Initial Kronos-small custom deployment

.gitattributes

@@ -0,0 +1 @@
+*.safetensors filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,4 @@
+__pycache__/
+*.pyc
+.env*
+.DS_Store

README.md

@@ -0,0 +1,120 @@
+## `faizack/kronos-small-custom`
+
+Custom Hugging Face model repo for deploying the **Kronos-small** time-series forecasting model as an Inference Endpoint.
+
+This folder is structured so you can **zip it or `git init` it and push directly** to Hugging Face under your account [`faizack`](https://huggingface.co/faizack).
+
+### 1. Files expected in this repo
+
+You will need the following files in the root of the Hugging Face repo:
+
+- `config.json` – copied or downloaded from `NeoQuasar/Kronos-small`
+- `model.safetensors` – weights from `NeoQuasar/Kronos-small`
+- `model.py` – Kronos model definition (from the official GitHub repo)
+- `tokenizer.py` – Kronos tokenizer implementation
+- `predictor.py` – `KronosPredictor` wrapper
+- `inference.py` – entrypoint used by Hugging Face Inference Endpoints (already provided here)
+- `requirements.txt` – Python dependencies (already provided here)
+
+This folder currently includes:
+
+- `README.md` (this file)
+- `inference.py`
+- `requirements.txt`
+- `.env.example`
+- `.gitattributes` (Git LFS for safetensors)
+- `.gitignore`
+
+You still need to add:
+
+- `config.json`
+- `model.safetensors`
+- `model.py`
+- `tokenizer.py`
+- `predictor.py`
+
+### 2. How to prepare and push to Hugging Face
+
+From this folder:
+
+```bash
+cd kronos-small-custom
+
+# (optional) initialize git
+git init
+git lfs install
+
+# Log in to Hugging Face
+huggingface-cli login
+
+# Create the remote repo under your account
+huggingface-cli repo create faizack/kronos-small-custom --type model
+
+# Add HF remote
+git remote add origin https://huggingface.co/faizack/kronos-small-custom
+```
+
+Now copy in the Kronos implementation and weights:
+
+1. From the official Kronos GitHub repo, copy:
+   - `model.py`
+   - `tokenizer.py`
+   - `predictor.py`
+2. From `NeoQuasar/Kronos-small` on Hugging Face, download:
+   - `config.json`
+   - `model.safetensors`
+
+Then commit and push:
+
+```bash
+git add .
+git commit -m "Initial Kronos-small custom deployment"
+git push -u origin main
+```
+
+### 3. Inference contract
+
+`inference.py` exposes a `predict(request)` function that Hugging Face Inference Endpoints will call.
+
+Expected JSON body:
+
+```json
+{
+  "inputs": {
+    "df": [
+      {"open": 1.0, "high": 1.1, "low": 0.9, "close": 1.05},
+      {"open": 1.05, "high": 1.12, "low": 1.0, "close": 1.08}
+    ],
+    "x_timestamp": ["2024-01-01T00:00:00Z", "2024-01-01T01:00:00Z"],
+    "y_timestamp": ["2024-01-01T02:00:00Z", "2024-01-01T03:00:00Z"],
+    "pred_len": 2,
+    "T": 1.0,
+    "top_p": 0.9,
+    "sample_count": 1
+  }
+}
+```
+
+Response structure:
+
+```json
+{
+  "predictions": [
+    {
+      "open": ...,
+      "high": ...,
+      "low": ...,
+      "close": ...
+    },
+    {
+      "open": ...,
+      "high": ...,
+      "low": ...,
+      "close": ...
+    }
+  ]
+}
+```
+
+You can adapt this contract as needed, as long as `predict` returns JSON-serializable data.
+

config.json

@@ -0,0 +1,13 @@
+{
+  "attn_dropout_p": 0.1,
+  "d_model": 512,
+  "ff_dim": 1024,
+  "ffn_dropout_p": 0.25,
+  "learn_te": true,
+  "n_heads": 8,
+  "n_layers": 8,
+  "resid_dropout_p": 0.25,
+  "s1_bits": 10,
+  "s2_bits": 10,
+  "token_dropout_p": 0.1
+}

inference.py

@@ -0,0 +1,82 @@
+import os
+from typing import Any, Dict
+
+import pandas as pd
+import torch
+
+from kronos import Kronos, KronosTokenizer, KronosPredictor  # type: ignore
+
+
+DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+
+
+def _load_components(model_dir: str = "."):
+    """
+    Load tokenizer, model, and predictor from a local directory.
+
+    This is called once at module import time on HF Inference Endpoints.
+    """
+    tokenizer = KronosTokenizer.from_pretrained(model_dir)
+    model = Kronos.from_pretrained(model_dir).to(DEVICE)
+
+    max_context = int(os.getenv("KRONOS_MAX_CONTEXT", "512"))
+
+    predictor = KronosPredictor(
+        model=model,
+        tokenizer=tokenizer,
+        device=DEVICE,
+        max_context=max_context,
+    )
+
+    return tokenizer, model, predictor
+
+
+TOKENIZER, MODEL, PREDICTOR = _load_components(".")
+
+
+def predict(request: Dict[str, Any]) -> Dict[str, Any]:
+    """
+    Entry point for Hugging Face Inference Endpoints.
+
+    Expected input format:
+
+    {
+        "inputs": {
+            "df": [
+                {"open": ..., "high": ..., "low": ..., "close": ...},
+                ...
+            ],
+            "x_timestamp": [...],   # list of ISO8601 strings or timestamps
+            "y_timestamp": [...],   # list of ISO8601 strings or timestamps
+            "pred_len": 120,
+            "T": 1.0,               # optional
+            "top_p": 0.9,           # optional
+            "sample_count": 1       # optional
+        }
+    }
+    """
+    inputs = request.get("inputs", request)
+
+    df = pd.DataFrame(inputs["df"])
+    x_timestamp = pd.to_datetime(inputs["x_timestamp"])
+    y_timestamp = pd.to_datetime(inputs["y_timestamp"])
+
+    pred_len = int(inputs["pred_len"])
+    T = float(inputs.get("T", 1.0))
+    top_p = float(inputs.get("top_p", 0.9))
+    sample_count = int(inputs.get("sample_count", 1))
+
+    result_df = PREDICTOR.predict(
+        df=df,
+        x_timestamp=x_timestamp,
+        y_timestamp=y_timestamp,
+        pred_len=pred_len,
+        T=T,
+        top_p=top_p,
+        sample_count=sample_count,
+    )
+
+    # Return a plain dict for JSON serialization
+    return {
+        "predictions": result_df.to_dict(orient="records"),
+    }

