Upload folder using huggingface_hub

Dockerfile

@@ -0,0 +1,16 @@
+FROM python:3.11-slim
+
+WORKDIR /app
+
+# HF Spaces runs as a non-root user (uid 1000); cache HF downloads in a writable dir.
+ENV HF_HOME=/app/.cache/huggingface \
+    PYTHONUNBUFFERED=1
+
+COPY requirements.txt .
+RUN pip install --no-cache-dir -r requirements.txt
+
+COPY . .
+RUN mkdir -p /app/.cache && chmod -R 777 /app/.cache
+
+EXPOSE 7860
+CMD ["python", "app.py"]

README.md

@@ -1,10 +1,25 @@
 ---
-title: Kronos Us
-emoji: 🐨
-colorFrom: purple
-colorTo: indigo
+title: Kronos US Forecast
+emoji: 📈
+colorFrom: red
+colorTo: gray
 sdk: docker
+app_port: 7860
 pinned: false
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Kronos · US-stock Forecast Dashboard
+
+Interactive CPU dashboard for the [Kronos](https://github.com/shiyu-coder/Kronos)
+financial time-series model, applied to US stocks (data via yfinance).
+
+Pick a ticker and sampling parameters → the server draws an ensemble of sample
+paths from Kronos-small → the page shows the **median forecast** with a **10–90%
+uncertainty band**, plus accuracy metrics in `backtest` mode.
+
+- **Model / prediction**: 100% original Kronos (`model/`, unmodified).
+- **This app**: US data adapter, hold-out accuracy backtest, ensemble median+band,
+  metrics, and the interactive UI.
+
+⚠️ Runs on CPU — each sample path is ~1–2s, so an ensemble takes tens of seconds.
+The UI shows a spinner and elapsed time. `n_paths` is capped at 30 for hosting.

app.py

@@ -0,0 +1,62 @@
+"""
+Hosted US-stock Kronos dashboard for HuggingFace Spaces (Docker SDK).
+
+Same Flask app as ../dashboard.py, adapted for hosting:
+  * binds 0.0.0.0:7860 (the port HF Spaces expects),
+  * pre-loads the model at startup so the first request isn't slow,
+  * a lower default n_paths keeps a request comfortably under proxy timeouts.
+"""
+import os
+
+from flask import Flask, jsonify, render_template, request
+
+import engine
+
+app = Flask(__name__, template_folder="templates")
+
+# Ceiling on ensemble size so a single hosted request can't run for minutes.
+MAX_PATHS = int(os.environ.get("MAX_PATHS", "30"))
+
+
+@app.route("/")
+def index():
+    return render_template("dashboard.html")
+
+
+@app.route("/api/health")
+def health():
+    return jsonify({"ok": True, "model_loaded": engine.is_loaded()})
+
+
+@app.route("/api/forecast", methods=["POST"])
+def forecast():
+    p = request.get_json(force=True) or {}
+    try:
+        n_paths = min(int(p.get("n_paths", 20)), MAX_PATHS)
+        result = engine.run_ensemble(
+            ticker=str(p.get("ticker", "AAPL")).upper().strip(),
+            interval=str(p.get("interval", "1d")),
+            period=str(p.get("period", "3y")),
+            lookback=int(p.get("lookback", 400)),
+            pred_len=int(p.get("pred_len", 20)),
+            mode=str(p.get("mode", "backtest")),
+            n_paths=n_paths,
+            T=float(p.get("T", 0.7)),
+            top_p=float(p.get("top_p", 0.9)),
+            top_k=int(p.get("top_k", 0)),
+            seed=int(p.get("seed", 123)),
+        )
+        return jsonify({"ok": True, "result": result})
+    except Exception as exc:
+        return jsonify({"ok": False, "error": str(exc)}), 400
+
+
+if __name__ == "__main__":
+    # Warm the model before serving so the first user request is fast.
+    print("[startup] pre-loading Kronos model ...", flush=True)
+    try:
+        engine.get_predictor()
+        print("[startup] model ready.", flush=True)
+    except Exception as exc:
+        print(f"[startup] model preload failed (will retry on first request): {exc}", flush=True)
+    app.run(host="0.0.0.0", port=7860, threaded=True)

