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::{MitraConfig, Task};
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: &MitraConfig,
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()).with_context(|| {
65 format!(
66 "tensor {hf_name}: unsupported dtype {:?}",
67 tensor_view.dtype()
68 )
69 })?;
70
71 let raw_data = tensor_view.data();
72 let py_shape = tensor_view.shape();
73 let n_elems: usize = py_shape.iter().product();
74 let innermost = py_shape.last().copied().unwrap_or(1);
75
76 let (dst_dtype, gguf_shape, tensor_data) = if opts.output_dtype == GGMLType::Q8_0
77 && (innermost % 32 != 0 || n_elems % 32 != 0)
78 {
79 fallback_count += 1;
80 let data = cast_data(raw_data, src_dtype, GGMLType::F32)
81 .with_context(|| format!("tensor {hf_name}: cast failed"))?;
82 let gs = py_shape.iter().rev().map(|&d| d as u64).collect();
83 (GGMLType::F32, gs, data)
84 } else {
85 let dst = opts.output_dtype;
86 let data = cast_data(raw_data, src_dtype, dst)
87 .with_context(|| format!("tensor {hf_name}: cast failed"))?;
88 let gs = py_shape.iter().rev().map(|&d| d as u64).collect();
89 (dst, gs, data)
90 };
91
92 writer.add_tensor(gguf_name, gguf_shape, dst_dtype, tensor_data);
93 mapped += 1;
94 pb.inc(1);
95 }
96 }
97
98 pb.finish_with_message("tensors processed");
99
100 if !skipped.is_empty() {
101 eprintln!("\nWarning: {} tensor(s) skipped:", skipped.len());
102 for name in &skipped {
103 eprintln!(" {name}");
104 }
105 }
106 if fallback_count > 0 {
107 eprintln!("\nNote: {fallback_count} tensor(s) fell back to F32.");
108 }
109
110 println!("Writing {mapped} tensors to {} …", output_path.display());
111 let out_file =
112 File::create(output_path).with_context(|| format!("create {}", output_path.display()))?;
113 let mut buf_writer = BufWriter::new(out_file);
114 writer.write_to(&mut buf_writer)?;
115 println!("Done.");
116 Ok(())
117}
118
119fn write_metadata(writer: &mut GGUFWriter, config: &MitraConfig) {
120 writer.add_metadata(
121 "general.architecture",
122 GGUFMetaValue::String("mitra".into()),
123 );
124 writer.add_metadata(
125 "general.name",
126 GGUFMetaValue::String(
127 match config.task {
128 Task::Classification => "autogluon/mitra-classifier",
129 Task::Regression => "autogluon/mitra-regressor",
130 }
131 .into(),
132 ),
133 );
134 writer.add_metadata(
135 "task",
136 GGUFMetaValue::String(
137 match config.task {
138 Task::Classification => "classification",
139 Task::Regression => "regression",
140 }
141 .into(),
142 ),
143 );
144 writer.add_metadata("mitra.dim", GGUFMetaValue::Uint32(config.dim as u32));
145 writer.add_metadata(
146 "mitra.n_layers",
147 GGUFMetaValue::Uint32(config.n_layers as u32),
148 );
149 writer.add_metadata(
150 "mitra.n_heads",
151 GGUFMetaValue::Uint32(config.n_heads as u32),
152 );
153 writer.add_metadata(
154 "mitra.dim_output",
155 GGUFMetaValue::Uint32(config.dim_output as u32),
156 );
157}
158
159fn ggml_type_from_st(dtype: StDtype) -> anyhow::Result<GGMLType> {
160 match dtype {
161 StDtype::F32 => Ok(GGMLType::F32),
162 StDtype::F16 => Ok(GGMLType::F16),
163 StDtype::BF16 => Ok(GGMLType::BF16),
164 other => anyhow::bail!("unsupported safetensors dtype: {other:?}"),
165 }
166}
167
168fn cast_data(data: &[u8], src: GGMLType, dst: GGMLType) -> anyhow::Result<Vec<u8>> {
169 if src == dst {
170 return Ok(data.to_vec());
171 }
172 if dst == GGMLType::Q8_0 {
173 let f32_values = decode_to_f32(data, src)?;
174 return quantize_q8_0(&f32_values);
175 }
176 match (src, dst) {
177 (GGMLType::F32, GGMLType::F16) => {
178 let vals = parse_f32_le(data)?;
179 let mut out = Vec::with_capacity(vals.len() * 2);
180 for v in vals {
181 out.extend_from_slice(&f32_to_f16_bits(v).to_le_bytes());
182 }
183 Ok(out)
184 }
185 (GGMLType::F32, GGMLType::BF16) => {
186 let vals = parse_f32_le(data)?;
187 let mut out = Vec::with_capacity(vals.len() * 2);
188 for v in vals {
189 out.extend_from_slice(&((v.to_bits() >> 16) as u16).to_le_bytes());
190 }
191 Ok(out)
192 }
