tract_linalg/frame/block_quant/
mod.rs

1use downcast_rs::{Downcast, impl_downcast};
2use dyn_clone::{DynClone, clone_box};
3use dyn_hash::DynHash;
4use num_traits::Zero;
5use tract_data::internal::*;
6use tract_data::itertools::Itertools;
7
8use std::alloc::Layout;
9use std::borrow::Cow;
10use std::fmt::{Debug, Display};
11use std::hash::Hash;
12use std::sync::Arc;
13
14mod helpers;
15mod q4_0;
16mod q8_1;
17mod value;
18
19pub use helpers::{NibbleReader, NibbleWriter};
20pub use q4_0::Q4_0;
21pub use q8_1::Q8_1;
22pub use value::{BlockQuantFact, PackedBlockQuantFact};
23
24use crate::mmm::{EagerPackedInput, MMMInputFormat};
25use crate::pack::PackedFormat;
26
27use crate::WeightType;
28
29use super::mmm::MMMInputValue;
30
31pub trait BlockQuant: Debug + Display + Send + Sync + DynClone + DynHash + Downcast {
32    fn same_as(&self, other: &dyn BlockQuant) -> bool;
33
34    fn block_len(&self) -> usize;
35
36    fn block_bytes(&self) -> usize;
37
38    fn dequant_block_f32(&self, quant: &[u8], block: &mut [f32]);
39    fn dequant_block_f16(&self, quant: &[u8], block: &mut [f16]);
40    fn quant_block_f16(&self, block: &[f16], quant: &mut [u8]);
41    fn quant_block_f32(&self, block: &[f32], quant: &mut [u8]);
42
43    fn quant_f16(&self, input: &[f16]) -> TractResult<Blob> {
44        unsafe {
45            let blocks = input.len() / self.block_len();
46            let mut quant = Blob::for_layout(
47                Layout::from_size_align(blocks * self.block_bytes(), 128).unwrap(),
48            );
49            for b in 0..blocks {
50                let block = &input[b * self.block_len()..][..self.block_len()];
51                let qblock = &mut quant[b * self.block_bytes()..][..self.block_bytes()];
52                self.quant_block_f16(block, qblock);
53            }
54            Ok(quant)
55        }
56    }
57
58    fn quant_f32(&self, input: &[f32]) -> TractResult<Blob> {
59        unsafe {
60            let blocks = input.len() / self.block_len();
61            let mut quant = Blob::for_layout(
62                Layout::from_size_align(blocks * self.block_bytes(), 128).unwrap(),
63            );
64            for b in 0..blocks {
65                let block = &input[b * self.block_len()..][..self.block_len()];
66                let qblock = &mut quant[b * self.block_bytes()..][..self.block_bytes()];
67                self.quant_block_f32(block, qblock);
68            }
69            Ok(quant)
70        }
71    }
72
73    fn dequant_f32(&self, input: &[u8]) -> TractResult<Tensor> {
74        unsafe {
75            let blocks = input.len() / self.block_bytes();
76            let mut tensor = Tensor::uninitialized::<f32>(&[blocks * self.block_len()])?;
77            let slice = tensor.as_slice_mut::<f32>()?;
78            for b in 0..blocks {
79                let block = &mut slice[b * self.block_len()..][..self.block_len()];
80                let qblock = &input[b * self.block_bytes()..][..self.block_bytes()];
81                self.dequant_block_f32(qblock, block);
82            }
83            Ok(tensor)
84        }
85    }
86
87    fn dequant_f16(&self, input: &[u8]) -> TractResult<Tensor> {
88        unsafe {
89            let blocks = input.len() / self.block_bytes();
90            let mut tensor = Tensor::uninitialized::<f16>(&[blocks * self.block_len()])?;
91            let slice = tensor.as_slice_mut::<f16>()?;
92            for b in 0..blocks {
93                let block = &mut slice[b * self.block_len()..][..self.block_len()];
94                let qblock = &input[b * self.block_bytes()..][..self.block_bytes()];
95                self.dequant_block_f16(qblock, block);
96            }
97            Ok(tensor)
98        }
99    }
100
101    fn extract_at_offset_f16(&self, input: &[u8], offset: usize) -> f16 {
102        let len = self.block_len();
103        let block_id = offset / len;
104        let mut block = vec![f16::zero(); self.block_len()];
105        self.dequant_block_f16(
106            &input[block_id * self.block_bytes()..][..self.block_bytes()],
107            &mut block,
108        );
109        block[offset % len]
110    }
111
112    fn extract_at_offset_f32(&self, input: &[u8], offset: usize) -> f32 {