engine.py

@@ -0,0 +1,117 @@
+"""
+Ensemble inference engine for the hosted dashboard.
+Same logic as ../engine.py but with flat imports (Space repo is self-contained).
+"""
+import time
+
+import numpy as np
+
+from model import Kronos, KronosTokenizer, KronosPredictor
+from predict_us import fetch_ohlcv, future_timestamps, report_metrics
+
+_COLS = ["open", "high", "low", "close", "volume"]
+
+_PREDICTOR = None
+_PREDICTOR_KEY = None
+
+
+def get_predictor(device: str = "cpu", max_context: int = 512,
+                  model_id: str = "NeoQuasar/Kronos-small",
+                  tokenizer_id: str = "NeoQuasar/Kronos-Tokenizer-base"):
+    global _PREDICTOR, _PREDICTOR_KEY
+    key = (device, max_context, model_id, tokenizer_id)
+    if _PREDICTOR is None or _PREDICTOR_KEY != key:
+        tokenizer = KronosTokenizer.from_pretrained(tokenizer_id)
+        model = Kronos.from_pretrained(model_id)
+        tokenizer.eval()
+        model.eval()
+        _PREDICTOR = KronosPredictor(model, tokenizer, device=device, max_context=max_context)
+        _PREDICTOR_KEY = key
+    return _PREDICTOR
+
+
+def is_loaded() -> bool:
+    return _PREDICTOR is not None
+
+
+def run_ensemble(ticker: str = "AAPL", interval: str = "1d", period: str = "3y",
+                 lookback: int = 400, pred_len: int = 20, mode: str = "backtest",
+                 n_paths: int = 20, T: float = 0.7, top_p: float = 0.9, top_k: int = 0,
+                 seed: int = 123, device: str = "cpu", ctx_tail: int = 120) -> dict:
+    import random
+    import torch
+
+    random.seed(seed); np.random.seed(seed); torch.manual_seed(seed)
+
+    df = fetch_ohlcv(ticker, interval, period)
+    need = lookback + (pred_len if mode == "backtest" else 0)
+    if len(df) < need:
+        raise ValueError(
+            f"Not enough bars for {ticker}: have {len(df)}, need {need}. "
+            f"Lower lookback/pred_len or raise period."
+        )
+
+    if mode == "backtest":
+        hist = df.iloc[-(lookback + pred_len):-pred_len]
+        fut = df.iloc[-pred_len:]
+        y_ts = fut["timestamps"].reset_index(drop=True)
+    else:
+        hist = df.iloc[-lookback:]
+        fut = None
+        y_ts = future_timestamps(hist["timestamps"].iloc[-1], interval, pred_len)
+
+    x_df = hist[_COLS].reset_index(drop=True)
+    x_ts = hist["timestamps"].reset_index(drop=True)
+
+    predictor = get_predictor(device=device)
+
+    t0 = time.time()
+    paths = []
+    for i in range(n_paths):
+        torch.manual_seed(seed + i); np.random.seed(seed + i)
+        pred_df = predictor.predict(
+            df=x_df, x_timestamp=x_ts, y_timestamp=y_ts, pred_len=pred_len,
+            T=T, top_k=top_k, top_p=top_p, sample_count=1, verbose=False,
+        )
+        paths.append(pred_df["close"].values.astype(float))
+    paths = np.stack(paths, axis=0)
+    elapsed = time.time() - t0
+
+    median = np.median(paths, axis=0)
+    lo = np.quantile(paths, 0.1, axis=0)
+    hi = np.quantile(paths, 0.9, axis=0)
+
+    def _iso(ts_series):
+        return [pd_ts.isoformat() for pd_ts in ts_series]
+
+    ctx = hist.iloc[-ctx_tail:]
+    out = {
+        "meta": {
+            "ticker": ticker, "interval": interval, "period": period, "mode": mode,
+            "lookback": lookback, "pred_len": pred_len, "n_paths": n_paths,
+            "T": T, "top_p": top_p, "top_k": top_k, "elapsed": round(elapsed, 1),
+        },
+        "history": {
+            "t": _iso(ctx["timestamps"]),
+            "close": [float(v) for v in ctx["close"].values],
+        },
+        "forecast": {
+            "t": _iso(y_ts),
+            "median": [float(v) for v in median],
+            "lo": [float(v) for v in lo],
+            "hi": [float(v) for v in hi],
+            "paths": [[float(v) for v in p] for p in paths],
+        },
+        "truth": None,
+        "metrics": None,
+    }
+
+    if mode == "backtest":
+        truth = fut["close"].values.astype(float)
+        out["truth"] = {"t": _iso(fut["timestamps"]), "close": [float(v) for v in truth]}
+        m = report_metrics(truth, median)
+        cover = float(np.mean((truth >= lo) & (truth <= hi)) * 100)
+        m["coverage%"] = cover
+        out["metrics"] = {k: round(v, 3) for k, v in m.items()}
+
+    return out

