
+		self.model = QuietModel(config)
+		self.vocab_size = config.vocab_size
+		self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+		# self.router_aux_loss_coef = config.router_aux_loss_coef
+		# self.num_experts = config.num_experts
+		# self.num_experts_per_tok = config.num_experts_per_tok
+		self.max_thoughts = config.max_thoughts
+		self.merged_lm_and_talk_heads = config.merged_lm_and_talk_heads
+		self.use_concat_talk_head = config.use_concat_talk_head
+		self.use_shallow_talk = config.use_shallow_talk
+		self.use_complex_talk_head = config.use_complex_talk_head
+		self.use_weighted_talk_head = config.use_weighted_talk_head
+		# the weighted head will output a single value, so it can't be passed to the lm head
+		assert not (self.use_weighted_talk_head and self.use_shallow_talk)
+
+		self.n_ahead = 1
+		self.n_ahead_talk = 1
+		self.n_passes = 1
+		self.n_tokens_print = 1
+		self.gradient_accumulation_steps = 1
+		self.training_steps = 0
+		self.tokenizer = AutoTokenizer.from_pretrained("Crystalcareai/Quiet-Star-Custom")
+		self.start_token_id = None
+		self.end_token_id = None
+		self.rm_initialized = False
+		self.residual_talk_head = True
+		self.thought_init_std_scale = 1e-2
+
+		self.final_only_mode = False
+		self.first_and_last_mode = True
+		self.first_only = False
+		self.original_loss_weight = 0.5
+
+		self.cumulative_residual = False
+		self.clever_residual = False
+		self.skip_residual = False
+		self.no_residual = True
+
+		self.optimize_lm_head_only_at_start = False
+		self.optimize_model_only_at_start = False
+
+		if self.optimize_model_only_at_start:
+			raise NotImplementedError
+		self.train_only_thinking_embedding = False
+		self.weighted_embeddings = False
+		self.use_start_thought_token = True
+		self.use_end_thought_token = True
+		self.initialize_thought_embedding_to_normal = False
+		self.initial_start_token = "---"
+		self.initial_end_token = "---"
+		self.output_logits_at_the_end = True
+
+		self.wandb_enabled = False
+		self.gumbel_temperature = 0.001
+
+		self.use_policy_loss = True
+		self.include_policy_loss = True
+		self.trice_mode = True
+		self.remove_negative_rewards = True
+		self.use_policy_loss_for_end_thought = True
+		
+		self.base_original_mode = False
+		self.original_mode = False
+
+		self.thought_prefix = "(Let's think step by step"
+		self.tokenized_thought_prefix = None
+		self.log_dict = defaultdict(int)
+		self.eval_log_dict = defaultdict(int)
+		self.loss_mean = loss_mean
+
+		self.start_embedding = nn.Parameter(torch.zeros(2, self.model.config.hidden_size))
+		self.end_embedding = nn.Parameter(torch.zeros(2, self.model.config.hidden_size))
+
+		self.policy_loss_beta = 1e6
+		self.embedding_scale = 1e2
+		self.temperature = nn.Parameter(torch.ones(1))
+		self.max_temperature = config.max_temperature
+		self.reinforce_temperature = 3
+		self.base_loss_beta = 1
+		self.thinking_usefulness_head = nn.Linear(self.model.config.hidden_size, 1)
+		self.thinking_threshold = 0.5
+		self.thinking_usefulness_loss_weight = 1e-2
+
+		# Not used in the paper:
+		self.use_thought_prefix = False
+		self.use_reparam_for_thought_embeddings = False
+		self.use_upper_triangular = False
+		self.subtract_mean_reward = False
+		self.comparison_mode = False
+		self.gumbel_detach = False
+	
+		# For visualization
+		self.eval_mode = False
+
+		num_talk = 1
+		talk_input_dim = config.hidden_size if not self.use_concat_talk_head else config.hidden_size * 2
+		if self.use_weighted_talk_head:
+			talk_output_dim = 1
+		else:
+			talk_output_dim = config.hidden_size if self.use_shallow_talk else config.vocab_size
+
+		if not self.merged_lm_and_talk_heads:
+			if self.use_complex_talk_head:
+				self.talk_head = nn.ModuleList([nn.Sequential(
+					nn.Linear(talk_input_dim, config.hidden_size),
+					nn.ReLU(),
+					nn.Linear(config.hidden_size, config.hidden_size),
+					nn.ReLU(),
+					nn.Linear(config.hidden_size, talk_output_dim, bias=False)
+				)])
+			else:
+				self.talk_head = nn.ModuleList([nn.Sequential(
+					nn.Linear(talk_input_dim, talk_output_dim, bias=False)
+				)])
+				
+		self.apply(self._init_weights)
+
+		# Add dropout regularization
+		self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+		# 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 _init_weights(self, module):
+		if isinstance(module, nn.Linear):
+			nn.init.xavier_uniform_(module.weight)
+			if module.bias is not None:
+				nn.init.constant_(module.bias, 0)
+		elif isinstance(module, nn.Embedding):
+			nn.init.xavier_uniform_(module.weight)
+
+	@torch.no_grad()
+	def infer(
+		self,
+		input_ids: torch.LongTensor,
+		attention_mask: Optional[torch.Tensor] = None,
+		position_ids: Optional[torch.LongTensor] = None,
+		past_key_values: Optional[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,
+	):
+		batch_size, seq_len = input_ids.shape
+
+		# Save the original input_ids and attention_mask for later use
+		original_input_ids = input_ids.clone()
+		original_attention_mask = attention_mask.clone() if attention_mask is not None else None
+
+		# Append the start thought token to the input sequence
+		start_thought_token_id = self.tokenizer.convert_tokens_to_ids("<|startthought|>")
+		input_ids = torch.cat([input_ids, torch.tensor([[start_thought_token_id]] * batch_size).to(input_ids.device)], dim=-1)
+		seq_len += 1
+
+		# Update the attention mask
+		if attention_mask is not None:
+			attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)
+
+		# Generate the continuation
+		continuation_length = self.n_ahead - 2
+		new_key_values = past_key_values
+		
+		# Initialize next_token_id with a default value
+		next_token_id = torch.zeros(batch_size, dtype=torch.long).to(input_ids.device)
+		
+		start_time = time.time()
+		for continuation_idx in range(continuation_length):
+			outputs = self.model(
+				input_ids=input_ids if continuation_idx == 0 else next_token_id.unsqueeze(-1).to(input_ids.device),
+				attention_mask=attention_mask,
+				position_ids=position_ids,
+				past_key_values=new_key_values,
+				inputs_embeds=inputs_embeds,
+				use_cache=True,
+				output_attentions=output_attentions,
+				output_hidden_states=output_hidden_states,
+				return_dict=return_dict,
+			)
+			new_key_values = outputs.past_key_values
+
+			hidden_states = outputs[0]
+
+			logits = self.lm_head(hidden_states)
+			logits = logits[:, -1, :]  # Only consider the last token
+
+			# Apply Gumbel-Softmax to the logits
+			next_token_logits = F.gumbel_softmax(logits, tau=self.gumbel_temperature, hard=True, dim=-1)
+			next_token_id = torch.argmax(next_token_logits, dim=-1)
+
+			# Append the generated token to the input sequence
+			# input_ids = torch.cat([input_ids, next_token_id.unsqueeze(-1).to(input_ids.device)], dim=-1)
+			seq_len += 1
+
+			# Update the attention mask
+			if attention_mask is not None:
+				attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)
+
+		# Append the end thought token to the input sequence
+		end_thought_token_id = self.tokenizer.convert_tokens_to_ids("<|endthought|>")
+		input_ids = torch.cat([input_ids, torch.tensor([[end_thought_token_id]] * batch_size).to(input_ids.device)], dim=-1)
+		seq_len += 1
+
+		# Update the attention mask
+		if attention_mask is not None:
+			attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)
+
+		# Get the hidden states before and after the thought
+		outputs_before = self.model(
+			input_ids=original_input_ids,
+			attention_mask=original_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,
+		)
+		hidden_states_before = outputs_before[0][:, -1:, :]
+
+		# two new tokens: last continuation token and end thought token
+		outputs_after = self.model(
+			input_ids=torch.cat([next_token_id.unsqueeze(-1).to(input_ids.device), torch.tensor([[end_thought_token_id]] * batch_size).to(input_ids.device)], dim=-1),
+			attention_mask=torch.cat([attention_mask[:, -1:], torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1),
+			position_ids=position_ids,
+			past_key_values=new_key_values,
+			inputs_embeds=inputs_embeds,
+			use_cache=use_cache,
+			output_attentions=output_attentions,
+			output_hidden_states=output_hidden_states,
+			return_dict=return_dict,
+		)
+		hidden_states_after = outputs_after[0][:, -1:, :]
