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::MoiraiConfig;
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: &MoiraiConfig,
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"))?;
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 let _outermost = py_shape.first().copied().unwrap_or(1);
72
73 let (dst_dtype, gguf_shape, tensor_data) = if opts.output_dtype == GGMLType::Q8_0
74 && (innermost % 32 != 0 || n_elems % 32 != 0)
75 {
76 fallback_count += 1;
77 let data = cast_data(raw_data, src_dtype, GGMLType::F32)?;
78 let gs = py_shape.iter().rev().map(|&d| d as u64).collect();
79 (GGMLType::F32, gs, data)
80 } else {
81 let dst = opts.output_dtype;
82 let data = cast_data(raw_data, src_dtype, dst)?;
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 } pb.finish_with_message("tensors processed");
94
95 if !skipped.is_empty() {
96 eprintln!("\nWarning: {} tensor(s) skipped:", 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)?;
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: &MoiraiConfig) {
114 writer.add_metadata(
115 "general.architecture",
116 GGUFMetaValue::String("moirai".into()),
117 );
118 writer.add_metadata(
119 "general.name",
120 GGUFMetaValue::String("Moirai-1.0-R-large".into()),
121 );
122 writer.add_metadata(
123 "moirai.d_model",
124 GGUFMetaValue::Uint32(config.d_model as u32),
125 );
126 writer.add_metadata(
127 "moirai.n_layers",
128 GGUFMetaValue::Uint32(config.n_layers as u32),
129 );
130 writer.add_metadata(
131 "moirai.n_heads",
132 GGUFMetaValue::Uint32(config.n_heads as u32),
133 );
134 writer.add_metadata(
135 "moirai.head_dim",
136 GGUFMetaValue::Uint32(config.head_dim as u32),
137 );
138 writer.add_metadata("moirai.d_ff", GGUFMetaValue::Uint32(config.d_ff as u32));
139 writer.add_metadata(
140 "moirai.max_seq_len",
141 GGUFMetaValue::Uint32(config.max_seq_len as u32),
142 );
143 writer.add_metadata(
144 "moirai.max_patch_size",
145 GGUFMetaValue::Uint32(config.max_patch_size as u32),
146 );
147}
148
149fn ggml_type_from_st(dtype: StDtype) -> anyhow::Result<GGMLType> {
150 match dtype {
151 StDtype::F32 => Ok(GGMLType::F32),
152 StDtype::F16 => Ok(GGMLType::F16),
153 StDtype::BF16 => Ok(GGMLType::BF16),
154 other => anyhow::bail!("unsupported dtype: {other:?}"),
155 }
156}
157
158fn cast_data(data: &[u8], src: GGMLType, dst: GGMLType) -> anyhow::Result<Vec<u8>> {
159 if src == dst {
160 return Ok(data.to_vec());
161 }
162 if dst == GGMLType::Q8_0 {
163 return quantize_q8_0(&decode_to_f32(data, src)?);
164 }
165 match (src, dst) {
166 (GGMLType::F32, GGMLType::F16) => {
167 let vals = parse_f32_le(data)?;
168 let mut out = Vec::with_capacity(vals.len() * 2);
169 for v in vals {
170 out.extend_from_slice(&f32_to_f16_bits(v).to_le_bytes());
171 }
172 Ok(out)
173 }
174 (GGMLType::F32, GGMLType::BF16) => {
175 let vals = parse_f32_le(data)?;
176 let mut out = Vec::with_capacity(vals.len() * 2);
177 for v in vals {
178 out.extend_from_slice(&((v.to_bits() >> 16) as u16).to_le_bytes());
179 }
180 Ok(out)
181 }
182 (GGMLType::F16, GGMLType::BF16) => {
183 let mut out = Vec::with_capacity(data.len());
184 for c in data.chunks_exact(2) {
185 let f32_val = f16_to_f32(u16::from_le_bytes([c[0], c[1]]));
186 out.extend_from_slice(&((f32_val.to_bits() >> 16) as u16).to_le_bytes());
187 }
188 Ok(out)
189 }
190 (GGMLType::BF16, GGMLType::F32) => {
191 let mut out = Vec::with_capacity(data.len() * 2);
192 for c in data.chunks_exact(2) {
193 let bits = (u16::from_le_bytes([c[0], c[1]]) as u32) << 16;