model/__init__.py

@@ -0,0 +1,17 @@
+from .kronos import KronosTokenizer, Kronos, KronosPredictor
+
+model_dict = {
+    'kronos_tokenizer': KronosTokenizer,
+    'kronos': Kronos,
+    'kronos_predictor': KronosPredictor
+}
+
+
+def get_model_class(model_name):
+    if model_name in model_dict:
+        return model_dict[model_name]
+    else:
+        print(f"Model {model_name} not found in model_dict")
+        raise NotImplementedError
+
+

model/kronos.py

@@ -0,0 +1,662 @@
+import numpy as np
+import pandas as pd
+import torch
+from huggingface_hub import PyTorchModelHubMixin
+import sys
+
+from tqdm import trange
+
+sys.path.append("../")
+from model.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. / (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=half, collect_metrics=False)
+        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.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 = torch.topk(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():
+        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)
+        
+        initial_seq_len = x.size(1)
+        batch_size = x_token[0].size(0)
+        total_seq_len = initial_seq_len + pred_len
+        full_stamp = torch.cat([x_stamp, y_stamp], dim=1)
+
+        generated_pre = x_token[0].new_empty(batch_size, pred_len)
+        generated_post = x_token[1].new_empty(batch_size, pred_len)
+
+        pre_buffer = x_token[0].new_zeros(batch_size, max_context)
+        post_buffer = x_token[1].new_zeros(batch_size, max_context)
+        buffer_len = min(initial_seq_len, max_context)
+        if buffer_len > 0:
+            start_idx = max(0, initial_seq_len - max_context)
+            pre_buffer[:, :buffer_len] = x_token[0][:, start_idx:start_idx + buffer_len]
+            post_buffer[:, :buffer_len] = x_token[1][:, start_idx:start_idx + buffer_len]
+
+        if verbose:
+            ran = trange
+        else:
+            ran = range
+        for i in ran(pred_len):
+            current_seq_len = initial_seq_len + i
+            window_len = min(current_seq_len, max_context)
+
+            if current_seq_len <= max_context:
+                input_tokens = [
+                    pre_buffer[:, :window_len],
+                    post_buffer[:, :window_len]
+                ]
+            else:
+                input_tokens = [pre_buffer, post_buffer]
+
+            context_end = current_seq_len
+            context_start = max(0, context_end - max_context)
+            current_stamp = full_stamp[:, context_start:context_end, :].contiguous()
+
+            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)
+
+            generated_pre[:, i] = sample_pre.squeeze(-1)
+            generated_post[:, i] = sample_post.squeeze(-1)
+
+            if current_seq_len < max_context:
+                pre_buffer[:, current_seq_len] = sample_pre.squeeze(-1)
+                post_buffer[:, current_seq_len] = sample_post.squeeze(-1)
+            else:
+                pre_buffer.copy_(torch.roll(pre_buffer, shifts=-1, dims=1))
+                post_buffer.copy_(torch.roll(post_buffer, shifts=-1, dims=1))
+                pre_buffer[:, -1] = sample_pre.squeeze(-1)
+                post_buffer[:, -1] = sample_post.squeeze(-1)
+
+        full_pre = torch.cat([x_token[0], generated_pre], dim=1)
+        full_post = torch.cat([x_token[1], generated_post], dim=1)
+
+        context_start = max(0, total_seq_len - max_context)
+        input_tokens = [
+            full_pre[:, context_start:total_seq_len].contiguous(),
+            full_post[:, context_start:total_seq_len].contiguous()
+        ]
+        z = tokenizer.decode(input_tokens, half=True)
+        z = z.reshape(-1, 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=None, 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']
+        
+        # Auto-detect device if not specified
+        if device is None:
+            if torch.cuda.is_available():
+                device = "cuda:0"
+            elif hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
+                device = "mps"
+            else:
+                device = "cpu"
+        
+        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/module.py