kronos.py

@@ -0,0 +1,835 @@
+import numpy as np
+import pandas as pd
+from huggingface_hub import PyTorchModelHubMixin
+from tqdm import trange
+from module import *
+
+
+class KronosTokenizer(nn.Module, PyTorchModelHubMixin):
+    """
+    KronosTokenizer module for tokenizing input data using a hybrid quantization approach.
+
+    This tokenizer utilizes a combination of encoder and decoder Transformer blocks
+    along with the Binary Spherical Quantization (BSQuantizer) to compress and decompress input data.
+
+    Args:
+           d_in (int): Input dimension.
+           d_model (int): Model dimension.
+           n_heads (int): Number of attention heads.
+           ff_dim (int): Feed-forward dimension.
+           n_enc_layers (int): Number of encoder layers.
+           n_dec_layers (int): Number of decoder layers.
+           ffn_dropout_p (float): Dropout probability for feed-forward networks.
+           attn_dropout_p (float): Dropout probability for attention mechanisms.
+           resid_dropout_p (float): Dropout probability for residual connections.
+           s1_bits (int): Number of bits for the pre token in BSQuantizer.
+           s2_bits (int): Number of bits for the post token in BSQuantizer.
+           beta (float): Beta parameter for BSQuantizer.
+           gamma0 (float): Gamma0 parameter for BSQuantizer.
+           gamma (float): Gamma parameter for BSQuantizer.
+           zeta (float): Zeta parameter for BSQuantizer.
+           group_size (int): Group size parameter for BSQuantizer.
+
+    """
+
+    def __init__(
+        self,
+        d_in,
+        d_model,
+        n_heads,
+        ff_dim,
+        n_enc_layers,
+        n_dec_layers,
+        ffn_dropout_p,
+        attn_dropout_p,
+        resid_dropout_p,
+        s1_bits,
+        s2_bits,
+        beta,
+        gamma0,
+        gamma,
+        zeta,
+        group_size,
+    ):
+
+        super().__init__()
+        self.d_in = d_in
+        self.d_model = d_model
+        self.n_heads = n_heads
+        self.ff_dim = ff_dim
+        self.enc_layers = n_enc_layers
+        self.dec_layers = n_dec_layers
+        self.ffn_dropout_p = ffn_dropout_p
+        self.attn_dropout_p = attn_dropout_p
+        self.resid_dropout_p = resid_dropout_p
+
+        self.s1_bits = s1_bits
+        self.s2_bits = s2_bits
+        self.codebook_dim = (
+            s1_bits + s2_bits
+        )  # Total dimension of the codebook after quantization
+        self.embed = nn.Linear(self.d_in, self.d_model)
+        self.head = nn.Linear(self.d_model, self.d_in)
+
+        # Encoder Transformer Blocks
+        self.encoder = nn.ModuleList(
+            [
+                TransformerBlock(
+                    self.d_model,
+                    self.n_heads,
+                    self.ff_dim,
+                    self.ffn_dropout_p,
+                    self.attn_dropout_p,
+                    self.resid_dropout_p,
+                )
+                for _ in range(self.enc_layers - 1)
+            ]
+        )
+        # Decoder Transformer Blocks
+        self.decoder = nn.ModuleList(
+            [
+                TransformerBlock(
+                    self.d_model,
+                    self.n_heads,
+                    self.ff_dim,
+                    self.ffn_dropout_p,
+                    self.attn_dropout_p,
+                    self.resid_dropout_p,
+                )
+                for _ in range(self.dec_layers - 1)
+            ]
+        )
+        self.quant_embed = nn.Linear(
+            in_features=self.d_model, out_features=self.codebook_dim
+        )  # Linear layer before quantization
+        self.post_quant_embed_pre = nn.Linear(
+            in_features=self.s1_bits, out_features=self.d_model
+        )  # Linear layer after quantization (pre part - s1 bits)
+        self.post_quant_embed = nn.Linear(
+            in_features=self.codebook_dim, out_features=self.d_model
+        )  # Linear layer after quantization (full codebook)
+        self.tokenizer = BSQuantizer(
+            self.s1_bits, self.s2_bits, beta, gamma0, gamma, zeta, group_size
+        )  # BSQuantizer module
+
+    def forward(self, x):
+        """
+        Forward pass of the KronosTokenizer.
+
+        Args:
+            x (torch.Tensor): Input tensor of shape (batch_size, seq_len, d_in).
+
+        Returns:
+            tuple: A tuple containing:
+                - tuple: (z_pre, z) - Reconstructed outputs from decoder with s1_bits and full codebook respectively,
+                         both of shape (batch_size, seq_len, d_in).
+                - torch.Tensor: bsq_loss - Loss from the BSQuantizer.
+                - torch.Tensor: quantized - Quantized representation from BSQuantizer.
+                - torch.Tensor: z_indices - Indices from the BSQuantizer.
+        """
+        z = self.embed(x)
+
+        for layer in self.encoder:
+            z = layer(z)
+
+        z = self.quant_embed(z)  # (B, T, codebook)
+
+        bsq_loss, quantized, z_indices = self.tokenizer(z)
+
+        quantized_pre = quantized[
+            :, :, : self.s1_bits
+        ]  # Extract the first part of quantized representation (s1_bits)
+        z_pre = self.post_quant_embed_pre(quantized_pre)
+
+        z = self.post_quant_embed(quantized)
+
+        # Decoder layers (for pre part - s1 bits)
+        for layer in self.decoder:
+            z_pre = layer(z_pre)
+        z_pre = self.head(z_pre)
+
+        # Decoder layers (for full codebook)
+        for layer in self.decoder:
+            z = layer(z)
+        z = self.head(z)
+
+        return (z_pre, z), bsq_loss, quantized, z_indices
+
+    def indices_to_bits(self, x, half=False):
+        """
+        Converts indices to bit representations and scales them.
+
+        Args:
+            x (torch.Tensor): Indices tensor.
+            half (bool, optional): Whether to process only half of the codebook dimension. Defaults to False.
+
+        Returns:
+            torch.Tensor: Bit representation tensor.
+        """
+        if half:
+            x1 = x[0]  # Assuming x is a tuple of indices if half is True
+            x2 = x[1]
+            mask = 2 ** torch.arange(
+                self.codebook_dim // 2, device=x1.device, dtype=torch.long
+            )  # Create a mask for bit extraction
+            x1 = (x1.unsqueeze(-1) & mask) != 0  # Extract bits for the first half
+            x2 = (x2.unsqueeze(-1) & mask) != 0  # Extract bits for the second half
+            x = torch.cat([x1, x2], dim=-1)  # Concatenate the bit representations
+        else:
+            mask = 2 ** torch.arange(
+                self.codebook_dim, device=x.device, dtype=torch.long
+            )  # Create a mask for bit extraction
+            x = (x.unsqueeze(-1) & mask) != 0  # Extract bits
+
+        x = x.float() * 2 - 1  # Convert boolean to bipolar (-1, 1)
+        q_scale = 1.0 / (self.codebook_dim**0.5)  # Scaling factor
+        x = x * q_scale
+        return x
+
+    def encode(self, x, half=False):
+        """
+        Encodes the input data into quantized indices.
+
+        Args:
+            x (torch.Tensor): Input tensor of shape (batch_size, seq_len, d_in).
+            half (bool, optional): Whether to use half quantization in BSQuantizer. Defaults to False.
+
+        Returns:
+            torch.Tensor: Quantized indices from BSQuantizer.
+        """
+        z = self.embed(x)
+        for layer in self.encoder:
+            z = layer(z)
+        z = self.quant_embed(z)
+
+        bsq_loss, quantized, z_indices = self.tokenizer(z, half)
+        return z_indices
+
+    def decode(self, x, half=False):
+        """
+        Decodes quantized indices back to the input data space.
+
+        Args:
+            x (torch.Tensor): Quantized indices tensor.