+
+		# Apply the talk head to get the mixing weight
+		mixing_weight = self.talk_head[0](torch.cat([hidden_states_before, hidden_states_after], dim=-1))
+
+		# Apply the mixing weight to the hidden states
+		mixed_hidden_states = (1 - mixing_weight) * hidden_states_before + mixing_weight * hidden_states_after
+
+		# Apply the language model head to get the final logits
+		logits = self.lm_head(mixed_hidden_states)
+		return logits
+
+	@torch.no_grad()
+	def generate(
+		self,
+		input_ids: torch.LongTensor,
+		attention_mask: Optional[torch.Tensor] = None,
+		max_new_tokens: Optional[int] = None,
+		temperature: float = 1.1,
+		**kwargs,
+	):
+		if attention_mask is None:
+			# Create a default attention mask if not provided
+			attention_mask = torch.ones_like(input_ids)
+
+		from .generate import generate
+		return generate(self, input_ids=input_ids, attention_mask=attention_mask, max_new_tokens=max_new_tokens, temperature=temperature, **kwargs)
+
+	@add_start_docstrings_to_model_forward(QUIET_INPUTS_DOCSTRING)
+	@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+	def forward(
+		self,
+		input_ids: torch.LongTensor = None,
+		attention_mask: Optional[torch.Tensor] = None,
+		position_ids: Optional[torch.LongTensor] = None,
+		past_key_values: Optional[List[torch.FloatTensor]] = None,
+		inputs_embeds: Optional[torch.FloatTensor] = None,
+		labels: Optional[torch.LongTensor] = None,
+		use_cache: Optional[bool] = None,
+		# output_router_logits: Optional[bool] = None,
+		output_attentions: Optional[bool] = None,
+		output_hidden_states: Optional[bool] = None,
+		return_dict: Optional[bool] = None,
+	) -> Union[Tuple, CausalLMOutputWithPast]:
+		r"""
+		Args:
+			labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+				Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+				config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+				(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+		Returns:
+		Example:
+		```python
+		>>> from transformers import AutoTokenizer, QuietForCausalLM
+		>>> model = QuietForCausalLM.from_pretrained("quietai/Quiet-7B-v0.1")
+		>>> tokenizer = AutoTokenizer.from_pretrained("quietai/Quiet-7B-v0.1")
+		>>> prompt = "Hey, are you conscious? Can you talk to me?"
+		>>> inputs = tokenizer(prompt, return_tensors="pt")
+		>>> # Generate
+		>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+		>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+		"Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+		```"""
+
+		if not self.training:
+			n_ahead_talk_to_restore = self.n_ahead_talk
+			n_passes_to_restore = self.n_passes
+			self.n_ahead_talk = 1
+			self.n_passes = 1
+		
+		# aux_loss = None
+		# output_router_logits = output_router_logits if output_router_logits is not None else self.config.output_router_logits
+		# if output_router_logits:
+		#     router_logits = outputs.router_logits if return_dict else outputs[-1]
+		#     if router_logits is not None:
+		#         aux_loss = load_balancing_loss_func(
+		#             router_logits,
+		#             self.num_experts,
+		#             self.num_experts_per_tok,
+		#             attention_mask,
+		#         )
+		#         if labels is not None:
+		#             loss += self.router_aux_loss_coef * aux_loss.to(loss.device)
+					
+		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
+
+		assert self.cumulative_residual or self.clever_residual or self.skip_residual or self.no_residual
+		assert not (self.skip_residual and self.use_policy_loss)
+
+		if self.tokenized_thought_prefix is None and self.use_thought_prefix:
+			self.tokenized_thought_prefix = self.tokenizer(self.thought_prefix, return_tensors="pt", add_special_tokens=False)["input_ids"]
+
+		def apply_head(head, states, detach=False):
+			if detach:
+				head_weight = head.weight.detach()
+			else:
+				head_weight = head.weight
+			head_weight = head_weight.to(states.device)
+			return (head_weight @ states.transpose(-1, -2)).transpose(-1, -2).contiguous()
+	
+		def idx_if_sequential(head, idx=0):
+			if isinstance(head, nn.Sequential) or isinstance(head, nn.ModuleList):
+				return idx_if_sequential(head[idx], idx=idx)
+			return head
+
+		def none_repeat_interleave(x, n):
+			if x is None:
+				return x
+			return x.repeat_interleave(n, dim=0)
+
+		if self.n_passes > 1:
+			input_ids = none_repeat_interleave(input_ids, self.n_passes)
+			attention_mask = none_repeat_interleave(attention_mask, self.n_passes)
+			position_ids = none_repeat_interleave(position_ids, self.n_passes)
+			inputs_embeds = none_repeat_interleave(inputs_embeds, self.n_passes)
+			labels = none_repeat_interleave(labels, self.n_passes)
+			if past_key_values is not None:
+				past_key_values = [none_repeat_interleave(p, self.n_passes) for p in past_key_values]
+		cur_token_indices = torch.arange(input_ids.shape[1], device=input_ids.device)
+
+		self.tokenizer_has_start_thought_token = True
+		self.tokenizer_has_end_thought_token = True
+		if self.start_token_id is None:
+			self.start_token_id = self.tokenizer.convert_tokens_to_ids("<|startthought|>")
+			if self.start_token_id == 0:
+				self.start_token_id = self.tokenizer.bos_token_id
+				self.tokenizer_has_start_thought_token = False
+			elif self.use_start_thought_token:
+				# base_start_id = self.tokenizer.convert_tokens_to_ids(self.initial_start_token)
+				base_start_id = self.tokenizer.encode(self.initial_start_token, add_special_tokens=False)[0]
+				if self.initialize_thought_embedding_to_normal:
+					self.start_embedding.data = torch.zeros_like(self.start_embedding.data)
+				else:
+					self.start_embedding.data[0] = self.model.embed_tokens.weight.data[base_start_id].clone().detach() / self.embedding_scale
+				self.start_embedding.data[1] = torch.log(self.model.embed_tokens.weight.data.std(dim=0) * self.thought_init_std_scale / self.embedding_scale)
+		if self.end_token_id is None:
+			self.end_token_id = self.tokenizer.convert_tokens_to_ids("<|endthought|>")
+			if self.end_token_id == 0:
+				self.end_token_id = self.tokenizer.eos_token_id
+				self.tokenizer_has_end_thought_token = False
+			elif self.use_end_thought_token:
+				# base_end_id = self.tokenizer.convert_tokens_to_ids(self.initial_end_token)
+				base_end_id = self.tokenizer.encode(self.initial_end_token, add_special_tokens=False)[0]
+				if self.initialize_thought_embedding_to_normal:
+					self.end_embedding.data = torch.zeros_like(self.end_embedding.data)
+				else:
+					self.end_embedding.data[0] = self.model.embed_tokens.weight.data[base_end_id].clone().detach() / self.embedding_scale
+				self.end_embedding.data[1] = torch.log(self.model.embed_tokens.weight.data.std(dim=0) * self.thought_init_std_scale / self.embedding_scale)
+
+		if not self.rm_initialized and (self.n_ahead > 1 or not self.base_original_mode):
+			self.rm_initialized = True                        
+			if not self.use_shallow_talk:
+				head = self.talk_head[0]
+				cur_head = head[-1] if isinstance(head, nn.Sequential) else head
+				talk_input_dim = cur_head.weight.data.shape[1]
+				talk_output_dim = 1 if self.use_weighted_talk_head else self.lm_head.weight.data.shape[0]
+				cur_head.weight.data = torch.zeros(talk_output_dim, talk_input_dim, device=cur_head.weight.device, dtype=cur_head.weight.dtype)
+			else:
+				# convert to identity transform
+				def lambda_transform(cur_head):
+					# pdb.set_trace()
+					if cur_head.weight.data.shape[0] != cur_head.weight.data.shape[1]:
+						return torch.cat([
+						torch.eye(
+							cur_head.weight.data.shape[0],
+							device=cur_head.weight.device,
+							dtype=cur_head.weight.dtype
+						),
+						torch.zeros(
+							cur_head.weight.data.shape[0],
+							cur_head.weight.data.shape[1] - cur_head.weight.data.shape[0],
+							device=cur_head.weight.device,
+							dtype=cur_head.weight.dtype
+						)], dim=1)
+					return torch.eye(
+						cur_head.weight.data.shape[0],
+						device=cur_head.weight.device,
+						dtype=cur_head.weight.dtype
+					)
+				if isinstance(self.talk_head[0], nn.Sequential):
+					for cur_head in self.talk_head[0]:
+						# if it has weights
+						if hasattr(cur_head, "weight"):
+							cur_head.weight.data = lambda_transform(cur_head)