@@ -0,0 +1,570 @@
+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, collect_metrics=True):
+        # if self.input_format == 'bchw':
+        #     z = rearrange(z, 'b c h w -> b h w c')
+        zq = self.quantize(z)
+
+        q_scale = 1. / (self.embed_dim ** 0.5) if self.l2_norm else 1.
+
+        zq = zq * q_scale
+
+        if not collect_metrics:
+            return zq, zq.new_zeros(()), {}
+
+        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
+
+        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
+
+        # 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, collect_metrics=True):
+        z = F.normalize(z, dim=-1)
+        quantized, bsq_loss, metrics = self.bsq(z, collect_metrics=collect_metrics)
+        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)
+
+
+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 = F.scaled_dot_product_attention(
+            q, k, v,
+            attn_mask=attn_mask,
+            dropout_p=self.attn_dropout_p if self.training else 0.0,
+            is_causal=True
+        )
+
+        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 = F.scaled_dot_product_attention(
+            q, k, v,
+            attn_mask=attn_mask,
+            dropout_p=self.attn_dropout_p if self.training else 0.0,
+            is_causal=is_causal_flag
+        )
+
+        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 split_token(self, token_ids: torch.Tensor, s2_bits: int):
+        """Inputs:
+            token_ids (torch.Tensor): Composite token IDs of shape [batch_size, seq_len] or [N], each in range [0, 2^(s1_bits + s2_bits) - 1].
+            s2_bits (int): Number of low bits used for the fine token (s2).
+        """
+        assert isinstance(s2_bits, int) and s2_bits > 0, "s2_bits must be a positive integer"
+
+        t = token_ids.long()
+        mask = (1 << s2_bits) - 1
+        s2_ids = t & mask           # extract low bits
+        s1_ids = t >> s2_bits       # extract high bits
+        return s1_ids, s2_ids
+
+    def forward(self, token_ids):
+        """Inputs:
+        token_ids:
+            - tuple or list: (s1_ids, s2_ids), each of shape [batch_size, seq_len], or
+            - torch.Tensor: composite token IDs of shape [batch_size, seq_len], which will be split into (s1_ids, s2_ids) internally.
+        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
+
+
+
+
+
+
+
+