113        let len = self.block_len();
114        let block_id = offset / len;
115        let mut block = vec![f32::zero(); self.block_len()];
116        self.dequant_block_f32(
117            &input[block_id * self.block_bytes()..][..self.block_bytes()],
118            &mut block,
119        );
120        block[offset % len]
121    }
122
123    fn simulate_precision_loss(
124        &self,
125        mut tensor: Tensor,
126        block_axis: usize,
127    ) -> TractResult<Tensor> {
128        ensure!(block_axis == tensor.rank() - 1);
129        ensure!(tensor.shape()[block_axis] % self.block_len() == 0);
130        let mut scratch = vec![0u8; self.block_bytes()];
131        if tensor.datum_type() == f32::datum_type() {
132            for block in tensor.as_slice_mut::<f32>()?.chunks_mut(self.block_len()) {
133                self.quant_block_f32(block, &mut scratch);
134                self.dequant_block_f32(&scratch, block);
135            }
136            Ok(tensor)
137        } else if tensor.datum_type() == f16::datum_type() {
138            for block in tensor.as_slice_mut::<f16>()?.chunks_mut(self.block_len()) {
139                self.quant_block_f16(block, &mut scratch);
140                self.dequant_block_f16(&scratch, block);
141            }
142            Ok(tensor)
143        } else {
144            todo!()
145        }
146    }
147
148    fn pack(
149        &self,
150        input: &[u8],
151        k: usize,
152        r: usize,
153        zip: usize,
154        scales_at_end: bool,
155    ) -> TractResult<EagerPackedInput>;
156
157    unsafe fn extract_packed_panel(
158        &self,
159        value: &EagerPackedInput,
160        target: &PackedFormat,
161        panel: usize,
162        scratch: *mut u8,
163    ) -> TractResult<()>;
164
165    fn extract_at_mn_f16(
166        &self,
167        value: &EagerPackedInput,
168        mn: usize,
169        target: &mut [f16],
170    ) -> TractResult<()>;
171
172    fn extract_at_mn_f32(
173        &self,
174        value: &EagerPackedInput,
175        mn: usize,
176        target: &mut [f32],
177    ) -> TractResult<()>;
178}
179
180dyn_clone::clone_trait_object!(BlockQuant);
181dyn_hash::hash_trait_object!(BlockQuant);
182impl_downcast!(BlockQuant);
183
184#[allow(clippy::derived_hash_with_manual_eq)]
185#[derive(Clone, Hash)]
186pub struct PackedBlockQuantFormat {
187    pub bq: Box<dyn BlockQuant>,
188    pub r: usize,
189    pub zip: usize,
190    pub scales_at_end: bool,
191}
192
193impl PartialEq for PackedBlockQuantFormat {
194    fn eq(&self, other: &Self) -> bool {
195        self.bq.same_as(&*other.bq)
196            && self.r == other.r
197            && self.zip == other.zip
198            && self.scales_at_end == other.scales_at_end
199    }
200}
201
202impl Eq for PackedBlockQuantFormat {}
203
204impl Display for PackedBlockQuantFormat {
205    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
206        write!(f, "Packed{}[{}]", &*self.bq, self.r)?;
207        if self.zip != 0 {
208            write!(f, "Z{}", self.zip)?;
209        }
210        if self.scales_at_end {
211            write!(f, "Se")?;
212        }
213        Ok(())
214    }
215}
216
217impl Debug for PackedBlockQuantFormat {
218    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
219        <Self as Display>::fmt(self, f)
220    }
221}
222
223impl PackedBlockQuantFormat {
224    pub fn new(bq: &dyn BlockQuant, r: usize, zip: usize, scales_at_end: bool) -> Self {
225        PackedBlockQuantFormat { bq: clone_box(bq), r, zip, scales_at_end }
226    }
227
228    pub fn simulate_precision_loss(
229        &self,
230        tensor: Tensor,
231        block_axis: usize,
232    ) -> TractResult<Tensor> {
233        self.bq.simulate_precision_loss(tensor, block_axis)
234    }
235
236    pub fn pack(&self, input: &[u8], k: usize) -> TractResult<EagerPackedInput> {
237        self.bq.pack(input, k, self.r, self.zip, self.scales_at_end)
238    }
239}
240
241impl MMMInputFormat for PackedBlockQuantFormat {
242    fn prepare_tensor(&self, t: &Tensor, _k_axis: usize, _mn_axis: usize) -> TractResult<Tensor> {
243        let packed = t
244            .as_slice::<Opaque>()?