+				else:
+					self.talk_head[-1].weight.data = lambda_transform(self.talk_head[0])
+
+		loss = None
+		prev_rm_tokens = None
+		cur_rm_tokens = None
+		prev_rm_logits = None
+		prev_sample_probs = None
+		did_skip_sampling = None
+		skip_sampling = None
+		sample_probs = None
+		hidden_states = None
+		logits = None
+		talk_kl_penalty = None
+		rm_logits = None
+		residual_logits = None
+		probabilities_2d = None
+		prev_probabilities_2d = None
+		policy_reward = None
+		logits_to_output = None
+		batch_size, seq_len = input_ids.shape
+		base_input_ids = input_ids.clone()
+		loss_list = []
+		dqn_loss_list = []
+		sampled_token_history = []
+		sample_probs_history = []
+		action_loglikelihoods_list = []
+
+		temperature = self.temperature
+
+		if self.use_end_thought_token or self.use_start_thought_token:
+			if not self.use_reparam_for_thought_embeddings:
+				start_embedding = self.start_embedding[0].unsqueeze(0) * self.embedding_scale * temperature
+				end_embedding = self.end_embedding[0].unsqueeze(0) * self.embedding_scale * temperature
+			else:
+				start_embedding = self.start_embedding * self.embedding_scale * temperature
+				end_embedding = self.end_embedding * self.embedding_scale * temperature
+			base_embeddings = self.model.embed_tokens.weight
+			if self.train_only_thinking_embedding:
+				base_embeddings = base_embeddings.detach()
+				
+		# # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+		fwd_iters = 1 if self.original_mode else self.n_ahead + self.n_ahead_talk - 1
+		for ahead_idx in range(fwd_iters):
+			past_key_values_length = 0
+			if past_key_values is not None:
+				use_legacy_cache = not isinstance(past_key_values, Cache)
+				if use_legacy_cache:
+					past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+				past_key_values_length = past_key_values.get_usable_length(seq_len)
+
+			if position_ids is None:
+				device = input_ids.device if input_ids is not None else inputs_embeds.device
+				position_ids = torch.arange(
+					past_key_values_length, seq_len + past_key_values_length, dtype=torch.long, device=device
+				)
+				position_ids = position_ids.unsqueeze(0).view(-1, seq_len)
+			else:
+				position_ids = position_ids.view(-1, seq_len).long()
+
+			if inputs_embeds is None:
+				contains_start = self.use_start_thought_token and (input_ids == self.start_token_id).any()
+				contains_end = self.use_end_thought_token and (input_ids == self.end_token_id).any()
+				contains_thought = contains_start or contains_end
+				if contains_thought:
+					thought_id = self.start_token_id if contains_start else self.end_token_id
+					cur_thought_embedding = start_embedding if contains_start else end_embedding
+					if self.use_reparam_for_thought_embeddings:
+						inputs_embeds = torch.randn(batch_size, seq_len, self.model.config.hidden_size, device=input_ids.device, dtype=cur_thought_embedding.dtype)
+						inputs_embeds = inputs_embeds.detach() * torch.exp(cur_thought_embedding[1]) + cur_thought_embedding[0]
+						if contains_start:
+							sampled_start = inputs_embeds.clone().detach()
+						if contains_end:
+							sampled_end = inputs_embeds.clone().detach()
+					else:
+						inputs_embeds = cur_thought_embedding.unsqueeze(0).repeat(batch_size, seq_len, 1)
+				else:
+					with torch.set_grad_enabled(not self.train_only_thinking_embedding):
+						inputs_embeds = self.model.embed_tokens(input_ids)
+			
+			if self.n_ahead != 1 or self.n_ahead_talk != 1 or self.comparison_mode:
+				if attention_mask is None:
+					base_attention_mask = torch.triu(torch.ones(seq_len, seq_len), diagonal=0).to(input_ids.device)
+					base_attention_mask = base_attention_mask.view(1, 1, seq_len, seq_len)
+					base_attention_mask = base_attention_mask.repeat(input_ids.shape[0], 1, 1, 1)
+					attention_mask = base_attention_mask
+					# breakpoint()
+				elif attention_mask.dim() == 2:
+					if seq_len + past_key_values_length != attention_mask.shape[-1]:
+						# breakpoint()
+						attention_mask = torch.cat(
+							[torch.ones((attention_mask.shape[0], past_key_values_length), dtype=attention_mask.dtype, device=attention_mask.device), attention_mask],
+							dim=-1
+						)
+					# # if the attention mask 
+					attention_mask = _prepare_4d_causal_attention_mask(
+						attention_mask,
+						(batch_size, seq_len),
+						inputs_embeds,
+						past_key_values_length,
+						sliding_window=self.config.sliding_window,
+					)
+
+			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,
+				# output_router_logits=output_router_logits,
+				return_dict=return_dict,
+			)
+
+			prev_hidden_states = hidden_states
+			hidden_states = outputs[0]
+			prev_rm_logits = rm_logits  # for policy gradient
+			prev_rm_tokens = cur_rm_tokens  # for policy gradient
+
+			if ahead_idx == 0:
+				hidden_states_lm = hidden_states
+				logits = self.lm_head(hidden_states_lm)
+				base_hidden_states = hidden_states.clone()
+				initial_loss_logits = logits.clone()
+				if self.optimize_lm_head_only_at_start or self.optimize_model_only_at_start:
+					logits = logits.detach()
+					base_hidden_states = base_hidden_states.detach()
+				if self.optimize_model_only_at_start:
+					hidden_states = hidden_states.detach()
+				base_logits = logits.clone()
+			else:
+				talk_hidden_states = hidden_states
+				if self.merged_lm_and_talk_heads:
+					assert self.no_residual
+					residual_logits = self.lm_head(hidden_states)
+					talk_hidden_states = hidden_states
+				else:
+					if ahead_idx > self.n_ahead - 1:
+						cur_base_hidden = torch.cat([
+							base_hidden_states[..., ahead_idx - self.n_ahead + 1:, :],
+							base_hidden_states[..., :ahead_idx - self.n_ahead + 1, :]
+						], dim=-2)
+					else:
+						cur_base_hidden = base_hidden_states
+
+					if self.use_concat_talk_head:
+						# concatenate the hidden states with the original hidden states
+						head_input_hidden_states = torch.cat([cur_base_hidden, talk_hidden_states], dim=-1)
+					else:
+						head_input_hidden_states = talk_hidden_states
+
+					residual_logits = self.talk_head[0](head_input_hidden_states)
+					if self.use_shallow_talk:
+						residual_logits = apply_head(self.lm_head, residual_logits, detach=self.optimize_lm_head_only_at_start)                        
+					residual_logits = residual_logits.to(logits.device)
+					if self.use_weighted_talk_head:
+						# combine the cur_base_hidden with the talk_hidden_states according to the weighted head
+						residual_logits = cur_base_hidden * (1 - residual_logits) + talk_hidden_states * residual_logits
+						residual_logits = apply_head(self.lm_head, residual_logits, detach=self.optimize_lm_head_only_at_start)
+
+				assert sum([self.cumulative_residual, self.clever_residual, self.skip_residual, self.no_residual]) == 1
+				if self.clever_residual:
+					if ahead_idx >= self.n_ahead - 1:
+						# get the logits shifted according to the current talk ahead
+						cur_base_logits = torch.cat([
+							base_logits[..., ahead_idx - self.n_ahead + 1:, :],
+							base_logits[..., :ahead_idx - self.n_ahead + 1, :]
+						], dim=-2)
+						if self.optimize_lm_head_only_at_start:
+							cur_base_logits = cur_base_logits.detach()
+						logits = cur_base_logits + residual_logits
+					else:
+						logits += residual_logits / self.n_ahead
+				elif self.cumulative_residual:
+					if self.residual_talk_head:
+						if ahead_idx < self.n_ahead:
+							logits += residual_logits
+						else:
+							# get the logits shifted according to the current talk ahead
+							cur_base_logits = torch.cat([
+								base_logits[..., ahead_idx - self.n_ahead + 1:, :],
+								base_logits[..., :ahead_idx - self.n_ahead + 1, :]
+							], dim=-2)
+							if self.optimize_lm_head_only_at_start:
+								cur_base_logits = cur_base_logits.detach()
+							logits = cur_base_logits + residual_logits
+					else:
+						if ahead_idx < self.n_ahead:
+							logits += residual_logits
+						else:
+							logits = residual_logits
+				elif self.skip_residual:
+					if ahead_idx >= self.n_ahead:
+						# get the logits shifted according to the current talk ahead
+						cur_base_logits = torch.cat([
+							base_logits[..., ahead_idx - self.n_ahead + 1:, :],
