llama-rs 0.17.0

use std::collections::HashMap;
use std::fs::File;
use std::path::Path;

use memmap2::Mmap;
use serde_json::Value;

use super::{SafeTensorsError, SafeTensorsResult};

/// SafeTensors data types
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum SafeTensorsDtype {
    F32,
    F16,
    BF16,
    F64,
    I8,
    I16,
    I32,
    I64,
    U8,
    Bool,
}

impl SafeTensorsDtype {
    /// Parse dtype from string (e.g., "F32", "BF16")
    pub fn from_str(s: &str) -> SafeTensorsResult<Self> {
        match s {
            "F32" => Ok(Self::F32),
            "F16" => Ok(Self::F16),
            "BF16" => Ok(Self::BF16),
            "F64" => Ok(Self::F64),
            "I8" => Ok(Self::I8),
            "I16" => Ok(Self::I16),
            "I32" => Ok(Self::I32),
            "I64" => Ok(Self::I64),
            "U8" => Ok(Self::U8),
            "BOOL" => Ok(Self::Bool),
            _ => Err(SafeTensorsError::UnsupportedDtype(s.to_string())),
        }
    }

    /// Get the size in bytes of a single element of this dtype
    pub fn element_size(&self) -> usize {
        match self {
            Self::F32 | Self::I32 => 4,
            Self::F16 | Self::BF16 | Self::I16 => 2,
            Self::F64 | Self::I64 => 8,
            Self::I8 | Self::U8 | Self::Bool => 1,
        }
    }
}

/// Metadata about a single tensor in a SafeTensors file
#[derive(Debug, Clone)]
pub struct SafeTensorInfo {
    pub dtype: SafeTensorsDtype,
    pub shape: Vec<usize>,
    pub data_start: usize,  // byte offset relative to data section start
    pub data_end: usize,    // byte offset relative to data section start
}

impl SafeTensorInfo {
    pub fn byte_size(&self) -> usize {
        self.data_end - self.data_start
    }

    pub fn n_elements(&self) -> usize {
        self.shape.iter().product()
    }
}

/// A parsed SafeTensors file with memory-mapped data
pub struct SafeTensorsFile {
    mmap: Mmap,
    tensors: HashMap<String, SafeTensorInfo>,
    data_offset: usize,  // = 8 + header_size
}

impl SafeTensorsFile {
    /// Open and parse a SafeTensors file
    pub fn open(path: impl AsRef<Path>) -> SafeTensorsResult<Self> {
        let file = File::open(path.as_ref())?;
        let mmap = unsafe { Mmap::map(&file)? };

        // Read 8-byte header size (u64 little-endian)
        if mmap.len() < 8 {
            return Err(SafeTensorsError::InvalidFormat(
                "File too small for header".to_string(),
            ));
        }

        let header_size = u64::from_le_bytes(mmap[0..8].try_into().unwrap()) as usize;

        // Validate header size
        if header_size > mmap.len() - 8 {
            return Err(SafeTensorsError::InvalidFormat(format!(
                "Header size {} exceeds file size {}",
                header_size,
                mmap.len() - 8
            )));
        }

        let data_offset = 8 + header_size;

        // Parse JSON header
        let header_bytes = &mmap[8..data_offset];
        let header_json: HashMap<String, Value> = serde_json::from_slice(header_bytes)?;

        let mut tensors = HashMap::new();

        for (name, value) in header_json.iter() {
            // Skip metadata entries
            if name == "__metadata__" {
                continue;
            }

            let obj = value.as_object().ok_or_else(|| {
                SafeTensorsError::InvalidFormat(format!("Tensor {} is not an object", name))
            })?;

            // Parse dtype
            let dtype_str = obj
                .get("dtype")
                .and_then(|v| v.as_str())
                .ok_or_else(|| {
                    SafeTensorsError::InvalidFormat(format!("Missing dtype for tensor {}", name))
                })?;
            let dtype = SafeTensorsDtype::from_str(dtype_str)?;

            // Parse shape
            let shape_array = obj
                .get("shape")
                .and_then(|v| v.as_array())
                .ok_or_else(|| {
                    SafeTensorsError::InvalidFormat(format!("Missing shape for tensor {}", name))
                })?;
            let shape: Vec<usize> = shape_array
                .iter()
                .map(|v| {
                    v.as_u64()
                        .ok_or_else(|| {
                            SafeTensorsError::InvalidFormat(format!(
                                "Invalid shape value for tensor {}",
                                name
                            ))
                        })
                        .map(|x| x as usize)
                })
                .collect::<SafeTensorsResult<_>>()?;