245            .iter()
246            .map(|o| {
247                let bwf = o.downcast_ref::<BlobWithFact>().unwrap();
248                let bqf = bwf.fact.downcast_ref::<BlockQuantFact>().unwrap();
249                let packed = self.pack(&bwf.value, bqf.k())?;
250                Ok(Opaque(Arc::new(Box::new(packed) as Box<dyn MMMInputValue>)))
251            })
252            .collect::<TractResult<Vec<Opaque>>>()?;
253        tensor1(&packed).into_shape(t.shape())
254    }
255
256    fn prepare_one(
257        &self,
258        t: &Tensor,
259        k_axis: usize,
260        mn_axis: usize,
261    ) -> TractResult<Box<dyn MMMInputValue>> {
262        // this code path is essentially there for test scenarios
263        let t = if t.datum_type().is_number() {
264            let k = t.shape()[k_axis];
265            let m = t.shape()[mn_axis];
266            assert!(k % self.bq.block_len() == 0);
267            let t: Cow<Tensor> = if k_axis == 1 && mn_axis == 0 {
268                Cow::Borrowed(t)
269            } else {
270                Cow::Owned(t.clone().move_axis(1, 0)?)
271            };
272            let quant = if t.datum_type() == f32::datum_type() {
273                self.bq.quant_f32(t.as_slice()?)?
274            } else if t.datum_type() == f16::datum_type() {
275                self.bq.quant_f16(t.as_slice()?)?
276            } else {
277                todo!()
278            };
279            Cow::Owned(tensor0(Opaque(Arc::new(BlobWithFact {
280                value: Arc::new(quant),
281                fact: Box::new(BlockQuantFact::new(self.bq.clone(), tvec!(m, k))),
282            }))))
283        } else {
284            Cow::Borrowed(t)
285        };
286        ensure!(mn_axis == 0);
287        ensure!(k_axis == 1);
288        let bwf = t.to_scalar::<Opaque>()?.downcast_ref::<BlobWithFact>().unwrap();
289        let bqf = bwf.fact.downcast_ref::<BlockQuantFact>().unwrap();
290        let packed = self.pack(&bwf.value, bqf.k())?;
291        Ok(Box::new(packed))
292    }
293
294    fn precursor(&self) -> WeightType {
295        WeightType::BlockQuant(self.bq.clone())
296    }
297
298    fn k_alignment(&self) -> usize {
299        self.bq.block_len()
300    }
301
302    fn r(&self) -> usize {
303        self.r
304    }
305
306    fn mem_size(&self, k: TDim, mn: TDim) -> TDim {
307        k * mn * self.bq.block_bytes() / self.bq.block_len()
308    }
309
310    fn same_as(&self, other: &dyn MMMInputFormat) -> bool {
311        other.downcast_ref::<Self>().is_some_and(|other| self == other)
312    }
313
314    fn extract_at_mn_f16(
315        &self,
316        data: &EagerPackedInput,
317        mn: usize,
318        slice: &mut [f16],
319    ) -> TractResult<()> {
320        self.bq.extract_at_mn_f16(data, mn, slice)
321    }
322
323    fn extract_at_mn_f32(
324        &self,
325        data: &EagerPackedInput,
326        mn: usize,
327        slice: &mut [f32],
328    ) -> TractResult<()> {
329        self.bq.extract_at_mn_f32(data, mn, slice)
330    }
331}
tract_linalg/frame/block_quant/mod.rs

tract_linalg/frame/block_quant/
mod.rs