+            half (bool, optional): Whether the indices were generated with half quantization. Defaults to False.
+
+        Returns:
+            torch.Tensor: Reconstructed output tensor of shape (batch_size, seq_len, d_in).
+        """
+        quantized = self.indices_to_bits(x, half)
+        z = self.post_quant_embed(quantized)
+        for layer in self.decoder:
+            z = layer(z)
+        z = self.head(z)
+        return z
+
+
+class Kronos(nn.Module, PyTorchModelHubMixin):
+    """
+    Kronos Model.
+
+    Args:
+        s1_bits (int): Number of bits for pre tokens.
+        s2_bits (int): Number of bits for post tokens.
+        n_layers (int): Number of Transformer blocks.
+        d_model (int): Dimension of the model's embeddings and hidden states.
+        n_heads (int): Number of attention heads in the MultiheadAttention layers.
+        ff_dim (int): Dimension of the feedforward network in the Transformer blocks.
+        ffn_dropout_p (float): Dropout probability for the feedforward network.
+        attn_dropout_p (float): Dropout probability for the attention layers.
+        resid_dropout_p (float): Dropout probability for residual connections.
+        token_dropout_p (float): Dropout probability for token embeddings.
+        learn_te (bool): Whether to use learnable temporal embeddings.
+    """
+
+    def __init__(
+        self,
+        s1_bits,
+        s2_bits,
+        n_layers,
+        d_model,
+        n_heads,
+        ff_dim,
+        ffn_dropout_p,
+        attn_dropout_p,
+        resid_dropout_p,
+        token_dropout_p,
+        learn_te,
+    ):
+        super().__init__()
+        self.s1_bits = s1_bits
+        self.s2_bits = s2_bits
+        self.n_layers = n_layers
+        self.d_model = d_model
+        self.n_heads = n_heads
+        self.learn_te = learn_te
+        self.ff_dim = ff_dim
+        self.ffn_dropout_p = ffn_dropout_p
+        self.attn_dropout_p = attn_dropout_p
+        self.resid_dropout_p = resid_dropout_p
+        self.token_dropout_p = token_dropout_p
+
+        self.s1_vocab_size = 2**self.s1_bits
+        self.token_drop = nn.Dropout(self.token_dropout_p)
+        self.embedding = HierarchicalEmbedding(self.s1_bits, self.s2_bits, self.d_model)
+        self.time_emb = TemporalEmbedding(self.d_model, self.learn_te)
+        self.transformer = nn.ModuleList(
+            [
+                TransformerBlock(
+                    self.d_model,
+                    self.n_heads,
+                    self.ff_dim,
+                    self.ffn_dropout_p,
+                    self.attn_dropout_p,
+                    self.resid_dropout_p,
+                )
+                for _ in range(self.n_layers)
+            ]
+        )
+        self.norm = RMSNorm(self.d_model)
+        self.dep_layer = DependencyAwareLayer(self.d_model)
+        self.head = DualHead(self.s1_bits, self.s2_bits, self.d_model)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, module):
+
+        if isinstance(module, nn.Linear):
+            nn.init.xavier_normal_(module.weight)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, mean=0, std=self.embedding.d_model**-0.5)
+        elif isinstance(module, nn.LayerNorm):
+            nn.init.ones_(module.weight)
+            nn.init.zeros_(module.bias)
+        elif isinstance(module, RMSNorm):
+            nn.init.ones_(module.weight)
+
+    def forward(
+        self,
+        s1_ids,
+        s2_ids,
+        stamp=None,
+        padding_mask=None,
+        use_teacher_forcing=False,
+        s1_targets=None,
+    ):
+        """
+        Args:
+            s1_ids (torch.Tensor): Input tensor of s1 token IDs. Shape: [batch_size, seq_len]
+            s2_ids (torch.Tensor): Input tensor of s2 token IDs. Shape: [batch_size, seq_len]
+            stamp (torch.Tensor, optional): Temporal stamp tensor. Shape: [batch_size, seq_len]. Defaults to None.
+            padding_mask (torch.Tensor, optional): Mask for padding tokens. Shape: [batch_size, seq_len]. Defaults to None.
+            use_teacher_forcing (bool, optional): Whether to use teacher forcing for s1 decoding. Defaults to False.
+            s1_targets (torch.Tensor, optional): Target s1 token IDs for teacher forcing. Shape: [batch_size, seq_len]. Defaults to None.
+
+        Returns:
+            Tuple[torch.Tensor, torch.Tensor]:
+                - s1 logits: Logits for s1 token predictions. Shape: [batch_size, seq_len, s1_vocab_size]
+                - s2_logits: Logits for s2 token predictions, conditioned on s1. Shape: [batch_size, seq_len, s2_vocab_size]
+        """
+        x = self.embedding([s1_ids, s2_ids])
+        if stamp is not None:
+            time_embedding = self.time_emb(stamp)
+            x = x + time_embedding
+        x = self.token_drop(x)
+
+        for layer in self.transformer:
+            x = layer(x, key_padding_mask=padding_mask)
+
+        x = self.norm(x)
+
+        s1_logits = self.head(x)
+
+        if use_teacher_forcing:
+            sibling_embed = self.embedding.emb_s1(s1_targets)
+        else:
+            s1_probs = F.softmax(s1_logits.detach(), dim=-1)
+            sample_s1_ids = torch.multinomial(
+                s1_probs.view(-1, self.s1_vocab_size), 1
+            ).view(s1_ids.shape)
+            sibling_embed = self.embedding.emb_s1(sample_s1_ids)
+
+        x2 = self.dep_layer(
+            x, sibling_embed, key_padding_mask=padding_mask
+        )  # Dependency Aware Layer: Condition on s1 embeddings
+        s2_logits = self.head.cond_forward(x2)
+        return s1_logits, s2_logits
+
+    def decode_s1(self, s1_ids, s2_ids, stamp=None, padding_mask=None):
+        """
+        Decodes only the s1 tokens.
+
+        This method performs a forward pass to predict only s1 tokens. It returns the s1 logits
+        and the context representation from the Transformer, which can be used for subsequent s2 decoding.
+
+        Args:
+            s1_ids (torch.Tensor): Input tensor of s1 token IDs. Shape: [batch_size, seq_len]
+            s2_ids (torch.Tensor): Input tensor of s2 token IDs. Shape: [batch_size, seq_len]
+            stamp (torch.Tensor, optional): Temporal stamp tensor. Shape: [batch_size, seq_len]. Defaults to None.
+            padding_mask (torch.Tensor, optional): Mask for padding tokens. Shape: [batch_size, seq_len]. Defaults to None.
+
+        Returns:
+            Tuple[torch.Tensor, torch.Tensor]:
+                - s1 logits: Logits for s1 token predictions. Shape: [batch_size, seq_len, s1_vocab_size]
+                - context: Context representation from the Transformer. Shape: [batch_size, seq_len, d_model]
+        """
+        x = self.embedding([s1_ids, s2_ids])
+        if stamp is not None:
+            time_embedding = self.time_emb(stamp)
+            x = x + time_embedding
+        x = self.token_drop(x)
+
+        for layer in self.transformer:
+            x = layer(x, key_padding_mask=padding_mask)
+
+        x = self.norm(x)
+
+        s1_logits = self.head(x)
+        return s1_logits, x
+
+    def decode_s2(self, context, s1_ids, padding_mask=None):
+        """
+        Decodes the s2 tokens, conditioned on the context and s1 tokens.
+
+        This method decodes s2 tokens based on a pre-computed context representation (typically from `decode_s1`)
+        and the s1 token IDs. It uses the dependency-aware layer and the conditional s2 head to predict s2 tokens.
+
+        Args:
+            context (torch.Tensor): Context representation from the transformer (output of decode_s1).
+                                     Shape: [batch_size, seq_len, d_model]
+            s1_ids (torch.torch.Tensor): Input tensor of s1 token IDs. Shape: [batch_size, seq_len]
+            padding_mask (torch.Tensor, optional): Mask for padding tokens. Shape: [batch_size, seq_len]. Defaults to None.
+
+        Returns:
+            torch.Tensor: s2 logits. Shape: [batch_size, seq_len, s2_vocab_size]
+        """
+        sibling_embed = self.embedding.emb_s1(s1_ids)
+        x2 = self.dep_layer(context, sibling_embed, key_padding_mask=padding_mask)
+        return self.head.cond_forward(x2)
+
+
+def top_k_top_p_filtering(
+    logits,
+    top_k: int = 0,
+    top_p: float = 1.0,
+    filter_value: float = -float("Inf"),
+    min_tokens_to_keep: int = 1,
+):
+    """Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
+    Args:
+        logits: logits distribution shape (batch size, vocabulary size)
+        if top_k > 0: keep only top k tokens with highest probability (top-k filtering).
+        if top_p < 1.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
+            Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
+        Make sure we keep at least min_tokens_to_keep per batch example in the output
+    From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
+    """
+    if top_k > 0:
+        top_k = min(max(top_k, min_tokens_to_keep), logits.size(-1))  # Safety check
+        # Remove all tokens with a probability less than the last token of the top-k
+        indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+        logits[indices_to_remove] = filter_value
+        return logits
+
+    if top_p < 1.0:
+        sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+        cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
+
+        # Remove tokens with cumulative probability above the threshold (token with 0 are kept)
+        sorted_indices_to_remove = cumulative_probs > top_p
+        if min_tokens_to_keep > 1:
+            # Keep at least min_tokens_to_keep (set to min_tokens_to_keep-1 because we add the first one below)
+            sorted_indices_to_remove[..., :min_tokens_to_keep] = 0
+        # Shift the indices to the right to keep also the first token above the threshold
+        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+        sorted_indices_to_remove[..., 0] = 0
+
+        # scatter sorted tensors to original indexing
+        indices_to_remove = sorted_indices_to_remove.scatter(
+            1, sorted_indices, sorted_indices_to_remove
+        )
+        logits[indices_to_remove] = filter_value
+        return logits
+
+
+def sample_from_logits(
+    logits, temperature=1.0, top_k=None, top_p=None, sample_logits=True
+):
+    logits = logits / temperature
+    if top_k is not None or top_p is not None:
+        if top_k > 0 or top_p < 1.0:
+            logits = top_k_top_p_filtering(logits, top_k=top_k, top_p=top_p)
+
+    probs = F.softmax(logits, dim=-1)
+
+    if not sample_logits:
+        _, x = top_k(probs, k=1, dim=-1)
+    else:
+        x = torch.multinomial(probs, num_samples=1)
+
+    return x
+
+
+def auto_regressive_inference(
+    tokenizer,
+    model,
+    x,
+    x_stamp,
+    y_stamp,
+    max_context,
+    pred_len,
+    clip=5,
+    T=1.0,
+    top_k=0,
+    top_p=0.99,
+    sample_count=5,
+    verbose=False,
+):
+    with torch.no_grad():
+        batch_size = x.size(0)
+        initial_seq_len = x.size(1)
+        x = torch.clip(x, -clip, clip)
+
+        device = x.device
+        x = (
+            x.unsqueeze(1)
+            .repeat(1, sample_count, 1, 1)
+            .reshape(-1, x.size(1), x.size(2))
+            .to(device)
+        )
+        x_stamp = (
+            x_stamp.unsqueeze(1)
+            .repeat(1, sample_count, 1, 1)
+            .reshape(-1, x_stamp.size(1), x_stamp.size(2))
+            .to(device)
+        )
+        y_stamp = (
+            y_stamp.unsqueeze(1)
+            .repeat(1, sample_count, 1, 1)
+            .reshape(-1, y_stamp.size(1), y_stamp.size(2))
+            .to(device)
+        )
+
+        x_token = tokenizer.encode(x, half=True)
+
+        def get_dynamic_stamp(x_stamp, y_stamp, current_seq_len, pred_step):
+
+            if current_seq_len <= max_context - pred_step:
+                return torch.cat([x_stamp, y_stamp[:, :pred_step, :]], dim=1)
+            else:
+                start_idx = max_context - pred_step
+                return torch.cat(
+                    [x_stamp[:, -start_idx:, :], y_stamp[:, :pred_step, :]], dim=1
+                )
+
+        if verbose:
+            ran = trange
+        else:
+            ran = range
+        for i in ran(pred_len):
+            current_seq_len = initial_seq_len + i
+
+            if current_seq_len <= max_context:
+                input_tokens = x_token
+            else:
+                input_tokens = [t[:, -max_context:].contiguous() for t in x_token]
+
+            current_stamp = get_dynamic_stamp(x_stamp, y_stamp, current_seq_len, i)
+
+            s1_logits, context = model.decode_s1(
+                input_tokens[0], input_tokens[1], current_stamp
+            )
+            s1_logits = s1_logits[:, -1, :]
+            sample_pre = sample_from_logits(
+                s1_logits, temperature=T, top_k=top_k, top_p=top_p, sample_logits=True
+            )
+
+            s2_logits = model.decode_s2(context, sample_pre)
+            s2_logits = s2_logits[:, -1, :]
+            sample_post = sample_from_logits(
+                s2_logits, temperature=T, top_k=top_k, top_p=top_p, sample_logits=True
+            )
+
+            x_token[0] = torch.cat([x_token[0], sample_pre], dim=1)
+            x_token[1] = torch.cat([x_token[1], sample_post], dim=1)
+
+            torch.cuda.empty_cache()
+
+        input_tokens = [t[:, -max_context:].contiguous() for t in x_token]
+        z = tokenizer.decode(input_tokens, half=True)
+        z = z.reshape(batch_size, sample_count, z.size(1), z.size(2))
+        preds = z.cpu().numpy()
+        preds = np.mean(preds, axis=1)
+
+        return preds
+
+
+def calc_time_stamps(x_timestamp):
+    time_df = pd.DataFrame()
+    time_df["minute"] = x_timestamp.dt.minute
+    time_df["hour"] = x_timestamp.dt.hour
+    time_df["weekday"] = x_timestamp.dt.weekday
+    time_df["day"] = x_timestamp.dt.day
+    time_df["month"] = x_timestamp.dt.month
+    return time_df
+
+
+class KronosPredictor:
+
+    def __init__(self, model, tokenizer, device="cuda:0", max_context=512, clip=5):
+        self.tokenizer = tokenizer
+        self.model = model
+        self.max_context = max_context
+        self.clip = clip
+        self.price_cols = ["open", "high", "low", "close"]
+        self.vol_col = "volume"
+        self.amt_vol = "amount"
+        self.time_cols = ["minute", "hour", "weekday", "day", "month"]
+        self.device = device
+
+        self.tokenizer = self.tokenizer.to(self.device)
+        self.model = self.model.to(self.device)
+
+    def generate(
+        self, x, x_stamp, y_stamp, pred_len, T, top_k, top_p, sample_count, verbose
+    ):
+
+        x_tensor = torch.from_numpy(np.array(x).astype(np.float32)).to(self.device)
+        x_stamp_tensor = torch.from_numpy(np.array(x_stamp).astype(np.float32)).to(
+            self.device
+        )
+        y_stamp_tensor = torch.from_numpy(np.array(y_stamp).astype(np.float32)).to(
+            self.device
+        )
+
+        preds = auto_regressive_inference(
+            self.tokenizer,
+            self.model,
+            x_tensor,
+            x_stamp_tensor,
+            y_stamp_tensor,
+            self.max_context,
+            pred_len,
+            self.clip,
+            T,
+            top_k,
+            top_p,
+            sample_count,
+            verbose,
+        )
+        preds = preds[:, -pred_len:, :]
+        return preds
+
+    def predict(
+        self,
+        df,
+        x_timestamp,
+        y_timestamp,
+        pred_len,
+        T=1.0,
+        top_k=0,
+        top_p=0.9,
+        sample_count=1,
+        verbose=True,
+    ):
+
+        if not isinstance(df, pd.DataFrame):
+            raise ValueError("Input must be a pandas DataFrame.")
+