            // Parse data_offsets
            let data_offsets = obj
                .get("data_offsets")
                .and_then(|v| v.as_array())
                .ok_or_else(|| {
                    SafeTensorsError::InvalidFormat(format!(
                        "Missing data_offsets for tensor {}",
                        name
                    ))
                })?;

            if data_offsets.len() != 2 {
                return Err(SafeTensorsError::InvalidFormat(format!(
                    "Expected 2 data_offsets for tensor {}, got {}",
                    name,
                    data_offsets.len()
                )));
            }

            let data_start = data_offsets[0].as_u64().ok_or_else(|| {
                SafeTensorsError::InvalidFormat(format!(
                    "Invalid data_start for tensor {}",
                    name
                ))
            })? as usize;

            let data_end = data_offsets[1].as_u64().ok_or_else(|| {
                SafeTensorsError::InvalidFormat(format!("Invalid data_end for tensor {}", name))
            })? as usize;

            tensors.insert(
                name.clone(),
                SafeTensorInfo {
                    dtype,
                    shape,
                    data_start,
                    data_end,
                },
            );
        }

        Ok(Self {
            mmap,
            tensors,
            data_offset,
        })
    }

    /// Get the raw bytes for a tensor
    pub fn tensor_data(&self, name: &str) -> Option<&[u8]> {
        let info = self.tensors.get(name)?;
        let start = self.data_offset + info.data_start;
        let end = self.data_offset + info.data_end;

        if end > self.mmap.len() {
            return None;
        }

        Some(&self.mmap[start..end])
    }

    /// Get tensor metadata
    pub fn tensor_info(&self, name: &str) -> Option<&SafeTensorInfo> {
        self.tensors.get(name)
    }

    /// Iterator over all tensor names
    pub fn tensor_names(&self) -> impl Iterator<Item = &str> {
        self.tensors.keys().map(|s| s.as_str())
    }

    /// Get the number of tensors in this file
    pub fn num_tensors(&self) -> usize {
        self.tensors.len()
    }
}

/// Manages a sharded SafeTensors model
pub struct ShardedSafeTensors {
    shards: Vec<SafeTensorsFile>,
    tensor_to_shard: HashMap<String, usize>,  // tensor name -> shard index
}

impl ShardedSafeTensors {
    /// Open a sharded or single-file SafeTensors model
    ///
    /// First looks for model.safetensors.index.json, then falls back to model.safetensors
    pub fn open(dir: impl AsRef<Path>) -> SafeTensorsResult<Self> {
        let dir = dir.as_ref();
        let index_path = dir.join("model.safetensors.index.json");
        let single_path = dir.join("model.safetensors");

        if index_path.exists() {
            // Sharded model
            let index_file = File::open(&index_path)?;
            let index_json: Value = serde_json::from_reader(index_file)?;

            let weight_map = index_json
                .get("weight_map")
                .and_then(|v| v.as_object())
                .ok_or_else(|| {
                    SafeTensorsError::InvalidFormat(
                        "Missing or invalid weight_map in index.json".to_string(),
                    )
                })?;

            // Collect unique shard filenames
            let mut shard_filenames: Vec<String> = weight_map
                .values()
                .filter_map(|v| v.as_str())
                .map(|s| s.to_string())
                .collect();
            shard_filenames.sort();
            shard_filenames.dedup();

            // Open each shard
            let mut shards = Vec::new();
            let mut shard_name_to_idx = HashMap::new();

            for (idx, filename) in shard_filenames.iter().enumerate() {
                let shard_path = dir.join(filename);
                let shard = SafeTensorsFile::open(&shard_path)?;
                shards.push(shard);
                shard_name_to_idx.insert(filename.clone(), idx);
            }

            // Build tensor -> shard index mapping
            let mut tensor_to_shard = HashMap::new();
            for (tensor_name, shard_filename) in weight_map.iter() {
                let shard_filename_str = shard_filename.as_str().ok_or_else(|| {
                    SafeTensorsError::InvalidFormat(format!(
                        "Invalid shard filename for tensor {}",
                        tensor_name
                    ))
                })?;

                let shard_idx = shard_name_to_idx.get(shard_filename_str).ok_or_else(|| {
                    SafeTensorsError::InvalidFormat(format!(
                        "Shard {} not found for tensor {}",
                        shard_filename_str, tensor_name
                    ))
                })?;

                tensor_to_shard.insert(tensor_name.clone(), *shard_idx);
            }

            Ok(Self {
                shards,
                tensor_to_shard,
            })
        } else if single_path.exists() {
            // Single-file model
            let shard = SafeTensorsFile::open(&single_path)?;
            let tensor_to_shard: HashMap<String, usize> = shard
                .tensor_names()
                .map(|name| (name.to_string(), 0))
                .collect();