+							base_logits[..., :ahead_idx - self.n_ahead + 1, :]
+						], dim=-2)
+						if self.optimize_lm_head_only_at_start:
+							cur_base_logits = cur_base_logits.detach()
+						logits = cur_base_logits
+				elif self.no_residual:
+					logits = residual_logits
+				else:
+					logits = base_logits + residual_logits
+
+			attempted = False
+			talk_loss_list = []
+			if self.original_mode or (self.n_ahead == 1) or (self.comparison_mode and ahead_idx == 0):# or (self.optimize_lm_head_only_at_start and ahead_idx == 0):
+				loss = None
+				attempted = True
+
+				if labels is not None:
+					for shift_amount in range(self.n_ahead_talk):
+						# Shift so that tokens < n predict n
+						#  ab[cde]f
+						# abc[def]
+						if ahead_idx == 0 and self.optimize_lm_head_only_at_start:
+							loss_logits = initial_loss_logits
+						else:
+							loss_logits = logits
+						shift_logits = loss_logits[..., shift_amount:-1, :].contiguous()
+						shift_labels = labels[..., 1 + shift_amount:].contiguous()
+						# Flatten the tokens
+						loss_fct = CrossEntropyLoss(reduction="none")
+						# print("Shift logits before:", shift_logits)
+						shift_logits = shift_logits.view(-1, self.config.vocab_size)
+						shift_labels = shift_labels.view(-1).clone()
+						# print("shift logits after:", shift_logits)
+						# Enable model parallelism
+						shift_labels[shift_labels == self.tokenizer.pad_token_id] = -100
+						shift_labels = shift_labels.to(shift_logits.device)
+						loss = loss_fct(shift_logits, shift_labels)
+						if not self.comparison_mode and not (self.optimize_lm_head_only_at_start and (self.n_ahead + self.n_ahead_talk > 2)) or self.original_mode:
+							loss_list.append(loss)
+						talk_loss_list.append(nonzero_mean(loss).detach())
+			
+			if not attempted or self.comparison_mode:
+				rm_hidden_states = hidden_states
+				# print("Magnitude of RM hidden states before RM head", rm_hidden_states.norm())
+				rm_logits = apply_head(self.lm_head, rm_hidden_states, detach=self.optimize_lm_head_only_at_start)
+					
+				# don't allow it to predict the thinking token
+				if self.tokenizer_has_start_thought_token:                    
+					rm_logits[..., self.start_token_id] = -1e10
+				if self.tokenizer_has_end_thought_token:
+					rm_logits[..., self.end_token_id] = -1e10
+				probabilities = rm_logits
+				if probabilities_2d is not None:
+					prev_probabilities_2d = probabilities_2d.clone()
+				probabilities_2d = probabilities.view(-1, probabilities.size(-1))
+
+				did_skip_sampling = skip_sampling
+				skip_sampling = False
+				if ahead_idx == 0 and self.use_start_thought_token:
+					override_token = self.start_token_id
+				elif self.use_thought_prefix and ahead_idx < self.tokenized_thought_prefix.shape[-1]:
+					override_token = self.tokenized_thought_prefix[..., ahead_idx]
+				elif ahead_idx == self.n_ahead - 2 and self.use_end_thought_token:
+					override_token = self.end_token_id
+				else:
+					override_token = None
+				if override_token is not None and self.n_ahead > 1:
+					# always start with the start token
+					probabilities_2d = torch.zeros_like(probabilities_2d)
+					probabilities_2d[:, override_token] = 1.0
+					skip_sampling = True
+				elif ahead_idx >= self.n_ahead - 1:
+					if labels is not None:  # we're in the talk phase
+						cur_talk_n = ahead_idx - (self.n_ahead - 1) + 1
+						# print("Setting rm to labels", cur_talk_n, "during", ahead_idx)
+						shift_labels = labels[..., cur_talk_n:].contiguous().to(probabilities_2d.device)
+						padding = torch.full_like(
+							labels[..., :cur_talk_n],
+							self.tokenizer.pad_token_id,
+							dtype=torch.long,
+							device=shift_labels.device
+						)
+						new_rm_tokens = torch.cat(
+							[shift_labels, padding],
+							dim=-1
+						)
+						
+						# print((new_rm_tokens > self.vocab_size - 1).any().item())
+						new_rm_tokens = torch.clamp(new_rm_tokens, 0, self.vocab_size - 1)
+
+						# Now safely convert rm tokens to one-hot
+						probabilities_2d = F.one_hot(new_rm_tokens, num_classes=self.vocab_size).reshape(-1, self.vocab_size).to(probabilities_2d.dtype)
+					else:
+						continue
+				temperature = self.gumbel_temperature if self.training else 0.001
+				prev_sample_probs = sample_probs
+				sample_probs = probabilities_2d
+				if ahead_idx < self.n_ahead - 1 and not skip_sampling:
+					probabilities_2d = F.gumbel_softmax(sample_probs, tau=temperature, hard=True, dim=-1)
+					if self.gumbel_detach:
+						probabilities_2d = probabilities_2d.detach()
+				sampled_token_history.append(probabilities_2d.argmax(dim=-1).detach().cpu())
+				# convert rm logits directly to embeddings
+				contains_start = self.use_start_thought_token and (probabilities_2d[..., self.start_token_id].sum() > 0)
+				contains_end = self.use_end_thought_token and (probabilities_2d[..., self.end_token_id].sum() > 0)
+				contains_thought = contains_start or contains_end
+
+
+				if not contains_thought:
+					with torch.set_grad_enabled(not self.train_only_thinking_embedding):
+						inputs_embeds = probabilities_2d @ (self.model.embed_tokens.weight.to(probabilities.device).to(probabilities.dtype) * temperature)
+				else:
+					thought_id = self.start_token_id if contains_start else self.end_token_id
+					cur_thought_embedding = start_embedding if contains_start else end_embedding
+					if self.use_reparam_for_thought_embeddings:
+						inputs_embeds = torch.randn(batch_size, seq_len, self.model.config.hidden_size, device=input_ids.device, dtype=cur_thought_embedding.dtype)
+						inputs_embeds = inputs_embeds * torch.exp(cur_thought_embedding[1]) + cur_thought_embedding[0]
+						if contains_start:
+							sampled_start = inputs_embeds.clone().detach()
+						else:
+							sampled_end = inputs_embeds.clone().detach()
+					else:
+						inputs_embeds = cur_thought_embedding.unsqueeze(0).repeat(batch_size, seq_len, 1)
+				inputs_embeds = inputs_embeds.view(probabilities.size(0), probabilities.size(1), -1).to(self.model.embed_tokens.weight.dtype)
+				inputs_embeds = inputs_embeds.view(probabilities.size(0), probabilities.size(1), -1).to(self.model.embed_tokens.weight.dtype)
+				
+				# Predict the usefulness of thinking at each token position
+				thinking_usefulness = self.thinking_usefulness_head(hidden_states).squeeze(-1)
+
+				# Apply a threshold to decide where to generate thoughts
+				generate_thought_mask = thinking_usefulness > self.thinking_threshold
+
+				# Compute the regularization loss for thinking usefulness prediction
+				thinking_usefulness_loss = torch.mean(thinking_usefulness * (1 - generate_thought_mask.float()))
+
+				# Add the regularization loss to the total loss
+				if loss is not None:
+					loss = loss + self.thinking_usefulness_loss_weight * thinking_usefulness_loss
+				else:
+					loss = self.thinking_usefulness_loss_weight * thinking_usefulness_loss
+
+
+				if len(attention_mask.shape) == 2:
+					breakpoint()
+				else:
+					original_attention = attention_mask[..., :attention_mask.shape[-2]]
+					if self.use_upper_triangular:
+						new_attention = original_attention
+					else:
+						original_attention = original_attention == attention_mask.max()
+						# because eye isn't implemented for BF16, we need to handle the case
+						if not attention_mask.dtype == torch.bfloat16:
+							new_attention = torch.eye(
+								seq_len, dtype=attention_mask.dtype, device=attention_mask.device
+							)
+						else:
+							new_attention = torch.eye(
+								seq_len, dtype=torch.float32, device=attention_mask.device
+							).to(attention_mask.dtype)
+
+						new_attention = new_attention.view(1, 1, seq_len, seq_len).repeat(input_ids.shape[0], 1, 1, 1)
+						new_attention = new_attention * original_attention
+						new_attention[new_attention == 0] = attention_mask.min()
+						new_attention[new_attention == 1] = attention_mask.max()
+					attention_mask = torch.cat([attention_mask, new_attention], dim=-1)
+				past_key_values = outputs.past_key_values
+				position_ids = position_ids + 1
+
+				if labels is not None and (self.n_ahead > 1 or not self.base_original_mode):
+					# Shift so that tokens < n predict n
+					# logits: abcdef -> bcdef? -> cdef??