+        if not all(col in df.columns for col in self.price_cols):
+            raise ValueError(f"Price columns {self.price_cols} not found in DataFrame.")
+
+        df = df.copy()
+        if self.vol_col not in df.columns:
+            df[self.vol_col] = 0.0  # Fill missing volume with zeros
+            df[self.amt_vol] = 0.0  # Fill missing amount with zeros
+        if self.amt_vol not in df.columns and self.vol_col in df.columns:
+            df[self.amt_vol] = df[self.vol_col] * df[self.price_cols].mean(axis=1)
+
+        if df[self.price_cols + [self.vol_col, self.amt_vol]].isnull().values.any():
+            raise ValueError(
+                "Input DataFrame contains NaN values in price or volume columns."
+            )
+
+        x_time_df = calc_time_stamps(x_timestamp)
+        y_time_df = calc_time_stamps(y_timestamp)
+
+        x = df[self.price_cols + [self.vol_col, self.amt_vol]].values.astype(np.float32)
+        x_stamp = x_time_df.values.astype(np.float32)
+        y_stamp = y_time_df.values.astype(np.float32)
+
+        x_mean, x_std = np.mean(x, axis=0), np.std(x, axis=0)
+
+        x = (x - x_mean) / (x_std + 1e-5)
+        x = np.clip(x, -self.clip, self.clip)
+
+        x = x[np.newaxis, :]
+        x_stamp = x_stamp[np.newaxis, :]
+        y_stamp = y_stamp[np.newaxis, :]
+
+        preds = self.generate(
+            x, x_stamp, y_stamp, pred_len, T, top_k, top_p, sample_count, verbose
+        )
+
+        preds = preds.squeeze(0)
+        preds = preds * (x_std + 1e-5) + x_mean
+
+        pred_df = pd.DataFrame(
+            preds,
+            columns=self.price_cols + [self.vol_col, self.amt_vol],
+            index=y_timestamp,
+        )
+        return pred_df
+
+    def predict_batch(
+        self,
+        df_list,
+        x_timestamp_list,
+        y_timestamp_list,
+        pred_len,
+        T=1.0,
+        top_k=0,
+        top_p=0.9,
+        sample_count=1,
+        verbose=True,
+    ):
+        """
+        Perform parallel (batch) prediction on multiple time series. All series must have the same historical length and prediction length (pred_len).
+
+        Args:
+            df_list (List[pd.DataFrame]): List of input DataFrames, each containing price columns and optional volume/amount columns.
+            x_timestamp_list (List[pd.DatetimeIndex or Series]): List of timestamps corresponding to historical data, length should match the number of rows in each DataFrame.
+            y_timestamp_list (List[pd.DatetimeIndex or Series]): List of future prediction timestamps, length should equal pred_len.
+            pred_len (int): Number of prediction steps.
+            T (float): Sampling temperature.
+            top_k (int): Top-k filtering threshold.
+            top_p (float): Top-p (nucleus sampling) threshold.
+            sample_count (int): Number of parallel samples per series, automatically averaged internally.
+            verbose (bool): Whether to display autoregressive progress.
+
+        Returns:
+            List[pd.DataFrame]: List of prediction results in the same order as input, each DataFrame contains
+                                `open, high, low, close, volume, amount` columns, indexed by corresponding `y_timestamp`.
+        """
+        # Basic validation
+        if (
+            not isinstance(df_list, (list, tuple))
+            or not isinstance(x_timestamp_list, (list, tuple))
+            or not isinstance(y_timestamp_list, (list, tuple))
+        ):
+            raise ValueError(
+                "df_list, x_timestamp_list, y_timestamp_list must be list or tuple types."
+            )
+        if not (len(df_list) == len(x_timestamp_list) == len(y_timestamp_list)):
+            raise ValueError(
+                "df_list, x_timestamp_list, y_timestamp_list must have consistent lengths."
+            )
+
+        num_series = len(df_list)
+
+        x_list = []
+        x_stamp_list = []
+        y_stamp_list = []
+        means = []
+        stds = []
+        seq_lens = []
+        y_lens = []
+
+        for i in range(num_series):
+            df = df_list[i]
+            if not isinstance(df, pd.DataFrame):
+                raise ValueError(f"Input at index {i} is not a pandas DataFrame.")
+            if not all(col in df.columns for col in self.price_cols):
+                raise ValueError(
+                    f"DataFrame at index {i} is missing price columns {self.price_cols}."
+                )
+
+            df = df.copy()
+            if self.vol_col not in df.columns:
+                df[self.vol_col] = 0.0
+                df[self.amt_vol] = 0.0
+            if self.amt_vol not in df.columns and self.vol_col in df.columns:
+                df[self.amt_vol] = df[self.vol_col] * df[self.price_cols].mean(axis=1)
+
+            if df[self.price_cols + [self.vol_col, self.amt_vol]].isnull().values.any():
+                raise ValueError(
+                    f"DataFrame at index {i} contains NaN values in price or volume columns."
+                )
+
+            x_timestamp = x_timestamp_list[i]
+            y_timestamp = y_timestamp_list[i]
+
+            x_time_df = calc_time_stamps(x_timestamp)
+            y_time_df = calc_time_stamps(y_timestamp)
+
+            x = df[self.price_cols + [self.vol_col, self.amt_vol]].values.astype(
+                np.float32
+            )
+            x_stamp = x_time_df.values.astype(np.float32)
+            y_stamp = y_time_df.values.astype(np.float32)
+
+            if x.shape[0] != x_stamp.shape[0]:
+                raise ValueError(
+                    f"Inconsistent lengths at index {i}: x has {x.shape[0]} vs x_stamp has {x_stamp.shape[0]}."
+                )
+            if y_stamp.shape[0] != pred_len:
+                raise ValueError(
+                    f"y_timestamp length at index {i} should equal pred_len={pred_len}, got {y_stamp.shape[0]}."
+                )
+
+            x_mean, x_std = np.mean(x, axis=0), np.std(x, axis=0)
+            x_norm = (x - x_mean) / (x_std + 1e-5)
+            x_norm = np.clip(x_norm, -self.clip, self.clip)
+
+            x_list.append(x_norm)
+            x_stamp_list.append(x_stamp)
+            y_stamp_list.append(y_stamp)
+            means.append(x_mean)
+            stds.append(x_std)
+
+            seq_lens.append(x_norm.shape[0])
+            y_lens.append(y_stamp.shape[0])
+
+        # Require all series to have consistent historical and prediction lengths for batch processing
+        if len(set(seq_lens)) != 1:
+            raise ValueError(
+                f"Parallel prediction requires all series to have consistent historical lengths, got: {seq_lens}"
+            )
+        if len(set(y_lens)) != 1:
+            raise ValueError(
+                f"Parallel prediction requires all series to have consistent prediction lengths, got: {y_lens}"
+            )
+
+        x_batch = np.stack(x_list, axis=0).astype(np.float32)  # (B, seq_len, feat)
+        x_stamp_batch = np.stack(x_stamp_list, axis=0).astype(
+            np.float32
+        )  # (B, seq_len, time_feat)
+        y_stamp_batch = np.stack(y_stamp_list, axis=0).astype(
+            np.float32
+        )  # (B, pred_len, time_feat)
+
+        preds = self.generate(
+            x_batch,
+            x_stamp_batch,
+            y_stamp_batch,
+            pred_len,
+            T,
+            top_k,
+            top_p,
+            sample_count,
+            verbose,
+        )
+        # preds: (B, pred_len, feat)
+
+        pred_dfs = []
+        for i in range(num_series):
+            preds_i = preds[i] * (stds[i] + 1e-5) + means[i]
+            pred_df = pd.DataFrame(
+                preds_i,
+                columns=self.price_cols + [self.vol_col, self.amt_vol],
+                index=y_timestamp_list[i],
+            )
+            pred_dfs.append(pred_df)
+
+        return pred_dfs