predict_us.py

@@ -0,0 +1,50 @@
+"""
+Data + metric helpers for the hosted dashboard (trimmed, no matplotlib/CLI).
+Mirrors the functions of the same name in ../predict_us.py.
+"""
+import numpy as np
+import pandas as pd
+
+
+def fetch_ohlcv(ticker: str, interval: str, period: str) -> pd.DataFrame:
+    """Download OHLCV via yfinance and normalise to Kronos' expected schema."""
+    import yfinance as yf
+
+    raw = yf.download(ticker, interval=interval, period=period,
+                      auto_adjust=True, progress=False)
+    if raw is None or raw.empty:
+        raise RuntimeError(f"No data returned for {ticker} ({interval}, {period}).")
+
+    if isinstance(raw.columns, pd.MultiIndex):
+        raw.columns = raw.columns.get_level_values(0)
+
+    raw = raw.rename(columns=str.lower)
+    df = raw[["open", "high", "low", "close", "volume"]].copy()
+    df = df.dropna().reset_index()
+    ts_col = "Date" if "Date" in df.columns else "Datetime"
+    df = df.rename(columns={ts_col: "timestamps"})
+    df["timestamps"] = pd.to_datetime(df["timestamps"]).dt.tz_localize(None)
+    return df
+
+
+def future_timestamps(last_ts: pd.Timestamp, interval: str, n: int) -> pd.Series:
+    """Generate n future timestamps after last_ts, respecting the bar interval."""
+    if interval.endswith("d"):
+        idx = pd.bdate_range(start=last_ts, periods=n + 1, freq="C")[1:]
+    elif interval.endswith("h"):
+        idx = pd.date_range(start=last_ts, periods=n + 1, freq="h")[1:]
+    else:
+        step = int("".join(c for c in interval if c.isdigit()) or "1")
+        idx = pd.date_range(start=last_ts, periods=n + 1, freq=f"{step}min")[1:]
+    return pd.Series(idx)
+
+
+def report_metrics(truth: np.ndarray, pred: np.ndarray) -> dict:
+    """Accuracy metrics on the close price."""
+    mae = float(np.mean(np.abs(pred - truth)))
+    rmse = float(np.sqrt(np.mean((pred - truth) ** 2)))
+    mape = float(np.mean(np.abs((pred - truth) / truth)) * 100)
+    dir_truth = np.sign(np.diff(truth))
+    dir_pred = np.sign(np.diff(pred))
+    da = float(np.mean(dir_truth == dir_pred) * 100) if len(dir_truth) else float("nan")
+    return {"MAE": mae, "RMSE": rmse, "MAPE%": mape, "DirAcc%": da}

requirements.txt

@@ -0,0 +1,10 @@
+--extra-index-url https://download.pytorch.org/whl/cpu
+torch
+numpy
+pandas
+einops==0.8.1
+huggingface_hub==0.33.1
+safetensors==0.6.2
+tqdm==4.67.1
+flask==3.0.3
+yfinance

