Skip to main content

zsfm_tabdpt/
convert.rs

1use std::fs::File;
2use std::io::BufWriter;
3use std::path::{Path, PathBuf};
4
5use anyhow::Context;
6use indicatif::{ProgressBar, ProgressStyle};
7use safetensors::Dtype as StDtype;
8use safetensors::SafeTensors;
9
10use zsfm_gguf::{GGMLType, GGUFMetaValue, GGUFWriter};
11
12use crate::config::TabDptConfig;
13use crate::tensor_map::map_tensor_name;
14
15pub struct ConvertOptions {
16    pub output_dtype: GGMLType,
17}
18
19pub fn convert(
20    shard_paths: &[PathBuf],
21    config: &TabDptConfig,
22    opts: &ConvertOptions,
23    output_path: &Path,
24) -> anyhow::Result<()> {
25    let mut writer = GGUFWriter::new();
26    write_metadata(&mut writer, config);
27
28    let shard_bytes: Vec<Vec<u8>> = shard_paths
29        .iter()
30        .map(|p| std::fs::read(p).with_context(|| format!("read {}", p.display())))
31        .collect::<anyhow::Result<_>>()?;
32    let shard_views: Vec<SafeTensors> = shard_bytes
33        .iter()
34        .map(|b| SafeTensors::deserialize(b).context("deserialize shard"))
35        .collect::<anyhow::Result<_>>()?;
36
37    let total: usize = shard_views.iter().map(|s| s.len()).sum();
38    let pb = ProgressBar::new(total as u64);
39    pb.set_style(
40        ProgressStyle::with_template(
41            "{spinner:.green} [{elapsed_precise}] [{bar:40.cyan/blue}] {pos}/{len} {msg}",
42        )
43        .unwrap()
44        .progress_chars("=>-"),
45    );
46
47    let mut mapped = 0usize;
48    let mut skipped: Vec<String> = Vec::new();
49    let mut fallback_count = 0usize;
50
51    for shard in &shard_views {
52        for (hf_name, tensor_view) in shard.tensors() {
53            pb.set_message(hf_name.to_string());
54
55            let gguf_name = match map_tensor_name(&hf_name) {
56                Some(n) => n,
57                None => {
58                    skipped.push(hf_name.to_string());
59                    pb.inc(1);
60                    continue;
61                }
62            };
63
64            let src_dtype = ggml_type_from_st(tensor_view.dtype())
65                .with_context(|| format!("tensor {hf_name}: unsupported dtype {:?}", tensor_view.dtype()))?;
66
67            let raw_data = tensor_view.data();
68            let py_shape = tensor_view.shape();
69            let n_elems: usize = py_shape.iter().product();
70            let innermost = py_shape.last().copied().unwrap_or(1);
71
72            let (dst_dtype, gguf_shape, tensor_data) =
73                if opts.output_dtype == GGMLType::Q8_0 && (innermost % 32 != 0 || n_elems % 32 != 0) {
74                    fallback_count += 1;
75                    let data = cast_data(raw_data, src_dtype, GGMLType::F32)
76                        .with_context(|| format!("tensor {hf_name}: cast failed"))?;
77                    let gs = py_shape.iter().rev().map(|&d| d as u64).collect();
78                    (GGMLType::F32, gs, data)
79                } else {
80                    let dst = opts.output_dtype;
81                    let data = cast_data(raw_data, src_dtype, dst)
82                        .with_context(|| format!("tensor {hf_name}: cast failed"))?;
83                    let gs = py_shape.iter().rev().map(|&d| d as u64).collect();
84                    (dst, gs, data)
85                };
86
87            writer.add_tensor(gguf_name, gguf_shape, dst_dtype, tensor_data);
88            mapped += 1;
89            pb.inc(1);
90        }
91    }
92
93    pb.finish_with_message("tensors processed");
94
95    if !skipped.is_empty() {
96        eprintln!("\nNote: {} tensor(s) skipped (scalar attention-scaling buffers, recomputed from config):", skipped.len());
97        for name in &skipped {
98            eprintln!("  {name}");
99        }
100    }
101    if fallback_count > 0 {
102        eprintln!("\nNote: {fallback_count} tensor(s) fell back to F32.");
103    }
104
105    println!("Writing {mapped} tensors to {} …", output_path.display());