193 (GGMLType::F16, GGMLType::BF16) => {
194 let mut out = Vec::with_capacity(data.len());
195 for c in data.chunks_exact(2) {
196 let f32_val = f16_to_f32(u16::from_le_bytes([c[0], c[1]]));
197 out.extend_from_slice(&((f32_val.to_bits() >> 16) as u16).to_le_bytes());
198 }
199 Ok(out)
200 }
201 (GGMLType::BF16, GGMLType::F32) => {
202 let mut out = Vec::with_capacity(data.len() * 2);
203 for c in data.chunks_exact(2) {
204 let bits = (u16::from_le_bytes([c[0], c[1]]) as u32) << 16;
205 out.extend_from_slice(&bits.to_le_bytes());
206 }
207 Ok(out)
208 }
209 (GGMLType::BF16, GGMLType::F16) => {
210 let mut out = Vec::with_capacity(data.len());
211 for c in data.chunks_exact(2) {
212 let bits = (u16::from_le_bytes([c[0], c[1]]) as u32) << 16;
213 out.extend_from_slice(&f32_to_f16_bits(f32::from_bits(bits)).to_le_bytes());
214 }
215 Ok(out)
216 }
217 (GGMLType::F16, GGMLType::F32) => {
218 let mut out = Vec::with_capacity(data.len() * 2);
219 for c in data.chunks_exact(2) {
220 out.extend_from_slice(
221 &f16_to_f32(u16::from_le_bytes([c[0], c[1]]))
222 .to_bits()
223 .to_le_bytes(),
224 );
225 }
226 Ok(out)
227 }
228 _ => anyhow::bail!("unsupported cast: {src:?} → {dst:?}"),
229 }
230}
231
232fn decode_to_f32(data: &[u8], src: GGMLType) -> anyhow::Result<Vec<f32>> {
233 match src {
234 GGMLType::F32 => parse_f32_le(data),
235 GGMLType::F16 => data
236 .chunks_exact(2)
237 .map(|c| Ok(f16_to_f32(u16::from_le_bytes([c[0], c[1]]))))
238 .collect(),
239 GGMLType::BF16 => data
240 .chunks_exact(2)
241 .map(|c| {
242 Ok(f32::from_bits(
243 (u16::from_le_bytes([c[0], c[1]]) as u32) << 16,
244 ))
245 })
246 .collect(),
247 GGMLType::Q8_0 => anyhow::bail!("Q8_0 as source not supported"),
248 }
249}
250
251fn quantize_q8_0(values: &[f32]) -> anyhow::Result<Vec<u8>> {
252 const BLOCK: usize = 32;
253 if values.len() % BLOCK != 0 {
254 anyhow::bail!("Q8_0 requires count divisible by {BLOCK}");
255 }
256 let n_blocks = values.len() / BLOCK;
257 let mut out = vec![0u8; n_blocks * 34];
258 for b in 0..n_blocks {
259 let blk = &values[b * BLOCK..(b + 1) * BLOCK];
260 let amax = blk.iter().copied().map(f32::abs).fold(0.0f32, f32::max);
261 let d = if amax == 0.0 { 0.0f32 } else { amax / 127.0 };
262 let d_inv = if d == 0.0 { 0.0f32 } else { 1.0 / d };
263 let base = b * 34;
264 out[base..base + 2].copy_from_slice(&f32_to_f16_bits(d).to_le_bytes());
265 for i in 0..BLOCK {
266 out[base + 2 + i] = (blk[i] * d_inv).round().clamp(-127.0, 127.0) as i8 as u8;
267 }
268 }
269 Ok(out)
270}
271
272fn parse_f32_le(data: &[u8]) -> anyhow::Result<Vec<f32>> {
273 if data.len() % 4 != 0 {
274 anyhow::bail!("f32 data length not divisible by 4");
275 }
276 Ok(data
277 .chunks_exact(4)
278 .map(|c| f32::from_le_bytes([c[0], c[1], c[2], c[3]]))
279 .collect())
280}
281
282fn f32_to_f16_bits(v: f32) -> u16 {
283 let bits = v.to_bits();
284 let sign = ((bits >> 16) & 0x8000) as u16;
285 let exp = ((bits >> 23) & 0xFF) as i32;
286 let mantissa = bits & 0x007F_FFFF;
287 if exp == 0xFF {
288 return sign | 0x7C00 | if mantissa != 0 { 0x0200 } else { 0 };
289 }
290 let new_exp = exp - 127 + 15;
291 if new_exp >= 31 {
292 return sign | 0x7C00;
293 }
294 if new_exp <= 0 {
295 if new_exp < -10 {
296 return sign;
297 }
298 let m = (mantissa | 0x0080_0000) >> (1 - new_exp);
299 return sign | (m >> 13) as u16;
300 }
301 sign | ((new_exp as u16) << 10) | (mantissa >> 13) as u16
302}
303
304fn f16_to_f32(bits: u16) -> f32 {
305 let sign = ((bits & 0x8000) as u32) << 16;
306 let exp = ((bits >> 10) & 0x1F) as i32;
307 let mantissa = (bits & 0x03FF) as u32;
308 let f32_bits = if exp == 0 {
309 if mantissa == 0 {
310 sign
311 } else {
312 let mut m = mantissa;
313 let mut e = 0i32;
314 while m & 0x0400 == 0 {
315 m <<= 1;
316 e += 1;
317 }
318 sign | ((127 - 15 - e + 1) as u32) << 23 | (m & 0x03FF) << 13
319 }
320 } else if exp == 31 {
321 sign | 0x7F80_0000 | (mantissa << 13)
322 } else {
323 sign | ((exp + 127 - 15) as u32) << 23 | (mantissa << 13)
324 };
325 f32::from_bits(f32_bits)
326}