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b082dfcbd8e8c142a725c8bbb99781802f38fec81210e13479effb32b3c3e020
+size 98980656

module.py

@@ -0,0 +1,581 @@
+import math
+
+from einops import rearrange, reduce
+import torch
+import torch.nn as nn
+from torch.autograd import Function
+import torch.nn.functional as F
+
+
+class DifferentiableEntropyFunction(Function):
+    @staticmethod
+    def forward(ctx, zq, basis, K, eps):
+        zb = (zq + 1) / 2
+        zi = ((zb * basis).sum(-1)).to(torch.int64)
+        cnt = torch.scatter_reduce(torch.zeros(2 ** K, device=zq.device, dtype=zq.dtype),
+                                   0,
+                                   zi.flatten(),
+                                   torch.ones_like(zi.flatten()).to(zq.dtype),
+                                   'sum')
+        prob = (cnt + eps) / (cnt + eps).sum()
+        H = -(prob * torch.log(prob)).sum()
+        ctx.save_for_backward(zq, zi, prob)
+        ctx.K = K
+        return H
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        zq, zi, prob = ctx.saved_tensors
+        grad_array = -grad_output * (torch.log(prob) + 1) / zi.numel() / ctx.K
+        reord_grad = grad_array[zi.flatten()].reshape(zi.shape)
+        grad_input = reord_grad.unsqueeze(-1) * zq
+        return grad_input, None, None, None, None
+
+
+def codebook_entropy(zq, basis, K, eps=1e-4):
+    return DifferentiableEntropyFunction.apply(zq, basis, K, eps)
+
+
+class BinarySphericalQuantizer(nn.Module):
+    def __init__(self, embed_dim, beta, gamma0, gamma, zeta,
+                 input_format='bchw',
+                 soft_entropy=True, group_size=9,
+                 persample_entropy_compute='analytical',
+                 cb_entropy_compute='group',
+                 l2_norm=True,
+                 inv_temperature=1):
+        """
+        Paper link: https://arxiv.org/pdf/2406.07548.pdf
+        Here we use the official implementation of the BinarySphericalQuantizer.
+        """
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.beta = beta  # loss weight for commit loss
+        self.gamma0 = gamma0  # loss weight for entropy penalty
+        self.gamma = gamma  # loss weight for entropy penalty
+        self.zeta = zeta  # loss weight for entire entropy penalty
+        self.input_format = input_format
+        assert self.embed_dim % group_size == 0, "embed_dim must be divisible by group_size"
+        self.num_groups = self.embed_dim // group_size
+        self.group_size = group_size
+        assert persample_entropy_compute in ['group', 'analytical'], "persample_entropy_compute must be either 'group' or 'analytical'"
+        assert cb_entropy_compute in ['group', 'nce'], "cb_entropy_compute must be either 'group' or 'nce'"
+        self.persample_entropy_compute = persample_entropy_compute
+        self.cb_entropy_compute = cb_entropy_compute
+        self.l2_norm = l2_norm
+        self.inv_temperature = inv_temperature
+
+        self.register_buffer('basis', 2 ** torch.arange(embed_dim - 1, -1, -1))
+        self.register_buffer('group_basis', 2 ** torch.arange(group_size - 1, -1, -1))
+
+        self.num_dimensions = 2 ** embed_dim
+        self.bits_per_index = embed_dim
+
+        # we only need to keep the codebook portion up to the group size
+        # because we approximate the H loss with this subcode
+        group_codes = torch.arange(2 ** self.group_size)
+        group_codebook = self.indexes_to_codes(group_codes).float()[:, -group_size:]
+        self.register_buffer('group_codebook', group_codebook, persistent=False)
+
+        self.soft_entropy = soft_entropy  # soft_entropy: Sec 3.2 of https://arxiv.org/pdf/1911.05894.pdf
+
+    def quantize(self, z):
+        assert z.shape[-1] == self.embed_dim, f"Expected {self.embed_dim} dimensions, got {z.shape[-1]}"
+
+        zhat = torch.where(z > 0,
+                           torch.tensor(1, dtype=z.dtype, device=z.device),
+                           torch.tensor(-1, dtype=z.dtype, device=z.device))
+        return z + (zhat - z).detach()
+
+    def forward(self, z):
+        # if self.input_format == 'bchw':
+        #     z = rearrange(z, 'b c h w -> b h w c')
+        zq = self.quantize(z)
+
+        indices = self.codes_to_indexes(zq.detach())
+        group_indices = self.codes_to_group_indexes(zq.detach())
+        if not self.training:
+            used_codes = torch.unique(indices, return_counts=False)
+        else:
+            used_codes = None
+
+        q_scale = 1. / (self.embed_dim ** 0.5) if self.l2_norm else 1.
+
+        if self.soft_entropy:
+            persample_entropy, cb_entropy, avg_prob = self.soft_entropy_loss(z)
+            entropy_penalty = self.gamma0 * persample_entropy - self.gamma * cb_entropy
+        else:
+            zb_by_sample = ((zq + 1) / 2).reshape(z.shape[0], -1, z.shape[-1]).to(torch.float32)
+            persample_entropy = self.get_hard_per_sample_entropy(zb_by_sample)
+            cb_entropy = codebook_entropy(zq, self.basis, self.embed_dim)
+            entropy_penalty = self.gamma0 * persample_entropy - self.gamma * cb_entropy
+
+        zq = zq * q_scale
+
+        # commit loss
+        commit_loss = self.beta * torch.mean(((zq.detach() - z) ** 2).sum(dim=-1))
+
+        # if self.input_format == 'bchw':
+        #     zq = rearrange(zq, 'b h w c -> b c h w')
+
+        return (
+            zq,
+            commit_loss + self.zeta * entropy_penalty / self.inv_temperature,
+            {"H": cb_entropy, "used_codes": used_codes, "indices": indices, "group_indices": group_indices,
+             "avg_prob": avg_prob}
+        )
+
+    def soft_entropy_loss(self, z):
+        # if we divide the code in subgroups of size group_size, the codebook will be of size 2 ** group_size
+        # the sub-code is the last group_size bits of the full code
+        group_code_book = self.group_codebook / (self.embed_dim ** 0.5 if self.l2_norm else 1)
+        divided_z = rearrange(z, '... (g c) -> ... g c', c=self.group_size)
+
+        # we calculate the distance between the divided_z and the codebook for each subgroup
+        distance = - 2 * torch.einsum('... g c, d c ->... g d', divided_z, group_code_book)
+        prob = (-distance * self.inv_temperature).softmax(dim=-1)
+        if self.persample_entropy_compute == 'analytical':
+            if self.l2_norm:
+                p = torch.sigmoid(-4 * z / (self.embed_dim ** 0.5) * self.inv_temperature)
+            else:
+                p = torch.sigmoid(-4 * z * self.inv_temperature)
+            prob = torch.stack([p, 1 - p], dim=-1)
+            per_sample_entropy = self.get_entropy(prob, dim=-1, normalize=False).sum(dim=-1).mean()
+        else:
+            per_sample_entropy = self.get_entropy(prob, dim=-1, normalize=False).sum(dim=-1).mean()
+
+        # macro average of the probability of each subgroup
+        avg_prob = reduce(prob, '... g d ->g d', 'mean')
+        codebook_entropy = self.get_entropy(avg_prob, dim=-1, normalize=False)
+
+        # the approximation of the entropy is the sum of the entropy of each subgroup
+        return per_sample_entropy, codebook_entropy.sum(), avg_prob
+
+    def get_hard_per_sample_entropy(self, zb_by_sample):
+        probs_per_dim = zb_by_sample.sum(1) / zb_by_sample.shape[1]
+        persample_entropy = - probs_per_dim * torch.log(probs_per_dim + 1e-8) - (1 - probs_per_dim) * torch.log(1 - probs_per_dim + 1e-8)
+        persample_entropy = persample_entropy.sum(-1)
+        return persample_entropy.mean()
+
+    def codes_to_indexes(self, zhat):
+        """Converts a `code` to an index in the codebook.
+        Args:
+            zhat: A tensor of shape (B, ..., C) containing the codes. must be in {-1, 1}
+        """
+        assert zhat.shape[-1] == self.embed_dim, f"Expected {self.embed_dim} dimensions, got {zhat.shape[-1]}"