            Ok(Self {
                shards: vec![shard],
                tensor_to_shard,
            })
        } else {
            Err(SafeTensorsError::InvalidFormat(
                "No model.safetensors or model.safetensors.index.json found".to_string(),
            ))
        }
    }

    /// Get the raw bytes for a tensor
    pub fn tensor_data(&self, name: &str) -> Option<&[u8]> {
        let shard_idx = self.tensor_to_shard.get(name)?;
        let shard = self.shards.get(*shard_idx)?;
        shard.tensor_data(name)
    }

    /// Get tensor metadata
    pub fn tensor_info(&self, name: &str) -> Option<&SafeTensorInfo> {
        let shard_idx = self.tensor_to_shard.get(name)?;
        let shard = self.shards.get(*shard_idx)?;
        shard.tensor_info(name)
    }

    /// Get all tensor names across all shards
    pub fn tensor_names(&self) -> Vec<String> {
        self.tensor_to_shard.keys().cloned().collect()
    }

    /// Get the number of tensors across all shards
    pub fn num_tensors(&self) -> usize {
        self.tensor_to_shard.len()
    }
}

/// Convert BF16 bytes to F32 vec
pub fn bf16_to_f32(data: &[u8]) -> Vec<f32> {
    let bf16s: &[u16] = bytemuck::cast_slice(data);
    bf16s
        .iter()
        .map(|&bits| f32::from_bits((bits as u32) << 16))
        .collect()
}

/// Convert F16 bytes to F32 vec
pub fn f16_to_f32(data: &[u8]) -> Vec<f32> {
    let f16s: &[u16] = bytemuck::cast_slice(data);
    f16s
        .iter()
        .map(|&bits| half::f16::from_bits(bits).to_f32())
        .collect()
}

#[cfg(test)]
mod tests {
    use super::*;
    use std::io::Write;
    use tempfile::NamedTempFile;

    #[test]
    fn test_dtype_from_str() {
        assert_eq!(SafeTensorsDtype::from_str("F32").unwrap(), SafeTensorsDtype::F32);
        assert_eq!(SafeTensorsDtype::from_str("F16").unwrap(), SafeTensorsDtype::F16);
        assert_eq!(SafeTensorsDtype::from_str("BF16").unwrap(), SafeTensorsDtype::BF16);
        assert_eq!(SafeTensorsDtype::from_str("F64").unwrap(), SafeTensorsDtype::F64);
        assert_eq!(SafeTensorsDtype::from_str("I8").unwrap(), SafeTensorsDtype::I8);
        assert_eq!(SafeTensorsDtype::from_str("I16").unwrap(), SafeTensorsDtype::I16);
        assert_eq!(SafeTensorsDtype::from_str("I32").unwrap(), SafeTensorsDtype::I32);
        assert_eq!(SafeTensorsDtype::from_str("I64").unwrap(), SafeTensorsDtype::I64);
        assert_eq!(SafeTensorsDtype::from_str("U8").unwrap(), SafeTensorsDtype::U8);
        assert_eq!(SafeTensorsDtype::from_str("BOOL").unwrap(), SafeTensorsDtype::Bool);
        assert!(SafeTensorsDtype::from_str("INVALID").is_err());
    }

    #[test]
    fn test_parse_single_file() {
        // Create a temp SafeTensors file with 2 F32 tensors
        let mut tmpfile = NamedTempFile::new().unwrap();

        // Tensor 1: [2, 3] F32 with values [1.0, 2.0, 3.0, 4.0, 5.0, 6.0]
        // Tensor 2: [4] F32 with values [10.0, 20.0, 30.0, 40.0]

        let tensor1_data: Vec<f32> = vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0];
        let tensor2_data: Vec<f32> = vec![10.0, 20.0, 30.0, 40.0];

        let tensor1_bytes: Vec<u8> = tensor1_data
            .iter()
            .flat_map(|f| f.to_le_bytes())
            .collect();
        let tensor2_bytes: Vec<u8> = tensor2_data
            .iter()
            .flat_map(|f| f.to_le_bytes())
            .collect();