106    let out_file = File::create(output_path).with_context(|| format!("create {}", output_path.display()))?;
107    let mut buf_writer = BufWriter::new(out_file);
108    writer.write_to(&mut buf_writer)?;
109    println!("Done.");
110    Ok(())
111}
112
113fn write_metadata(writer: &mut GGUFWriter, config: &TabDptConfig) {
114    writer.add_metadata("general.architecture", GGUFMetaValue::String("tabdpt".into()));
115    writer.add_metadata("general.name", GGUFMetaValue::String("Layer6/TabDPT".into()));
116    writer.add_metadata("tabdpt.dim", GGUFMetaValue::Uint32(config.dim as u32));
117    writer.add_metadata("tabdpt.n_layers", GGUFMetaValue::Uint32(config.n_layers as u32));
118    writer.add_metadata("tabdpt.n_heads", GGUFMetaValue::Uint32(config.n_heads as u32));
119    writer.add_metadata("tabdpt.ff_dim", GGUFMetaValue::Uint32(config.ff_dim as u32));
120    writer.add_metadata("tabdpt.y_encoder_dim", GGUFMetaValue::Uint32(config.y_encoder_dim as u32));
121    writer.add_metadata("tabdpt.max_num_classes", GGUFMetaValue::Uint32(config.max_num_classes as u32));
122    writer.add_metadata("tabdpt.regression_bin_count", GGUFMetaValue::Uint32(config.regression_bin_count as u32));
123    writer.add_metadata("tabdpt.regression_bin_min", GGUFMetaValue::Float32(config.regression_bin_min));
124    writer.add_metadata("tabdpt.regression_bin_max", GGUFMetaValue::Float32(config.regression_bin_max));
125    writer.add_metadata("tabdpt.max_num_features", GGUFMetaValue::Uint32(config.max_num_features as u32));
126    writer.add_metadata("tabdpt.base_len", GGUFMetaValue::Uint32(config.base_len as u32));
127    writer.add_metadata("tabdpt.max_len", GGUFMetaValue::Uint32(config.max_len as u32));
128    writer.add_metadata("tabdpt.n_thinking_rows", GGUFMetaValue::Uint32(config.n_thinking_rows as u32));
129}
130
131fn ggml_type_from_st(dtype: StDtype) -> anyhow::Result<GGMLType> {
132    match dtype {
133        StDtype::F32 => Ok(GGMLType::F32),
134        StDtype::F16 => Ok(GGMLType::F16),
135        StDtype::BF16 => Ok(GGMLType::BF16),
136        other => anyhow::bail!("unsupported safetensors dtype: {other:?}"),
137    }
138}
139
140fn cast_data(data: &[u8], src: GGMLType, dst: GGMLType) -> anyhow::Result<Vec<u8>> {
141    if src == dst {
142        return Ok(data.to_vec());
143    }
144    if dst == GGMLType::Q8_0 {
145        let f32_values = decode_to_f32(data, src)?;
146        return quantize_q8_0(&f32_values);
147    }
148    match (src, dst) {
149        (GGMLType::F32, GGMLType::F16) => {
150            let vals = parse_f32_le(data)?;
151            let mut out = Vec::with_capacity(vals.len() * 2);
152            for v in vals {
153                out.extend_from_slice(&f32_to_f16_bits(v).to_le_bytes());
154            }
155            Ok(out)
156        }
157        (GGMLType::F32, GGMLType::BF16) => {
158            let vals = parse_f32_le(data)?;
159            let mut out = Vec::with_capacity(vals.len() * 2);
160            for v in vals {
161                out.extend_from_slice(&((v.to_bits() >> 16) as u16).to_le_bytes());
162            }
163            Ok(out)
164        }
165        (GGMLType::F16, GGMLType::BF16) => {
166            let mut out = Vec::with_capacity(data.len());
167            for c in data.chunks_exact(2) {
168                let f32_val = f16_to_f32(u16::from_le_bytes([c[0], c[1]]));
169                out.extend_from_slice(&((f32_val.to_bits() >> 16) as u16).to_le_bytes());
170            }
171            Ok(out)
172        }
173        (GGMLType::BF16, GGMLType::F32) => {
174            let mut out = Vec::with_capacity(data.len() * 2);
175            for c in data.chunks_exact(2) {
176                let bits = (u16::from_le_bytes([c[0], c[1]]) as u32) << 16;
177                out.extend_from_slice(&bits.to_le_bytes());
178            }
179            Ok(out)
180        }
181        (GGMLType::BF16, GGMLType::F16) => {