+        return ((zhat + 1) / 2 * self.basis).sum(axis=-1).to(torch.int64)
+
+    def codes_to_group_indexes(self, zhat):
+        """Converts a `code` to a list of indexes (in groups) in the codebook.
+        Args:
+            zhat: A tensor of shape (B, ..., C) containing the codes. must be in {-1, 1}
+        """
+        zhat_in_group = rearrange(zhat, 'b ... (g c) -> b ... g c', c=self.group_size)
+        return ((zhat_in_group + 1) / 2 * self.group_basis).sum(axis=-1).to(torch.int64)
+
+    def indexes_to_codes(self, indices):
+        """Inverse of `indexes_to_codes`."""
+        indices = indices.unsqueeze(-1)
+        codes_non_centered = torch.remainder(
+            torch.floor_divide(indices, self.basis), 2
+        )
+        return codes_non_centered * 2 - 1
+
+    def group_indexes_to_codes(self, group_indices):
+        """Inverse of `group_indexes_to_codes`."""
+        group_indices = group_indices.unsqueeze(-1)
+        codes_non_centered = torch.remainder(
+            torch.floor_divide(group_indices, self.group_basis), 2
+        )
+        codes_non_centered = rearrange(codes_non_centered, 'b ... g c -> b ... (g c)')
+        return codes_non_centered * 2 - 1
+
+    def get_entropy(self, count, dim=-1, eps=1e-4, normalize=True):
+        if normalize:
+            probs = (count + eps) / (count + eps).sum(dim=dim, keepdim=True)
+        else:
+            probs = count
+        H = -(probs * torch.log(probs + 1e-8)).sum(dim=dim)
+        return H
+
+    def get_group_codebook_entry(self, group_indices):
+        z_q = self.group_indexes_to_codes(group_indices)
+        q_scale = 1. / (self.embed_dim ** 0.5) if self.l2_norm else 1.
+        z_q = z_q * q_scale
+        if self.input_format == 'bchw':
+            h, w = int(z_q.shape[1] ** 0.5)
+            assert h * w == z_q.shape[1], 'Invalid sequence length'
+            z_q = rearrange(z_q, 'b (h w) c -> b c h w', h=h)
+        return z_q
+
+    def get_codebook_entry(self, indices):
+        z_q = self.indexes_to_codes(indices)
+        q_scale = 1. / (self.embed_dim ** 0.5) if self.l2_norm else 1.
+        z_q = z_q * q_scale
+        if self.input_format == 'bchw':
+            h, w = int(z_q.shape[1] ** 0.5)
+            assert h * w == z_q.shape[1], 'Invalid sequence length'
+            z_q = rearrange(z_q, 'b (h w) c -> b c h w', h=h)
+        return z_q
+
+
+class BSQuantizer(nn.Module):
+
+    def __init__(self, s1_bits, s2_bits, beta, gamma0, gamma, zeta, group_size):
+        super().__init__()
+        self.codebook_dim = s1_bits + s2_bits
+        self.s1_bits = s1_bits
+        self.s2_bits = s2_bits
+        self.bsq = BinarySphericalQuantizer(self.codebook_dim, beta, gamma0, gamma, zeta, group_size=group_size)
+
+    def bits_to_indices(self, bits):
+        bits = (bits >= 0).to(torch.long)
+        indices = 2 ** torch.arange(
+            0,
+            bits.shape[-1],
+            1,
+            dtype=torch.long,
+            device=bits.device,
+        )
+        return (bits * indices).sum(-1)
+
+    def forward(self, z, half=False):
+        z = F.normalize(z, dim=-1)
+        quantized, bsq_loss, metrics = self.bsq(z)
+        if half:
+            q_pre = quantized[:, :, :self.s1_bits]
+            q_post = quantized[:, :, self.s1_bits:]
+            z_indices = [self.bits_to_indices(q_pre), self.bits_to_indices(q_post)]
+        else:
+            z_indices = self.bits_to_indices(quantized)
+        return bsq_loss, quantized, z_indices
+
+
+class RMSNorm(torch.nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-5):
+        super().__init__()
+        self.eps = eps
+        self.weight = nn.Parameter(torch.ones(dim))
+
+    def _norm(self, x):
+        return x * torch.rsqrt(torch.mean(x * x, dim=-1, keepdim=True) + self.eps)
+
+    def forward(self, x):
+        output = self._norm(x.float()).type_as(x)
+        return output * self.weight
+
+
+class FeedForward(nn.Module):
+    def __init__(self, d_model, ff_dim, ffn_dropout_p=0.0):
+        super().__init__()
+
+        self.w1 = nn.Linear(d_model, ff_dim, bias=False)
+        self.w3 = nn.Linear(d_model, ff_dim, bias=False)
+        self.w2 = nn.Linear(ff_dim, d_model, bias=False)
+        self.ffn_dropout = nn.Dropout(ffn_dropout_p)
+
+    def forward(self, x):
+        return self.ffn_dropout(self.w2(F.silu(self.w1(x)) * self.w3(x)))
+
+
+class RotaryPositionalEmbedding(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim))
+        self.register_buffer("inv_freq", inv_freq)
+        self.seq_len_cached = None
+        self.cos_cached = None
+        self.sin_cached = None
+
+    def _update_cos_sin_cache(self, x, seq_len):
+        if seq_len != self.seq_len_cached:
+            self.seq_len_cached = seq_len
+            t = torch.arange(seq_len, device=x.device).type_as(self.inv_freq)
+            freqs = torch.einsum('i,j->ij', t, self.inv_freq)
+            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
+            self.cos_cached = emb.cos()[None, None, :, :]
+            self.sin_cached = emb.sin()[None, None, :, :]
+        return self.cos_cached, self.sin_cached
+
+    def forward(self, q, k):
+        cos, sin = self._update_cos_sin_cache(q, q.shape[-2])
+        return (
+            (q * cos) + (self._rotate_half(q) * sin),
+            (k * cos) + (self._rotate_half(k) * sin),
+        )
+
+    def _rotate_half(self, x):
+        x1, x2 = x.chunk(2, dim=-1)
+        return torch.cat((-x2, x1), dim=-1)
+
+
+def scaled_dot_product_attention(query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False, scale=None, training=True) -> torch.Tensor:
+    L, S = query.size(-2), key.size(-2)
+    scale_factor = 1 / math.sqrt(query.size(-1)) if scale is None else scale
+    attn_bias = torch.zeros(L, S, dtype=query.dtype).to(query.device)
+
+    if is_causal:
+        assert attn_mask is None
+        temp_mask = torch.ones(L, S, dtype=torch.bool).tril(diagonal=0).to(query.device)
+        attn_bias.masked_fill_(temp_mask.logical_not(), float("-inf"))
+        attn_bias.to(query.dtype)
+
+    attn_weight = query @ key.transpose(-2, -1) * scale_factor
+    attn_weight += attn_bias
+
+    if attn_mask is not None:
+        attn_mask_bias = torch.zeros_like(attn_weight)
+        if attn_mask.dtype == torch.bool:
+            attn_mask_bias.masked_fill_(attn_mask, float("-inf"))
+        else:
+            attn_mask_bias += attn_mask
+        attn_weight += attn_mask_bias
+
+    attn_weight = torch.softmax(attn_weight, dim=-1)
+    attn_weight = torch.dropout(attn_weight, dropout_p, train=training)
+    return attn_weight @ value
+
+
+class MultiHeadAttentionWithRoPE(nn.Module):
+    def __init__(self, d_model, n_heads, attn_dropout_p=0.0, resid_dropout_p=0.0):
+        super().__init__()
+        self.d_model = d_model
+        self.n_heads = n_heads
+        self.head_dim = d_model // n_heads
+
+        self.q_proj = nn.Linear(d_model, d_model)
+        self.k_proj = nn.Linear(d_model, d_model)
+        self.v_proj = nn.Linear(d_model, d_model)
+        self.out_proj = nn.Linear(d_model, d_model)
+        self.rotary = RotaryPositionalEmbedding(self.head_dim)
+        self.attn_dropout_p = attn_dropout_p
+        self.resid_dropout = nn.Dropout(resid_dropout_p)
+
+    def forward(self, x, key_padding_mask=None):
+        batch_size, seq_len, _ = x.shape
+
+        q = self.q_proj(x).view(batch_size, seq_len, self.n_heads, self.head_dim).transpose(1, 2)
+        k = self.k_proj(x).view(batch_size, seq_len, self.n_heads, self.head_dim).transpose(1, 2)
+        v = self.v_proj(x).view(batch_size, seq_len, self.n_heads, self.head_dim).transpose(1, 2)
+
+        q, k = self.rotary(q, k)
+
+        if key_padding_mask is not None:
+            attn_mask = key_padding_mask.unsqueeze(1).unsqueeze(2)  # [batch, 1, 1, seq_len]
+            attn_mask = attn_mask.expand(-1, self.n_heads, seq_len, -1)  # [batch, n_heads, q_len, k_len]