+					# labels: abcdef -> ?bcdef -> ??cdef
+					if ahead_idx == 0 and self.optimize_lm_head_only_at_start:
+						loss_logits = initial_loss_logits
+					else:
+						loss_logits = logits
+					shift_idx = 1 + max(0, ahead_idx - (self.n_ahead - 1))
+					shift_logits = loss_logits[..., :-shift_idx, :].contiguous()
+					shift_labels = labels[..., shift_idx:].contiguous()
+					# Flatten the tokens
+					loss_fct = CrossEntropyLoss(reduction="none")
+					shift_logits = shift_logits.view(-1, self.config.vocab_size)
+					shift_labels = shift_labels.view(-1)
+					# Enable model parallelism
+					shift_labels = shift_labels.to(shift_logits.device)
+					# if shift_labels.min() == self.tokenizer.pad_token_id:
+					shift_labels = torch.where(shift_labels == self.tokenizer.pad_token_id, -100, shift_labels)
+					unreduced_loss = loss_fct(shift_logits, shift_labels)
+					# print("Loss:", unreduced_loss.item())  # Print the loss before checking for NaN values
+					if torch.any(unreduced_loss != unreduced_loss):
+						# pdb.set_trace()
+						raise ValueError("NaN loss")
+					unreduced_loss = unreduced_loss.reshape(logits.shape[0], -1)
+					loss_list.append(unreduced_loss)
+
+
+					if self.use_policy_loss and ahead_idx > 0 and (ahead_idx > 1 or not self.use_start_thought_token):
+						# we treat the change in loss as the reward
+						previous_loss = loss_list[-2]
+						# for example, suppose n_ahead = 3 and n_ahead_talk = 2
+						# note that we end at self.n_ahead + self.n_ahead_talk - 2
+						# in this case, 5 - 2 = 3, so we end at ahead_idx = 3
+						# we also predict the next token at ahead_idx = 2
+						# when we get to ahead_idx = 2, we predict ahead
+						# so we shift by 1
+						# note that this is ahead_idx = n_ahead - 1
+						# when we get to ahead_idx = 3, we predict ahead
+						# so we shift by 2
+						# note that this is ahead_idx = n_ahead
+						if ahead_idx < self.n_ahead - 1:
+							shift_amount = 0
+							reward_scale = 1.0
+							original_dqn_reward = torch.sign(previous_loss - unreduced_loss).detach() * reward_scale
+							if self.first_and_last_mode:
+								original_dqn_reward = original_dqn_reward * 0.0
+						else:
+							# logits vs cur_policy_shift_logits
+							# let's look at rm_logits and prev_rm_logits
+							shift_amount = max(0, ahead_idx - (self.n_ahead - 1))
+							# let's say shift_amount = 2
+							# abcdefg -> bcdefg? -> cdefg??
+							# logits = [a b]c d e f[g]
+							# labels = [a b c]d e f g
+							cur_policy_shift_logits = initial_loss_logits[..., shift_amount:-1, :].contiguous().detach()
+							cur_policy_shift_labels = labels[..., 1 + shift_amount:].contiguous()
+							# Flatten the tokens
+							cur_policy_loss_fct = CrossEntropyLoss(reduction="none")
+							cur_policy_shift_logits = cur_policy_shift_logits.view(-1, self.config.vocab_size)
+							cur_policy_shift_labels = cur_policy_shift_labels.view(-1).clone()
+							# Enable model parallelism
+							cur_policy_shift_labels[cur_policy_shift_labels == self.tokenizer.pad_token_id] = -100
+							cur_policy_shift_labels = cur_policy_shift_labels.to(cur_policy_shift_labels.device)
+							cur_policy_reward_base_loss = loss_fct(
+								cur_policy_shift_logits, cur_policy_shift_labels.to(cur_policy_shift_logits.device)
+							).reshape(logits.shape[0], -1)
+							original_dqn_reward = cur_policy_reward_base_loss.detach() - unreduced_loss
+								
+						if not did_skip_sampling:
+							nonzero_indices = prev_probabilities_2d.nonzero()
+							action_loglikelihoods = F.log_softmax(prev_sample_probs / self.reinforce_temperature, dim=-1)[nonzero_indices[:, 0], nonzero_indices[:, 1]]
+							action_loglikelihoods_2d = action_loglikelihoods.reshape(batch_size, -1)[:, :-1 - shift_amount]
+							action_loglikelihoods_list.append(action_loglikelihoods_2d)
+						if policy_reward is None:
+							policy_reward = original_dqn_reward[:, :-(self.n_ahead_talk - shift_amount)]
+						else:
+							if self.n_ahead_talk > shift_amount:
+								added_reward = original_dqn_reward[:, :-(self.n_ahead_talk - shift_amount)]
+							else:
+								added_reward = original_dqn_reward
+							policy_reward += added_reward
+
+						for action_loglikelihoods_2d in action_loglikelihoods_list:
+							train_policy_reward = policy_reward
+
+							# discard rewards below the mean
+							if self.trice_mode and self.n_passes > 1:
+								batched_policy_reward = train_policy_reward.reshape(-1, self.n_passes, train_policy_reward.shape[-1])
+								# average over the passes
+								train_policy_reward = batched_policy_reward - batched_policy_reward.mean(dim=1, keepdim=True)
+								train_policy_reward = train_policy_reward.reshape(-1, train_policy_reward.shape[-1])
+								
+							if self.subtract_mean_reward:
+								train_policy_reward = train_policy_reward - train_policy_reward.mean()
+							if self.remove_negative_rewards:
+								fixed_policy_reward = train_policy_reward.detach().clamp(min=0)
+							else:
+								fixed_policy_reward = train_policy_reward.detach()
+
+							# Normalize rewards
+							fixed_policy_reward = (fixed_policy_reward - fixed_policy_reward.mean()) / (fixed_policy_reward.std() + 1e-8)
+							actor_loss = -fixed_policy_reward * action_loglikelihoods_2d[:, :policy_reward.shape[-1]].to(policy_reward.device)
+							if action_loglikelihoods_2d.mean() < -1e4 and not self.use_policy_loss_just_for_thoughts:
+								# This will only happen when we force the next token to be the end of thought token
+								break
+							dqn_loss_list.append(actor_loss.mean())
+
+		if loss_list:
+			if self.first_and_last_mode:
+				loss = sum(
+					self.loss_mean(loss_list[-(i + 1)]) for i in range(self.n_ahead_talk)
+				) * (1 - self.original_loss_weight) / self.n_ahead_talk
+				loss = loss + self.loss_mean(loss_list[0]) * self.original_loss_weight
+				# Let's NaN out the others
+				# e.g. if n_ahead_talk = 2 and the list is 5 long, we want to NaN out 1, 2 but keep 0, 3, 4
+				for i in range(1, len(loss_list) - self.n_ahead_talk):
+					loss_list[i] = loss_list[i] * math.nan
+			elif self.first_only:
+				loss = self.loss_mean(loss_list[0])
+			elif self.final_only_mode:
+				loss = sum(
+					self.loss_mean(loss_list[-i]) for i in range(1, self.n_ahead_talk + 1)
+				) / self.n_ahead_talk   
+			else:
+				loss = None
+				for i in range(len(loss_list)):
+					cur_loss = self.loss_mean(loss_list[i])
+					if loss is not None:
+						loss = loss + cur_loss.to(loss.device)
+					else:
+						loss = cur_loss
+				loss = loss / len(loss_list)
+				loss = loss + thinking_usefulness_loss
+			
+			base_loss_scale = 0.6
+			policy_loss_scale = 0.03
+
+			loss = loss * base_loss_scale
+
+		if dqn_loss_list:
+			dqn_loss = sum(dqn_loss_list) / len(dqn_loss_list)
+			if self.include_policy_loss:
+				if loss is not None:
+					loss += dqn_loss * policy_loss_scale
+				else:
+					loss = dqn_loss * self.policy_loss_beta
+
+		if not return_dict:
+			output = (logits,) + outputs[1:]
+			return (loss,) + output if loss is not None else output
+	
+		base_log_dict = {
+			f"loss_{i}": nonzero_mean(loss_list[i]) for i in range(len(loss_list))
+		}
+
+		if loss is not None:
+			base_log_dict["loss_train"] = loss.item()
+
+		if not self.training:
+			self.n_ahead_talk = n_ahead_talk_to_restore
+			self.n_passes = n_passes_to_restore
+		
+		del start_embedding
+		del end_embedding
+		torch.cuda.empty_cache()
+		
+
+		return CausalLMOutputWithPast(
+			loss=loss if loss is not None else None,
+			logits=(rm_logits if self.n_ahead > 1 else logits) if not self.output_logits_at_the_end else 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
+	):
+		# Omit tokens covered by past_key_values
+		if past_key_values is not None:
+			if isinstance(past_key_values, Cache):
+				cache_length = past_key_values.get_seq_length()
+				past_length = past_key_values.seen_tokens
+				max_cache_length = past_key_values.get_max_length()
+			else:
+				cache_length = past_length = past_key_values[0][0].shape[2]
+				max_cache_length = None
+
+			# Keep only the unprocessed tokens:
+			# 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
+			# some of the inputs are exclusively passed as part of the cache (e.g. when passing inputs_embeds as
+			# input)
+			if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
+				input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
+			# 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
+			# input_ids based on the past_length.