182            let mut out = Vec::with_capacity(data.len());
183            for c in data.chunks_exact(2) {
184                let bits = (u16::from_le_bytes([c[0], c[1]]) as u32) << 16;
185                out.extend_from_slice(&f32_to_f16_bits(f32::from_bits(bits)).to_le_bytes());
186            }
187            Ok(out)
188        }
189        (GGMLType::F16, GGMLType::F32) => {
190            let mut out = Vec::with_capacity(data.len() * 2);
191            for c in data.chunks_exact(2) {
192                out.extend_from_slice(&f16_to_f32(u16::from_le_bytes([c[0], c[1]])).to_bits().to_le_bytes());
193            }
194            Ok(out)
195        }
196        _ => anyhow::bail!("unsupported cast: {src:?} → {dst:?}"),
197    }
198}
199
200fn decode_to_f32(data: &[u8], src: GGMLType) -> anyhow::Result<Vec<f32>> {
201    match src {
202        GGMLType::F32 => parse_f32_le(data),
203        GGMLType::F16 => data
204            .chunks_exact(2)
205            .map(|c| Ok(f16_to_f32(u16::from_le_bytes([c[0], c[1]]))))
206            .collect(),
207        GGMLType::BF16 => data
208            .chunks_exact(2)
209            .map(|c| Ok(f32::from_bits((u16::from_le_bytes([c[0], c[1]]) as u32) << 16)))
210            .collect(),
211        GGMLType::Q8_0 => anyhow::bail!("Q8_0 as source not supported"),
212    }
213}
214
215fn quantize_q8_0(values: &[f32]) -> anyhow::Result<Vec<u8>> {
216    const BLOCK: usize = 32;
217    if values.len() % BLOCK != 0 {
218        anyhow::bail!("Q8_0 requires count divisible by {BLOCK}");
219    }
220    let n_blocks = values.len() / BLOCK;
221    let mut out = vec![0u8; n_blocks * 34];
222    for b in 0..n_blocks {
223        let blk = &values[b * BLOCK..(b + 1) * BLOCK];
224        let amax = blk.iter().copied().map(f32::abs).fold(0.0f32, f32::max);
225        let d = if amax == 0.0 { 0.0f32 } else { amax / 127.0 };
226        let d_inv = if d == 0.0 { 0.0f32 } else { 1.0 / d };
227        let base = b * 34;
228        out[base..base + 2].copy_from_slice(&f32_to_f16_bits(d).to_le_bytes());
229        for i in 0..BLOCK {
230            out[base + 2 + i] = (blk[i] * d_inv).round().clamp(-127.0, 127.0) as i8 as u8;
231        }
232    }
233    Ok(out)
234}
235
236fn parse_f32_le(data: &[u8]) -> anyhow::Result<Vec<f32>> {
237    if data.len() % 4 != 0 {
238        anyhow::bail!("f32 data length not divisible by 4");
239    }
240    Ok(data.chunks_exact(4).map(|c| f32::from_le_bytes([c[0], c[1], c[2], c[3]])).collect())
241}
242
243fn f32_to_f16_bits(v: f32) -> u16 {
244    let bits = v.to_bits();
245    let sign = ((bits >> 16) & 0x8000) as u16;
246    let exp = ((bits >> 23) & 0xFF) as i32;
247    let mantissa = bits & 0x007F_FFFF;
248    if exp == 0xFF {
249        return sign | 0x7C00 | if mantissa != 0 { 0x0200 } else { 0 };
250    }
251    let new_exp = exp - 127 + 15;
252    if new_exp >= 31 {
253        return sign | 0x7C00;
254    }
255    if new_exp <= 0 {
256        if new_exp < -10 {
257            return sign;
258        }
259        let m = (mantissa | 0x0080_0000) >> (1 - new_exp);
260        return sign | (m >> 13) as u16;
261    }
262    sign | ((new_exp as u16) << 10) | (mantissa >> 13) as u16
263}
264
265fn f16_to_f32(bits: u16) -> f32 {
266    let sign = ((bits & 0x8000) as u32) << 16;
267    let exp = ((bits >> 10) & 0x1F) as i32;
268    let mantissa = (bits & 0x03FF) as u32;
269    let f32_bits = if exp == 0 {
270        if mantissa == 0 {
271            sign
272        } else {
273            let mut m = mantissa;
274            let mut e = 0i32;
275            while m & 0x0400 == 0 {
276                m <<= 1;
277                e += 1;
278            }
279            sign | ((127 - 15 - e + 1) as u32) << 23 | (m & 0x03FF) << 13
280        }
281    } else if exp == 31 {
282        sign | 0x7F80_0000 | (mantissa << 13)
283    } else {
284        sign | ((exp + 127 - 15) as u32) << 23 | (mantissa << 13)
285    };
286    f32::from_bits(f32_bits)
287}