194 out.extend_from_slice(&bits.to_le_bytes());
195 }
196 Ok(out)
197 }
198 (GGMLType::BF16, GGMLType::F16) => {
199 let mut out = Vec::with_capacity(data.len());
200 for c in data.chunks_exact(2) {
201 let bits = (u16::from_le_bytes([c[0], c[1]]) as u32) << 16;
202 out.extend_from_slice(&f32_to_f16_bits(f32::from_bits(bits)).to_le_bytes());
203 }
204 Ok(out)
205 }
206 (GGMLType::F16, GGMLType::F32) => {
207 let mut out = Vec::with_capacity(data.len() * 2);
208 for c in data.chunks_exact(2) {
209 out.extend_from_slice(
210 &f16_to_f32(u16::from_le_bytes([c[0], c[1]]))
211 .to_bits()
212 .to_le_bytes(),
213 );
214 }
215 Ok(out)
216 }
217 _ => anyhow::bail!("unsupported cast: {src:?} → {dst:?}"),
218 }
219}
220
221fn decode_to_f32(data: &[u8], src: GGMLType) -> anyhow::Result<Vec<f32>> {
222 match src {
223 GGMLType::F32 => parse_f32_le(data),
224 GGMLType::F16 => data
225 .chunks_exact(2)
226 .map(|c| Ok(f16_to_f32(u16::from_le_bytes([c[0], c[1]]))))
227 .collect(),
228 GGMLType::BF16 => data
229 .chunks_exact(2)
230 .map(|c| {
231 Ok(f32::from_bits(
232 (u16::from_le_bytes([c[0], c[1]]) as u32) << 16,
233 ))
234 })
235 .collect(),
236 GGMLType::Q8_0 => anyhow::bail!("Q8_0 as source not supported"),
237 }
238}
239
240fn quantize_q8_0(values: &[f32]) -> anyhow::Result<Vec<u8>> {
241 const BLOCK: usize = 32;
242 if values.len() % BLOCK != 0 {
243 anyhow::bail!("Q8_0 requires divisibility by {BLOCK}");
244 }
245 let n_blocks = values.len() / BLOCK;
246 let mut out = vec![0u8; n_blocks * 34];
247 for b in 0..n_blocks {
248 let blk = &values[b * BLOCK..(b + 1) * BLOCK];
249 let amax = blk.iter().copied().map(f32::abs).fold(0.0f32, f32::max);
250 let d = if amax == 0.0 { 0.0f32 } else { amax / 127.0 };
251 let d_inv = if d == 0.0 { 0.0f32 } else { 1.0 / d };
252 let base = b * 34;
253 out[base..base + 2].copy_from_slice(&f32_to_f16_bits(d).to_le_bytes());
254 for i in 0..BLOCK {
255 out[base + 2 + i] = (blk[i] * d_inv).round().clamp(-127.0, 127.0) as i8 as u8;
256 }
257 }
258 Ok(out)
259}
260
261fn parse_f32_le(data: &[u8]) -> anyhow::Result<Vec<f32>> {
262 if data.len() % 4 != 0 {
263 anyhow::bail!("f32 data length not divisible by 4");
264 }
265 Ok(data
266 .chunks_exact(4)
267 .map(|c| f32::from_le_bytes([c[0], c[1], c[2], c[3]]))
268 .collect())
269}
270
271fn f32_to_f16_bits(v: f32) -> u16 {
272 let bits = v.to_bits();
273 let sign = ((bits >> 16) & 0x8000) as u16;
274 let exp = ((bits >> 23) & 0xFF) as i32;
275 let mantissa = bits & 0x007F_FFFF;
276 if exp == 0xFF {
277 return sign | 0x7C00 | if mantissa != 0 { 0x0200 } else { 0 };
278 }
279 let new_exp = exp - 127 + 15;
280 if new_exp >= 31 {
281 return sign | 0x7C00;
282 }
283 if new_exp <= 0 {
284 if new_exp < -10 {
285 return sign;
286 }
287 let m = (mantissa | 0x0080_0000) >> (1 - new_exp);
288 return sign | (m >> 13) as u16;
289 }
290 sign | ((new_exp as u16) << 10) | (mantissa >> 13) as u16
291}
292
293fn f16_to_f32(bits: u16) -> f32 {
294 let sign = ((bits & 0x8000) as u32) << 16;
295 let exp = ((bits >> 10) & 0x1F) as i32;
296 let mantissa = (bits & 0x03FF) as u32;
297 let f32_bits = if exp == 0 {
298 if mantissa == 0 {
299 sign
300 } else {
301 let mut m = mantissa;
302 let mut e = 0i32;
303 while m & 0x0400 == 0 {
304 m <<= 1;
305 e += 1;
306 }
307 sign | ((127 - 15 - e + 1) as u32) << 23 | (m & 0x03FF) << 13
308 }
309 } else if exp == 31 {
310 sign | 0x7F80_0000 | (mantissa << 13)
311 } else {
312 sign | ((exp + 127 - 15) as u32) << 23 | (mantissa << 13)
313 };
314 f32::from_bits(f32_bits)
315}