+			elif past_length < input_ids.shape[1]:
+				input_ids = input_ids[:, past_length:]
+			# 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
+
+			# If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
+			if (
+				max_cache_length is not None
+				and attention_mask is not None
+				and cache_length + input_ids.shape[1] > max_cache_length
+			):
+				attention_mask = attention_mask[:, -max_cache_length:]
+
+		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[:, -input_ids.shape[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
+
+	@staticmethod
+	def _reorder_cache(past_key_values, beam_idx):
+		reordered_past = ()
+		for layer_past in past_key_values:
+			reordered_past += (
+				tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
+			)
+		return reordered_past
+	
+	
+
+
+@add_start_docstrings(
+	"""
+	The Quiet Model transformer with a sequence classification head on top (linear layer).
+	[`QuietForSequenceClassification`] uses the last token in order to do the classification, as other causal models
+	(e.g. GPT-2) do.
+	Since it does classification on the last token, it requires to know the position of the last token. If a
+	`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
+	no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
+	padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
+	each row of the batch).
+	""",
+	QUIET_START_DOCSTRING,
+)
+# Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->Quiet, LLAMA->QUIET
+class QuietForSequenceClassification(QuietPreTrainedModel):
+	def __init__(self, config):
+		super().__init__(config)
+		self.num_labels = config.num_labels
+		self.model = QuietModel(config)
+		self.score = nn.Linear(config.hidden_size, self.num_labels, 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
+
+	@add_start_docstrings_to_model_forward(QUIET_INPUTS_DOCSTRING)
+	def forward(
+		self,
+		input_ids: torch.LongTensor = None,
+		attention_mask: Optional[torch.Tensor] = None,
+		position_ids: Optional[torch.LongTensor] = None,
+		past_key_values: Optional[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,
+	) -> Union[Tuple, SequenceClassifierOutputWithPast]:
+		r"""
+		labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+			Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+			config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+			`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+		"""
+		return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+		transformer_outputs = self.model(
+			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,
+		)
+		hidden_states = transformer_outputs[0]
+		logits = self.score(hidden_states)
+
+		if input_ids is not None:
+			batch_size = input_ids.shape[0]
+		else:
+			batch_size = inputs_embeds.shape[0]
+
+		if self.config.pad_token_id is None and batch_size != 1:
+			raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
+		if self.config.pad_token_id is None:
+			sequence_lengths = -1
+		else:
+			if input_ids is not None:
+				# if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+				sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+				sequence_lengths = sequence_lengths % input_ids.shape[-1]
+				sequence_lengths = sequence_lengths.to(logits.device)
+			else:
+				sequence_lengths = -1
+
+		pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
+
+		loss = None
+		if labels is not None:
+			labels = labels.to(logits.device)
+			if self.config.problem_type is None:
+				if self.num_labels == 1:
+					self.config.problem_type = "regression"
+				elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+					self.config.problem_type = "single_label_classification"
+				else:
+					self.config.problem_type = "multi_label_classification"
+
+			if self.config.problem_type == "regression":
+				loss_fct = MSELoss()
+				if self.num_labels == 1:
+					loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
+				else:
+					loss = loss_fct(pooled_logits, labels)
+			elif self.config.problem_type == "single_label_classification":
+				loss_fct = CrossEntropyLoss()
+				loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
+			elif self.config.problem_type == "multi_label_classification":
+				loss_fct = BCEWithLogitsLoss()
+				loss = loss_fct(pooled_logits, labels)
+		if not return_dict:
+			output = (pooled_logits,) + transformer_outputs[1:]
+			return ((loss,) + output) if loss is not None else output
+
+		return SequenceClassifierOutputWithPast(
+			loss=loss,
+			logits=pooled_logits,
+			past_key_values=transformer_outputs.past_key_values,
+			hidden_states=transformer_outputs.hidden_states,
+			attentions=transformer_outputs.attentions,
+		)

optuna.py

@@ -0,0 +1,153 @@
+import optuna
+import torch
+import random
+from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments
+from datasets import load_dataset
+from trl import SFTTrainer
+import time
+
+# Set random seed for reproducibility
+random_seed = 42
+torch.manual_seed(random_seed)
+random.seed(random_seed)
+
+# Load dataset
+dataset = load_dataset("tatsu-lab/alpaca", split="train")
+
+
+def chatml_format(example):
+    """Format the dataset for training, accounting for empty columns."""