templates/dashboard.html

@@ -0,0 +1,225 @@
+<!doctype html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>Kronos · US Forecast Dashboard</title>
+<script src="https://cdn.jsdelivr.net/npm/chart.js@4.4.1/dist/chart.umd.min.js"></script>
+<style>
+  :root{
+    --bg:#0d1117; --panel:#161b22; --panel2:#1c2330; --line:#30363d;
+    --fg:#e6edf3; --muted:#8b949e; --accent:#e5484d; --blue:#4493f8; --good:#3fb950;
+  }
+  *{box-sizing:border-box}
+  body{margin:0;font:14px/1.5 -apple-system,BlinkMacSystemFont,"Segoe UI",Roboto,sans-serif;
+    background:var(--bg);color:var(--fg)}
+  header{padding:16px 22px;border-bottom:1px solid var(--line);display:flex;
+    align-items:baseline;gap:12px;flex-wrap:wrap}
+  header h1{font-size:17px;margin:0;font-weight:650}
+  header .sub{color:var(--muted);font-size:12.5px}
+  .wrap{display:grid;grid-template-columns:290px 1fr;gap:18px;padding:18px 22px;
+    align-items:start}
+  @media(max-width:880px){.wrap{grid-template-columns:1fr}}
+  .panel{background:var(--panel);border:1px solid var(--line);border-radius:10px;padding:16px}
+  .panel h2{font-size:12px;text-transform:uppercase;letter-spacing:.06em;color:var(--muted);
+    margin:0 0 12px}
+  label{display:block;font-size:12px;color:var(--muted);margin:10px 0 4px}
+  input,select{width:100%;background:var(--panel2);border:1px solid var(--line);color:var(--fg);
+    border-radius:7px;padding:8px 9px;font-size:13px}
+  input:focus,select:focus{outline:none;border-color:var(--blue)}
+  .row2{display:grid;grid-template-columns:1fr 1fr;gap:10px}
+  button{width:100%;margin-top:16px;background:var(--accent);color:#fff;border:none;
+    border-radius:8px;padding:11px;font-size:14px;font-weight:600;cursor:pointer}
+  button:disabled{opacity:.55;cursor:not-allowed}
+  .chk{display:flex;align-items:center;gap:8px;margin-top:12px;color:var(--muted);font-size:12.5px}
+  .chk input{width:auto}
+  .metrics{display:grid;grid-template-columns:repeat(5,1fr);gap:12px;margin-bottom:16px}
+  @media(max-width:880px){.metrics{grid-template-columns:repeat(2,1fr)}}
+  .card{background:var(--panel);border:1px solid var(--line);border-radius:10px;padding:14px}
+  .card .k{font-size:11px;color:var(--muted);text-transform:uppercase;letter-spacing:.05em}
+  .card .v{font-size:22px;font-weight:680;margin-top:5px}
+  .card .u{font-size:12px;color:var(--muted);font-weight:400}
+  .chartbox{position:relative;height:460px}
+  .status{color:var(--muted);font-size:13px;margin-top:10px;min-height:18px}
+  .status.err{color:var(--accent)}
+  .spin{display:inline-block;width:13px;height:13px;border:2px solid var(--muted);
+    border-top-color:transparent;border-radius:50%;animation:sp .8s linear infinite;
+    vertical-align:-2px;margin-right:7px}
+  @keyframes sp{to{transform:rotate(360deg)}}
+  .empty{color:var(--muted);text-align:center;padding:120px 0;font-size:13px}
+  /* info tooltips */
+  .info{display:inline-block;width:14px;height:14px;line-height:14px;text-align:center;
+    border-radius:50%;background:#30363d;color:#8b949e;font-size:10px;font-weight:700;
+    cursor:help;position:relative;margin-left:5px;vertical-align:1px}
+  .info:hover{background:var(--blue);color:#fff}
+  .info:hover::after{content:attr(data-tip);position:absolute;top:19px;left:0;width:230px;
+    background:#0d1117;border:1px solid var(--line);border-radius:7px;padding:9px 11px;
+    color:var(--fg);font-size:11.5px;line-height:1.5;font-weight:400;z-index:60;
+    box-shadow:0 8px 24px rgba(0,0,0,.55);white-space:normal;text-transform:none;letter-spacing:0}
+  .info.il:hover::after{left:auto;right:0}   /* right-column fields open leftward */
+</style>
+</head>
+<body>
+<header>
+  <h1>Kronos · US Forecast</h1>
+  <span class="sub">CPU ensemble · median + 10–90% band · yfinance data</span>
+</header>
+
+<div class="wrap">
+  <!-- controls -->
+  <div class="panel">
+    <h2>Parameters</h2>
+    <label>Ticker