+        else:
+            attn_mask = None
+
+        attn_output = scaled_dot_product_attention(
+            q, k, v,
+            attn_mask=attn_mask,
+            dropout_p=self.attn_dropout_p,
+            is_causal=True,
+            training=self.training
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous().view(batch_size, seq_len, self.d_model)
+        return self.resid_dropout(self.out_proj(attn_output))
+
+
+class MultiHeadCrossAttentionWithRoPE(nn.Module):
+    def __init__(self, d_model, n_heads, attn_dropout_p=0.0, resid_dropout=0.0):
+        super().__init__()
+        self.d_model = d_model
+        self.n_heads = n_heads
+        self.head_dim = d_model // n_heads
+
+        self.q_proj = nn.Linear(d_model, d_model)
+        self.k_proj = nn.Linear(d_model, d_model)
+        self.v_proj = nn.Linear(d_model, d_model)
+        self.out_proj = nn.Linear(d_model, d_model)
+        self.rotary = RotaryPositionalEmbedding(self.head_dim)
+        self.attn_dropout_p = attn_dropout_p
+        self.resid_dropout = nn.Dropout(resid_dropout)
+
+    def forward(self, query, key, value, key_padding_mask=None):
+        batch_size, q_len, _ = query.shape
+        _, seq_len, _ = key.shape
+
+        q = self.q_proj(query).view(batch_size, q_len, self.n_heads, self.head_dim).transpose(1, 2)
+        k = self.k_proj(key).view(batch_size, seq_len, self.n_heads, self.head_dim).transpose(1, 2)
+        v = self.v_proj(value).view(batch_size, seq_len, self.n_heads, self.head_dim).transpose(1, 2)
+
+        q, k = self.rotary(q, k)
+
+        if key_padding_mask is not None:
+            attn_mask = key_padding_mask.unsqueeze(1).unsqueeze(2)
+            attn_mask = attn_mask.expand(-1, self.n_heads, q_len, -1)
+        else:
+            attn_mask = None
+
+        is_causal_flag = self.training
+
+        attn_output = scaled_dot_product_attention(
+            q, k, v,
+            attn_mask=attn_mask,
+            dropout_p=self.attn_dropout_p,
+            is_causal=is_causal_flag,
+            training=self.training
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous().view(batch_size, q_len, self.d_model)
+        return self.resid_dropout(self.out_proj(attn_output))
+
+
+class HierarchicalEmbedding(nn.Module):
+    def __init__(self, s1_bits, s2_bits, d_model=256):
+        super().__init__()
+        self.s1_bits = s1_bits
+        self.s2_bits = s2_bits
+
+        vocab_s1 = 2 ** s1_bits
+        vocab_s2 = 2 ** s2_bits
+
+        self.emb_s1 = nn.Embedding(vocab_s1, d_model)
+        self.emb_s2 = nn.Embedding(vocab_s2, d_model)
+        self.d_model = d_model
+        self.fusion_proj = nn.Linear(d_model * 2, d_model)
+
+        nn.init.normal_(self.emb_s1.weight, mean=0, std=d_model ** -0.5)
+        nn.init.normal_(self.emb_s2.weight, mean=0, std=d_model ** -0.5)
+
+    def forward(self, token_ids):
+        """Inputs:
+        token_ids: [batch_size, seq_len] token ID
+        Output: [batch_size, seq_len, d_model]
+        """
+        if isinstance(token_ids, tuple) or isinstance(token_ids, list):
+            s1_ids, s2_ids = token_ids
+        else:
+            s1_ids, s2_ids = self.split_token(token_ids, self.s2_bits)
+        s1_emb = self.emb_s1(s1_ids) * math.sqrt(self.d_model)
+        s2_emb = self.emb_s2(s2_ids) * math.sqrt(self.d_model)
+        return self.fusion_proj(torch.cat([s1_emb, s2_emb], dim=-1))
+
+
+class DependencyAwareLayer(nn.Module):
+    def __init__(self, d_model, n_heads=4, attn_dropout_p=0.0, resid_dropout=0.0):
+        super().__init__()
+        self.cross_attn = MultiHeadCrossAttentionWithRoPE(d_model, n_heads, attn_dropout_p, resid_dropout)
+        self.norm = RMSNorm(d_model)
+
+    def forward(self, hidden_states, sibling_embed, key_padding_mask=None):
+        """hidden_states: [batch, seq_len, d_model]
+        sibling_embed: Embedding from another subtoken
+        """
+        attn_out = self.cross_attn(
+            query=sibling_embed,
+            key=hidden_states,
+            value=hidden_states,
+            key_padding_mask=key_padding_mask
+        )
+        return self.norm(hidden_states + attn_out)
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, d_model, n_heads, ff_dim=1024, ffn_dropout_p=0.0, attn_dropout_p=0.0, resid_dropout_p=0.0):
+        super().__init__()
+        self.norm1 = RMSNorm(d_model)
+        self.self_attn = MultiHeadAttentionWithRoPE(d_model, n_heads, attn_dropout_p, resid_dropout_p)
+        self.norm2 = RMSNorm(d_model)
+        self.ffn = FeedForward(d_model, ff_dim, ffn_dropout_p)
+
+    def forward(self, x, key_padding_mask=None):
+        residual = x
+        x = self.norm1(x)
+        attn_out = self.self_attn(x, key_padding_mask=key_padding_mask)
+        x = residual + attn_out
+
+        residual = x
+        x = self.norm2(x)
+        ffn_out = self.ffn(x)
+        x = residual + ffn_out
+        return x
+
+
+class DualHead(nn.Module):
+    def __init__(self, s1_bits, s2_bits, d_model):
+        super().__init__()
+        self.vocab_s1 = 2 ** s1_bits
+        self.vocab_s2 = 2 ** s2_bits
+        self.proj_s1 = nn.Linear(d_model, self.vocab_s1)
+        self.proj_s2 = nn.Linear(d_model, self.vocab_s2)
+
+    def compute_loss(self, s1_logits, s2_logits, s1_targets, s2_targets, padding_mask=None):
+        if padding_mask is not None:
+            valid_mask = (padding_mask == 0)
+            s1_logits = s1_logits[valid_mask]
+            s2_logits = s2_logits[valid_mask]
+            s1_targets = s1_targets[valid_mask]
+            s2_targets = s2_targets[valid_mask]
+            ce_s1 = F.cross_entropy(s1_logits, s1_targets)
+            ce_s2 = F.cross_entropy(s2_logits, s2_targets)
+        else:
+            ce_s1 = F.cross_entropy(s1_logits.reshape(-1, self.vocab_s1), s1_targets.reshape(-1))
+            ce_s2 = F.cross_entropy(s2_logits.reshape(-1, self.vocab_s2), s2_targets.reshape(-1))
+        ce_loss = (ce_s1 + ce_s2) / 2
+        return ce_loss, ce_s1, ce_s2
+
+    def forward(self, x):
+        return self.proj_s1(x)
+
+    def cond_forward(self, x2):
+        return self.proj_s2(x2)
+
+
+class FixedEmbedding(nn.Module):
+    def __init__(self, c_in, d_model):
+        super(FixedEmbedding, self).__init__()
+
+        w = torch.zeros(c_in, d_model).float()
+        w.require_grad = False
+
+        position = torch.arange(0, c_in).float().unsqueeze(1)
+        div_term = (torch.arange(0, d_model, 2).float() * -(math.log(10000.0) / d_model)).exp()
+
+        w[:, 0::2] = torch.sin(position * div_term)
+        w[:, 1::2] = torch.cos(position * div_term)
+
+        self.emb = nn.Embedding(c_in, d_model)
+        self.emb.weight = nn.Parameter(w, requires_grad=False)
+
+    def forward(self, x):
+        return self.emb(x).detach()
+
+
+class TemporalEmbedding(nn.Module):
+    def __init__(self, d_model, learn_pe):
+        super(TemporalEmbedding, self).__init__()
+
+        minute_size = 60
+        hour_size = 24
+        weekday_size = 7
+        day_size = 32
+        month_size = 13
+
+        Embed = FixedEmbedding if not learn_pe else nn.Embedding
+        self.minute_embed = Embed(minute_size, d_model)
+        self.hour_embed = Embed(hour_size, d_model)
+        self.weekday_embed = Embed(weekday_size, d_model)
+        self.day_embed = Embed(day_size, d_model)
+        self.month_embed = Embed(month_size, d_model)
+
+    def forward(self, x):
+        x = x.long()
+
+        minute_x = self.minute_embed(x[:, :, 0])
+        hour_x = self.hour_embed(x[:, :, 1])
+        weekday_x = self.weekday_embed(x[:, :, 2])
+        day_x = self.day_embed(x[:, :, 3])
+        month_x = self.month_embed(x[:, :, 4])
+
+        return hour_x + weekday_x + day_x + month_x + minute_x
+
+
+
+
+
+
+
+

requirements.txt

@@ -0,0 +1,6 @@
+torch
+pandas
+numpy
+safetensors
+huggingface_hub
+