+    return {
+        "instruction": example['instruction'] if 'instruction' in example else " \n",
+        "input": example['input'] if 'input' in example else " \n",
+        "system": example['system'] if 'system' in example else " \n",
+        "output": example['output'] if 'output' in example else " \n",
+    }
+
+
+# Format dataset
+dataset = dataset.map(chatml_format, remove_columns=dataset.column_names)
+
+# Define the model initialization function
+def model_init(trial=None):
+    original = False
+    params = {}
+    if trial is not None:
+        n_ahead = 1
+        n_ahead_talk = 1
+        n_passes = 1
+        gumbel_temperature = 1
+        use_start_thought_token = True
+        use_end_thought_token = True
+        include_policy_loss = True
+        gumbel_detach = True
+        merged_talk_heads = True
+        residual_think_head = False
+        optimize_lm_head_only_at_start = False
+
+    model_id = "Crystalcareai/Quiet-Star-Custom"
+    tokenizer_id = model_id
+
+    model = AutoModelForCausalLM.from_pretrained(
+        model_id,
+        torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
+        max_thoughts=n_ahead + n_ahead_talk + 1,
+        merged_talk_heads=merged_talk_heads,
+        merged_lm_and_talk_heads=False,
+        merged_lm_and_think_heads=True,
+        use_concat_talk_head=True,
+        use_shallow_think=True,
+        use_shallow_talk=False,
+        use_complex_think_head=False,
+        use_complex_talk_head=True,
+        use_weighted_talk_head=True,
+        trust_remote_code=True,  
+        device_map="auto",
+    )
+
+    tokenizer = AutoTokenizer.from_pretrained(tokenizer_id, truncation=True, padding="left")
+    tokenizer.pad_token_id = tokenizer.eos_token_id
+
+    special_tokens_to_add = []
+    if model.use_start_thought_token:
+        special_tokens_to_add.append("<|startthought|>")
+    if model.use_end_thought_token:
+        special_tokens_to_add.append("<|endthought|>")
+    if special_tokens_to_add:
+        tokenizer.add_special_tokens({"additional_special_tokens": special_tokens_to_add})
+        model.resize_token_embeddings(len(tokenizer))
+    model.tokenizer = tokenizer
+    for name, module in model.named_modules():
+        if "embed" in name:
+            print(module, flush=True)
+   
+    model.gumbel_detach = gumbel_detach
+    model.include_policy_loss = include_policy_loss
+    model.use_end_thought_token = use_end_thought_token
+    model.use_start_thought_token = use_start_thought_token
+    model.n_ahead = n_ahead
+    model.n_ahead_talk = n_ahead_talk
+    model.n_passes = n_passes
+    model.residual_think_head = residual_think_head
+    model.gumbel_temperature = gumbel_temperature
+    model.original_mode = original
+    model.config_params = params
+    model.run_start = int(time.time())
+    model.train()
+    return model
+
+# Define the objective function for Optuna
+# Define the objective function for Optuna
+def objective(trial):
+    # Hyperparameters to be optimized
+    learning_rate = trial.suggest_float("learning_rate", 1e-07, 1e-06, log=True)
+    max_grad_norm = trial.suggest_float("max_grad_norm", 0.3, 1.0)
+    warmup_steps = trial.suggest_int("warmup_steps", 0, 20)
+    gradient_accumulation_steps = trial.suggest_int("gradient_accumulation_steps", 4, 8)
+    
+    model = model_init(trial)
+
+    training_args = TrainingArguments(
+        output_dir="./out",
+        num_train_epochs=3,
+        max_steps=30,
+        per_device_train_batch_size=1,
+        logging_steps=1,
+        optim="lion_32bit",
+        save_strategy="steps",
+        save_steps=3000,
+        gradient_accumulation_steps=gradient_accumulation_steps,
+        learning_rate=learning_rate,
+        max_grad_norm=max_grad_norm,
+        warmup_steps=warmup_steps,
+        lr_scheduler_type="cosine",
+        report_to="none"  # Disable reporting to avoid errors related to WandB in this context
+    )
+
+    trainer = SFTTrainer(
+        args=training_args,
+        train_dataset=dataset,
+        model=model,
+        tokenizer=model.tokenizer,
+        max_seq_length=1024,
+        dataset_text_field="output",
+    )
+
+    # Train the model and get the training loss
+    train_result = trainer.train()
+    loss = train_result.training_loss
+
+    return loss
+
+
+# Create a study and optimize
+study = optuna.create_study(storage="sqlite:///db.sqlite3") 
+study.optimize(objective, n_trials=100)
+
+# Print the best trial
+print("Best trial:")
+trial = study.best_trial
+print(f"  Loss: {trial.value}")
+print("  Params: ")
+for key, value in trial.params.items():
+    print(f"    {key}: {value}")

sft-dora-alpaca.py

@@ -0,0 +1,162 @@
+import torch
+torch.backends.cuda.matmul.allow_tf32 = True
+import random
+from transformers import AutoTokenizer, AutoModelForCausalLM, TextGenerationPipeline, AutoConfig
+from datasets import load_dataset
+from transformers import TrainingArguments
+from trl import SFTTrainer
+from peft import LoraConfig
+# from accelerate import infer_auto_device_map, init_empty_weights, dispatch_model
+from torch.nn import CrossEntropyLoss
+
+import time
+random_seed = 42
+torch.manual_seed(random_seed)
+random.seed(random_seed)
+
+dataset = load_dataset("Vezora/Tested-22k-Python-Alpaca", split="train")
+
+def chatml_format(example):
+    """Format the dataset for training, accounting for empty columns."""
+    return {
+        "instruction": example['instruction'] if 'instruction' in example else " \n",
+        "input": example['input'] if 'input' in example else " \n",
+        "system": example['system'] if 'system' in example else " \n",
+        "output": example['output'] if 'output' in example else " \n",
+    }
+
+# Format dataset
+dataset = dataset.map(chatml_format, remove_columns=dataset.column_names)
+
+n_ahead_talk_global = 4
+n_passes_global = 2
+n_ahead_global = 8
+n_examples = 0
+
+def model_init(params):
+    original = False
+    if params is None:
+        params = {}
+    else:
+        params = params.params
+    # save params to file
+    n_ahead = params.get("n_ahead", n_ahead_global if not original else 1)
+    n_ahead_talk = params.get("n_ahead_talk", n_ahead_talk_global if not original else 1)
+    n_passes = params.get("n_passes", n_passes_global if not original else 1)
+    gumbel_temperature = params.get("gumbel_temperature", 1)
+    use_start_thought_token = params.get("use_start_thought_token", True)
+    use_end_thought_token = params.get("use_end_thought_token", True)
+    include_policy_loss = params.get("include_policy_loss", True)
+    gumbel_detach = params.get("gumbel_detach", True)
+    merged_talk_heads = params.get("merged_talk_heads", True)
+    residual_think_head = params.get("residual_think_head", False)
+    optimize_lm_head_only_at_start = params.get("optimize_lm_head_only_at_start", False)
+
+    model_id = "Crystalcareai/Quiet-Star-Custom"
+    tokenizer_id = model_id
+    print("Loading model")
+
+    model = AutoModelForCausalLM.from_pretrained(
+        model_id,
+        torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
+        max_thoughts=n_ahead + n_ahead_talk + 1,
+        merged_talk_heads=merged_talk_heads,
+        merged_lm_and_talk_heads=False,
+        merged_lm_and_think_heads=True,
+        use_concat_talk_head=True,
+        use_shallow_think=True,
+        use_shallow_talk=False,
+        use_complex_think_head=False,
+        use_complex_talk_head=True,
+        use_weighted_talk_head=True,
+        trust_remote_code=True,
+        device_map="auto",
+    )
+    print("Loaded model")
+
+    tokenizer = AutoTokenizer.from_pretrained(tokenizer_id, truncation=True, padding_side="right")
+    tokenizer.pad_token_id = tokenizer.eos_token_id
+
+    special_tokens_to_add = []
+    if model.use_start_thought_token:
+        special_tokens_to_add.append("<|startthought|>")
+    if model.use_end_thought_token:
+        special_tokens_to_add.append("<|endthought|>")
+    if special_tokens_to_add:
+        tokenizer.add_special_tokens({"additional_special_tokens": special_tokens_to_add})
+        model.resize_token_embeddings(len(tokenizer))
+    model.tokenizer = tokenizer
+    for name, module in model.named_modules():
+        if "embed" in name:
+            print(module, flush=True)
+
+    model.gumbel_detach = gumbel_detach
+    model.include_policy_loss = include_policy_loss
+    model.use_end_thought_token = use_end_thought_token
+    model.use_start_thought_token = use_start_thought_token
+    model.n_ahead = n_ahead
+    model.n_ahead_talk = n_ahead_talk
+    model.n_passes = n_passes
+    model.residual_think_head = residual_think_head
+    model.optimize_lm_head_only_at_start = optimize_lm_head_only_at_start
+    model.gumbel_temperature = gumbel_temperature
+    model.original_mode = original
+    model.config_params = params
+    model.run_start = int(time.time())
+    model.train()
+    return model
+
+max_seq_length = 1024
+run_id = int(time.time())
+training_args = TrainingArguments(
+    output_dir="./out",