+      <span class="info" data-tip="股票/ETF 代號,任何 yfinance 抓得到的美股皆可,如 AAPL、NVDA、MSFT、SPY、QQQ。自動轉大寫。冷門或剛上市的可能資料不足。"></span></label>
+    <input id="ticker" value="AAPL" autocomplete="off">
+    <div class="row2">
+      <div><label>Interval
+        <span class="info" data-tip="K 線週期。1d 日線最穩;1h/5m 的歷史 yfinance 只給最近一小段(5m≈60天),用它們時 Period 要調短,否則抓不到足夠根數會報錯。"></span></label>
+        <select id="interval"><option>1d</option><option>1h</option><option>5m</option></select>
+      </div>
+      <div><label>Period
+        <span class="info il" data-tip="往回下載多久的歷史。這是資料範圍,不是餵給模型的長度(那是 Lookback)。只要夠切出 Lookback+Pred len 根即可。別跨太多不同價格量級。"></span></label>
+        <select id="period"><option>1y</option><option selected>3y</option><option>5y</option></select>
+      </div>
+    </div>
+    <div class="row2">
+      <div><label>Lookback
+        <span class="info" data-tip="餵給模型看的歷史根數(model context)。模型上限 512,別超過 ~500;太短看不到趨勢。400 是實測甜蜜點,一般不用動。"></span></label>
+        <input id="lookback" type="number" value="400"></div>
+      <div><label>Pred len
+        <span class="info il" data-tip="要預測未來幾根。越長越不準(不確定性隨步數累積)。短線 5–10,中期 20–30。backtest 模式也用它決定藏起幾根當真值。"></span></label>
+        <input id="pred_len" type="number" value="20"></div>
+    </div>
+    <div class="row2">
+      <div><label>Mode
+        <span class="info" data-tip="backtest:藏起最後 Pred len 根真值讓模型預測、再對比打分(才有下方指標)。future:用全部最新資料預測真正的未來,無真值、無分數。"></span></label>
+        <select id="mode"><option value="backtest">backtest</option><option value="future">future</option></select>
+      </div>
+      <div><label>Paths
+        <span class="info il" data-tip="集成路徑數(此線上版上限 30)。每條是一次獨立抽樣;中位數與 10–90% 區間都由這 N 條算出。越多越穩越慢(CPU 每條 ~1–2s)。集成是把 MAPE 從 8.6% 降到 ~6% 的關鍵。"></span></label>
+        <input id="n_paths" type="number" value="20" max="30"></div>
+    </div>
+    <div class="row2">
+      <div><label>Temperature T
+        <span class="info" data-tip="抽樣亂度。低(0.5–0.7)保守、路徑集中、方向命中較高;高(1.0+)發散、區間更寬。官方預設 1.0,我們實測 0.7 方向命中最好。"></span></label>
+        <input id="T" type="number" step="0.05" value="0.7"></div>
+      <div><label>top_p
+        <span class="info il" data-tip="Nucleus 抽樣:只從累積機率前 p 的候選裡抽。低(0.8)砍掉長尾極端值、更保守;高(0.95+)允許更極端走勢。官方與我們都用 0.9。"></span></label>
+        <input id="top_p" type="number" step="0.05" value="0.9"></div>
+    </div>
+    <label class="chk"><input type="checkbox" id="showpaths"> show sample paths
+      <span class="info" data-tip="只是顯示開關,不影響計算。勾選會把 N 條原始抽樣路徑疊上圖,直觀看發散程度;不勾只留乾淨的中位數 + 區間。"></span></label>
+    <button id="run">Run forecast</button>
+    <div id="status" class="status"></div>
+  </div>
+
+  <!-- results -->
+  <div>
+    <div class="metrics" id="metrics" style="display:none">
+      <div class="card"><div class="k">MAE <span class="info" data-tip="平均絕對誤差(美元)。越小越好。會隨股價高低而變,不同標的之間別直接比。"></span></div><div class="v" id="m_mae">–</div></div>
+      <div class="card"><div class="k">MAPE <span class="info" data-tip="平均百分比誤差。跨標的可比的主要指標,~6% 算不錯。"></span></div><div class="v" id="m_mape">–</div></div>
+      <div class="card"><div class="k">Dir Acc <span class="info" data-tip="方向命中率(每步漲跌預測對不對)。50% = 瞎猜,>55% 才算有訊號。"></span></div><div class="v" id="m_dir">–</div></div>
+      <div class="card"><div class="k">Band cover <span class="info il" data-tip="真值落在 10–90% 區間內的比例,理想 ~80%。目前偏低 = 模型過度自信,是待辦裡『區間校準』要修的。"></span></div><div class="v" id="m_cov">–</div></div>
+      <div class="card"><div class="k">Runtime <span class="info il" data-tip="這次集成在 CPU 上花的秒數,主要由 Paths × Pred len 決定。"></span></div><div class="v" id="m_time">–</div></div>
+    </div>
+    <div class="panel">
+      <div class="chartbox">
+        <canvas id="chart" style="display:none"></canvas>
+        <div class="empty" id="empty">Set parameters and click <b>Run forecast</b>.<br>
+          First run downloads/loads the model (~20–40s), then each path ~1–2s on CPU.</div>
+      </div>
+    </div>
+  </div>
+</div>
+
+<script>