        // Build JSON header
        let header_json = serde_json::json!({
            "__metadata__": {"format": "pt"},
            "tensor1": {
                "dtype": "F32",
                "shape": [2, 3],
                "data_offsets": [0, tensor1_bytes.len()]
            },
            "tensor2": {
                "dtype": "F32",
                "shape": [4],
                "data_offsets": [tensor1_bytes.len(), tensor1_bytes.len() + tensor2_bytes.len()]
            }
        });

        let header_str = serde_json::to_string(&header_json).unwrap();
        let header_bytes = header_str.as_bytes();

        // Write file: 8-byte header size, JSON header, tensor data
        tmpfile.write_all(&(header_bytes.len() as u64).to_le_bytes()).unwrap();
        tmpfile.write_all(header_bytes).unwrap();
        tmpfile.write_all(&tensor1_bytes).unwrap();
        tmpfile.write_all(&tensor2_bytes).unwrap();
        tmpfile.flush().unwrap();

        // Parse
        let st = SafeTensorsFile::open(tmpfile.path()).unwrap();

        // Verify tensor count
        assert_eq!(st.num_tensors(), 2);

        // Verify tensor1
        let info1 = st.tensor_info("tensor1").unwrap();
        assert_eq!(info1.dtype, SafeTensorsDtype::F32);
        assert_eq!(info1.shape, vec![2, 3]);
        assert_eq!(info1.n_elements(), 6);
        assert_eq!(info1.byte_size(), 24);

        let data1 = st.tensor_data("tensor1").unwrap();
        // Convert bytes to f32 safely (handles potential alignment issues)
        let f32_data1: Vec<f32> = data1
            .chunks_exact(4)
            .map(|chunk| f32::from_le_bytes([chunk[0], chunk[1], chunk[2], chunk[3]]))
            .collect();
        assert_eq!(f32_data1, tensor1_data);

        // Verify tensor2
        let info2 = st.tensor_info("tensor2").unwrap();
        assert_eq!(info2.dtype, SafeTensorsDtype::F32);
        assert_eq!(info2.shape, vec![4]);
        assert_eq!(info2.n_elements(), 4);
        assert_eq!(info2.byte_size(), 16);

        let data2 = st.tensor_data("tensor2").unwrap();
        // Convert bytes to f32 safely (handles potential alignment issues)
        let f32_data2: Vec<f32> = data2
            .chunks_exact(4)
            .map(|chunk| f32::from_le_bytes([chunk[0], chunk[1], chunk[2], chunk[3]]))
            .collect();
        assert_eq!(f32_data2, tensor2_data);
    }

    #[test]
    fn test_bf16_to_f32() {
        // BF16 representation of 1.0 is 0x3F80
        // In little-endian bytes: [0x80, 0x3F]
        let bf16_bytes: Vec<u8> = vec![0x80, 0x3F, 0x00, 0x40]; // 1.0, 2.0 in BF16
        let f32_vec = bf16_to_f32(&bf16_bytes);

        assert_eq!(f32_vec.len(), 2);
        assert!((f32_vec[0] - 1.0).abs() < 1e-6);
        assert!((f32_vec[1] - 2.0).abs() < 1e-6);
    }

    #[test]
    fn test_f16_to_f32() {
        // F16 representation of 1.0
        let one_f16 = half::f16::from_f32(1.0);
        let two_f16 = half::f16::from_f32(2.0);

        let f16_bytes: Vec<u8> = vec![
            one_f16.to_bits().to_le_bytes()[0],
            one_f16.to_bits().to_le_bytes()[1],
            two_f16.to_bits().to_le_bytes()[0],
            two_f16.to_bits().to_le_bytes()[1],
        ];

        let f32_vec = f16_to_f32(&f16_bytes);

        assert_eq!(f32_vec.len(), 2);
        assert!((f32_vec[0] - 1.0).abs() < 1e-3);
        assert!((f32_vec[1] - 2.0).abs() < 1e-3);
    }

    #[test]
    fn test_tensor_not_found() {
        let mut tmpfile = NamedTempFile::new().unwrap();

        let header_json = serde_json::json!({
            "tensor1": {
                "dtype": "F32",
                "shape": [2],
                "data_offsets": [0, 8]
            }
        });

        let header_str = serde_json::to_string(&header_json).unwrap();
        let header_bytes = header_str.as_bytes();

        tmpfile.write_all(&(header_bytes.len() as u64).to_le_bytes()).unwrap();
        tmpfile.write_all(header_bytes).unwrap();
        tmpfile.write_all(&[0u8; 8]).unwrap(); // dummy data
        tmpfile.flush().unwrap();

        let st = SafeTensorsFile::open(tmpfile.path()).unwrap();

        assert!(st.tensor_info("nonexistent").is_none());
        assert!(st.tensor_data("nonexistent").is_none());
    }
}