+    num_train_epochs=3,
+    per_device_train_batch_size=1,
+    gradient_checkpointing=False,
+    gradient_accumulation_steps=8,
+    optim="lion_32bit",
+    logging_steps=1,
+    save_strategy="steps",
+    save_steps=300,
+    max_steps=1000,
+    bf16=True,
+    tf32=False,
+    learning_rate=6e-05,
+    max_grad_norm=0.3,
+    warmup_ratio=0.06,
+    lr_scheduler_type="cosine",
+    push_to_hub=False,
+    report_to="wandb"
+)
+
+peft_config = LoraConfig(
+          r = 16, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
+    target_modules = ["q_proj", "k_proj"],
+    lora_alpha = 16,
+    lora_dropout = 0, # Supports any, but = 0 is optimized
+    bias = "none", 
+    use_dora=True,
+)
+
+
+torch.autograd.set_detect_anomaly(True)
+
+# Set the device for each process
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+# torch.cuda.set_device(device)
+
+model = model_init(None)  # Initialize the model
+
+tokenizer = model.tokenizer
+
+trainer = SFTTrainer(
+    args=training_args,
+    train_dataset=dataset,
+    model=model,
+    tokenizer=tokenizer,
+    max_seq_length=max_seq_length,
+    dataset_text_field="output",
+    peft_config=peft_config,
+)
+
+trainer.train()

special_tokens_map.json

@@ -0,0 +1,40 @@
+{
+  "additional_special_tokens": [
+    {
+      "content": "<|endthought|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false
+    },
+    {
+      "content": "<|startthought|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false
+    }
+  ],
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "</s>",
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dadfd56d766715c61d2ef780a525ab43b8e6da4de6865bda3d95fdef5e134055
+size 493443

tokenizer_config.json

@@ -0,0 +1,61 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32000": {
+      "content": "<|endthought|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32001": {
+      "content": "<|startthought|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "additional_special_tokens": [
+    "<|endthought|>",
+    "<|startthought|>"
+  ],
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "legacy": true,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "</s>",
+  "sp_model_kwargs": {},
+  "spaces_between_special_tokens": false,
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}

train-h100-sharegpt-sft.py

@@ -0,0 +1,193 @@
+import torch
+torch.backends.cuda.matmul.allow_tf32 = True
+import random
+from transformers import AutoTokenizer, AutoModelForCausalLM, TextGenerationPipeline, AutoConfig, BitsAndBytesConfig
+from datasets import load_dataset
+from transformers import TrainingArguments
+from accelerate import infer_auto_device_map, init_empty_weights, dispatch_model
+from trl import SFTTrainer
+from peft import LoraConfig
+from torch.nn import CrossEntropyLoss
+import time
+import gc
+
+random_seed = 42
+torch.manual_seed(random_seed)
+random.seed(random_seed)
+
+dataset = load_dataset("HuggingFaceH4/orca-math-word-problems-200k", split="train_sft").select(range(1000))
+
+
+n_ahead_talk_global = 4
+n_passes_global = 1
+n_ahead_global = 4
+# n_examples = 1000
+# full_batch_size = 8
+
+def model_init(params):
+    original = False
+    if params is None:
+        params = {}
+    else:
+        params = params.params
+    # save params to file
+    n_ahead = params.get("n_ahead", n_ahead_global if not original else 1)
+    n_ahead_talk = params.get("n_ahead_talk", n_ahead_talk_global if not original else 1)
+    n_passes = params.get("n_passes", n_passes_global if not original else 1)
+    gumbel_temperature = params.get("gumbel_temperature", 1)
+    use_start_thought_token = params.get("use_start_thought_token", True)
+    use_end_thought_token = params.get("use_end_thought_token", True)
+    include_policy_loss = params.get("include_policy_loss", True)
+    gumbel_detach = params.get("gumbel_detach", True)
+    merged_talk_heads = params.get("merged_talk_heads", True)
+    residual_think_head = params.get("residual_think_head", False)
+    optimize_lm_head_only_at_start = params.get("optimize_lm_head_only_at_start", False)
+
+    model_id = "Crystalcareai/Quiet-Star-Custom"
+    tokenizer_id = model_id
+    print("Loading model")
+
+    model = AutoModelForCausalLM.from_pretrained(
+        model_id,
+        torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
+        max_thoughts=n_ahead + n_ahead_talk + 1,
+        merged_talk_heads=merged_talk_heads,
+        merged_lm_and_talk_heads=False,
+        merged_lm_and_think_heads=True,
+        use_concat_talk_head=True,
+        use_shallow_think=True,
+        use_shallow_talk=False,
+        use_complex_think_head=False,
+        use_complex_talk_head=True,
+        use_weighted_talk_head=True,
+        trust_remote_code=True,
+        device_map="auto",
+    )
+    print("Loaded model")
+
+    tokenizer = AutoTokenizer.from_pretrained(tokenizer_id, truncation=True, padding_side="right")
+    tokenizer.pad_token_id = tokenizer.eos_token_id
+
+    special_tokens_to_add = []
+    if model.use_start_thought_token:
+        special_tokens_to_add.append("<|startthought|>")
+    if model.use_end_thought_token:
+        special_tokens_to_add.append("<|endthought|>")
+    if special_tokens_to_add:
+        tokenizer.add_special_tokens({"additional_special_tokens": special_tokens_to_add})
+    model.tokenizer = tokenizer
+    for name, module in model.named_modules():
+        if "embed" in name:
+            print(module, flush=True)
+
+    model.gumbel_detach = gumbel_detach
+    model.include_policy_loss = include_policy_loss
+    model.use_end_thought_token = use_end_thought_token
+    model.use_start_thought_token = use_start_thought_token
+    model.n_ahead = n_ahead
+    model.n_ahead_talk = n_ahead_talk
+    model.n_passes = n_passes
+    model.residual_think_head = residual_think_head
+    model.optimize_lm_head_only_at_start = optimize_lm_head_only_at_start
+    model.gumbel_temperature = gumbel_temperature
+    model.original_mode = original
+    model.config_params = params
+    model.run_start = int(time.time())
+    model.train()
+    return model
+
+max_seq_length = 1024
+run_id = int(time.time())
+training_args = TrainingArguments(
+    output_dir="./out",
+    num_train_epochs=1,
+    per_device_train_batch_size=1,
+    gradient_checkpointing=False,
+    gradient_accumulation_steps=8,
+    optim="adamw_torch_fused",
+    logging_steps=1,
+    save_strategy="steps",
+    save_steps=100,
+    max_steps=-1,
+    # auto_find_batch_size=True,
+    weight_decay=0.001,
+    bf16=True,
+    
+    tf32=True,
+    learning_rate=2e-10,
+    max_grad_norm=0,
+    warmup_steps=20,
+    lr_scheduler_type="cosine",
+    push_to_hub=False,
+    report_to="wandb"
+)
+
+peft_config = LoraConfig(
+    r = 8, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
+    target_modules =["q_proj", "v_proj"],
+    lora_alpha = 32,
+    lora_dropout = 0, # Supports any, but = 0 is optimized
+    bias = "none", 
+    use_dora=True,
+    task_type="CAUSAL_LM"
+)
+
+torch.autograd.set_detect_anomaly(True)
+
+# class CustomSFTTrainer(SFTTrainer):
+#     def __init__(self, *args, **kwargs):
+#         super().__init__(*args, **kwargs)
+#         self.beta = 0.9  # momentum factor
+#         self.clip_factor = 1.0  # clipping factor
+#         self.moving_avg = 0.0
+
+#     def training_step(self, model, inputs):
+#         model.train()
+#         inputs = self._prepare_inputs(inputs)
+
+#         outputs = model(**inputs)
+#         loss = outputs.loss if isinstance(outputs, dict) else outputs[0]
+
+#         if self.args.gradient_accumulation_steps > 1:
+#             loss = loss / self.args.gradient_accumulation_steps
+
+#         loss.backward()
+
+#         # Compute gradients and their norm
+#         grad_norm = torch.sqrt(sum(p.grad.data.norm().to(model.device)**2 for p in model.parameters() if p.grad is not None))
+
+#         # Update moving average and apply gradient clipping
+#         if self.state.global_step == 0:
+#             self.moving_avg = grad_norm
+#         else:
+#             self.moving_avg = self.beta * self.moving_avg + (1 - self.beta) * grad_norm
+
+#         if grad_norm > self.clip_factor * self.moving_avg:
+#             clip_coef = (self.clip_factor * self.moving_avg / grad_norm).item()
+#             for param in model.parameters():
+#                 if param.grad is not None:
+#                     param.grad.data.mul_(clip_coef)
+
+#         if (self.state.global_step + 1) % self.args.gradient_accumulation_steps == 0:
+#             self.optimizer.step()
+#             self.lr_scheduler.step()
+#             model.zero_grad()
+#             self.state.global_step += 1
+
+#         # Return the loss as a Tensor
+#         return loss
+    
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+model = model_init(None)  
+
+trainer = SFTTrainer(
+    model=model,
+    args=training_args,
+    train_dataset=dataset,
+    tokenizer=model.tokenizer,
+    max_seq_length=max_seq_length,
+    peft_config=peft_config,
+)
+
+trainer.train()