+const $ = id => document.getElementById(id);
+let chart = null;
+const fmtDate = s => s.slice(0,10);
+
+async function run(){
+  const btn = $('run'), st = $('status');
+  const body = {
+    ticker:$('ticker').value, interval:$('interval').value, period:$('period').value,
+    lookback:+$('lookback').value, pred_len:+$('pred_len').value, mode:$('mode').value,
+    n_paths:+$('n_paths').value, T:+$('T').value, top_p:+$('top_p').value,
+  };
+  btn.disabled = true;
+  st.className = 'status';
+  const t0 = Date.now();
+  const tick = setInterval(()=>{ st.innerHTML =
+    `<span class="spin"></span>running ${body.n_paths} paths on CPU… ${((Date.now()-t0)/1000).toFixed(0)}s`;
+  }, 200);
+  try{
+    const r = await fetch('/api/forecast', {method:'POST',
+      headers:{'Content-Type':'application/json'}, body:JSON.stringify(body)});
+    const j = await r.json();
+    clearInterval(tick);
+    if(!j.ok){ st.className='status err'; st.textContent = 'Error: '+j.error; return; }
+    st.textContent = `done in ${j.result.meta.elapsed}s`;
+    render(j.result);
+  }catch(e){
+    clearInterval(tick); st.className='status err'; st.textContent = 'Request failed: '+e;
+  }finally{
+    btn.disabled = false;
+  }
+}
+
+function render(res){
+  $('empty').style.display='none';
+  $('chart').style.display='block';
+  const H = res.history, F = res.forecast, Tr = res.truth, M = res.metrics;
+  const labels = [...H.t, ...F.t].map(fmtDate);
+  const nH = H.t.length, nF = F.t.length;
+  const anchor = H.close[nH-1];               // connect history -> forecast
+  const pad = n => Array(n).fill(null);
+
+  // forecast series padded to sit under the forecast portion, anchored to last hist pt
+  const fc = arr => { const d = pad(nH-1); d.push(anchor); return d.concat(arr); };
+  const histLine = H.close.concat(pad(nF));
+
+  const ds = [
+    {label:'History', data:histLine, borderColor:'#8b949e', borderWidth:1.4,
+      pointRadius:0, tension:.15},
+    {label:'10–90% band (hi)', data:fc(F.hi), borderColor:'transparent',
+      backgroundColor:'rgba(229,72,77,.16)', pointRadius:0, fill:'+1'},
+    {label:'10–90% band (lo)', data:fc(F.lo), borderColor:'transparent',
+      pointRadius:0, fill:false},
+    {label:'Median forecast', data:fc(F.median), borderColor:'#e5484d',
+      borderWidth:2.4, pointRadius:0, tension:.15},
+  ];
+  if(Tr){ ds.push({label:'Ground truth', data:fc(Tr.close), borderColor:'#4493f8',
+      borderWidth:2.4, pointRadius:0, tension:.15}); }
+  if($('showpaths').checked){
+    F.paths.forEach(p => ds.push({label:'_path', data:fc(p), borderColor:'rgba(229,72,77,.13)',
+      borderWidth:.6, pointRadius:0, tension:.1}));
+  }
+
+  const cfg = {type:'line', data:{labels, datasets:ds}, options:{
+    responsive:true, maintainAspectRatio:false, animation:false,
+    interaction:{mode:'index', intersect:false},
+    plugins:{legend:{labels:{color:'#e6edf3', filter:i=>!i.text.startsWith('_')
+        && !i.text.includes('(lo)')}},
+      tooltip:{filter:i=>i.dataset.label!=='_path' && !i.dataset.label.includes('(lo)')}},
+    scales:{
+      x:{ticks:{color:'#8b949e',maxTicksLimit:12,autoSkip:true},grid:{color:'#21262d'}},
+      y:{ticks:{color:'#8b949e'},grid:{color:'#21262d'}},
+    }}};
+  if(chart) chart.destroy();
+  chart = new Chart($('chart'), cfg);
+
+  // metrics (accuracy cards only exist in backtest; runtime always shown)
+  $('metrics').style.display='grid';
+  const U = '<span class="u">%</span>';
+  $('m_mae').innerHTML  = M ? M['MAE'].toFixed(2) : '–';
+  $('m_mape').innerHTML = M ? M['MAPE%'].toFixed(2)+U : '–';
+  $('m_dir').innerHTML  = M ? M['DirAcc%'].toFixed(0)+U : '–';
+  $('m_cov').innerHTML  = M ? M['coverage%'].toFixed(0)+U : '–';
+  $('m_time').innerHTML = res.meta.elapsed+'<span class="u">s</span>';
+}
+
+$('run').addEventListener('click', run);
+$('ticker').addEventListener('keydown', e=>{ if(e.key==='Enter') run(); });
+</script>
+</body>
+</html>