mistralrs_core/diffusion_models/clip/
text.rs

1#![allow(clippy::cast_possible_truncation, clippy::cast_precision_loss)]
2
3use candle_core::{DType, Device, IndexOp, Result, Tensor, D};
4use candle_nn as nn;
5use candle_nn::Module;
6use mistralrs_quant::ShardedVarBuilder;
7use serde::Deserialize;
8
9use crate::layers::{self, MatMul};
10
11#[derive(Debug, Clone, Copy, Deserialize)]
12pub enum Activation {
13    #[serde(rename = "quick_gelu")]
14    QuickGelu,
15}
16
17impl Module for Activation {
18    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
19        match self {
20            Activation::QuickGelu => xs * nn::ops::sigmoid(&(xs * 1.702f64)?)?,
21        }
22    }
23}
24
25#[derive(Debug, Clone, Deserialize)]
26pub struct ClipTextConfig {
27    pub vocab_size: usize,
28    pub projection_dim: usize,
29    pub hidden_act: Activation,
30    pub intermediate_size: usize,
31    pub max_position_embeddings: usize,
32    pub num_hidden_layers: usize,
33    pub num_attention_heads: usize,
34}
35
36#[derive(Debug, Clone, Deserialize)]
37pub struct ClipConfig {
38    pub text_config: ClipTextConfig,
39}
40
41// ClipTextEmbeddings mostly based on the existing implementation in the stable diffision model.
42// TODO rewrite to be more similar to https://github.com/huggingface/transformers/blob/f6fa0f0bf0796ac66f201f23bdb8585de1609add/src/transformers/models/clip/modeling_clip.py#L142
43#[derive(Clone, Debug)]
44struct ClipTextEmbeddings {
45    token_embedding: candle_nn::Embedding,
46    position_embedding: candle_nn::Embedding,
47    position_ids: Tensor,
48}
49
50impl ClipTextEmbeddings {
51    fn new(vs: ShardedVarBuilder, c: &ClipTextConfig) -> Result<Self> {
52        let token_embedding =
53            layers::embedding(c.vocab_size, c.projection_dim, vs.pp("token_embedding"))?;
54        let position_embedding: nn::Embedding = layers::embedding(
55            c.max_position_embeddings,
56            c.projection_dim,
57            vs.pp("position_embedding"),
58        )?;
59        let position_ids =
60            Tensor::arange(0u32, c.max_position_embeddings as u32, vs.device())?.unsqueeze(0)?;
61        Ok(ClipTextEmbeddings {
62            token_embedding,
63            position_embedding,
64            position_ids,
65        })
66    }
67}
68
69impl Module for ClipTextEmbeddings {
70    fn forward(&self, input_ids: &Tensor) -> Result<Tensor> {
71        let seq_length = input_ids.dim(D::Minus1)?;
72        let inputs_embeds = self.token_embedding.forward(input_ids)?;
73        let position_ids = self.position_ids.narrow(1, 0, seq_length)?;
74        let position_embedding = self.position_embedding.forward(&position_ids)?;
75        inputs_embeds.broadcast_add(&position_embedding)
76    }
77}
78
79#[derive(Clone, Debug)]
80struct ClipAttention {
81    k_proj: candle_nn::Linear,
82    v_proj: candle_nn::Linear,
83    q_proj: candle_nn::Linear,
84    out_proj: candle_nn::Linear,
85    head_dim: usize,
86    scale: f64,
87    num_attention_heads: usize,
88}
89
90impl ClipAttention {
91    fn new(vs: ShardedVarBuilder, c: &ClipTextConfig) -> Result<Self> {
92        let projection_dim = c.projection_dim;
93        let num_attention_heads = c.num_attention_heads;
94        let k_proj = layers::linear(projection_dim, projection_dim, vs.pp("k_proj"))?;
95        let v_proj = layers::linear(projection_dim, projection_dim, vs.pp("v_proj"))?;
96        let q_proj = layers::linear(projection_dim, projection_dim, vs.pp("q_proj"))?;
97        let out_proj = layers::linear(projection_dim, projection_dim, vs.pp("out_proj"))?;
98        let head_dim = projection_dim / num_attention_heads;
99        let scale = (head_dim as f64).powf(-0.5);
100
101        Ok(ClipAttention {
102            k_proj,
103            v_proj,
104            q_proj,
105            out_proj,
106            head_dim,
107            scale,
108            num_attention_heads,
109        })
110    }
111
112    fn shape(&self, xs: &Tensor, seq_len: usize, bsz: usize) -> Result<Tensor> {
113        xs.reshape((bsz, seq_len, self.num_attention_heads, self.head_dim))?
114            .transpose(1, 2)?
115            .contiguous()
116    }
117
118    fn forward(&self, xs: &Tensor, causal_attention_mask: Option<&Tensor>) -> Result<Tensor> {
119        let in_dtype = xs.dtype();
120        let (bsz, seq_len, projection_dim) = xs.dims3()?;
121
122        let query_states = (self.q_proj.forward(xs)? * self.scale)?;
123        let proj_shape = (bsz * self.num_attention_heads, seq_len, self.head_dim);
124        let query_states = self
125            .shape(&query_states, seq_len, bsz)?
126            .reshape(proj_shape)?
127            .to_dtype(DType::F32)?;
128        let key_states = self
129            .shape(&self.k_proj.forward(xs)?, seq_len, bsz)?
130            .reshape(proj_shape)?
131            .to_dtype(DType::F32)?;
132        let value_states = self
133            .shape(&self.v_proj.forward(xs)?, seq_len, bsz)?
134            .reshape(proj_shape)?
135            .to_dtype(DType::F32)?;
136        let attn_weights = MatMul.matmul(&query_states, &key_states.transpose(1, 2)?)?;
137
138        let src_len = key_states.dim(1)?;
139
140        let attn_weights = if let Some(causal_attention_mask) = causal_attention_mask {
141            attn_weights
142                .reshape((bsz, self.num_attention_heads, seq_len, src_len))?
143                .broadcast_add(causal_attention_mask)?
144                .reshape((bsz * self.num_attention_heads, seq_len, src_len))?
145        } else {
146            attn_weights
147        };
148
149        let attn_weights = candle_nn::ops::softmax(&attn_weights, D::Minus1)?;
150
151        let attn_output = MatMul
152            .matmul(&attn_weights, &value_states)?
153            .to_dtype(in_dtype)?;
154        let attn_output = attn_output
155            .reshape((bsz, self.num_attention_heads, seq_len, self.head_dim))?
156            .transpose(1, 2)?
157            .reshape((bsz, seq_len, projection_dim))?;
158        self.out_proj.forward(&attn_output)
159    }
160}
161
162#[derive(Clone, Debug)]
163struct ClipMlp {
164    fc1: candle_nn::Linear,
165    fc2: candle_nn::Linear,
166    activation: Activation,
167}
168
169impl ClipMlp {
170    fn new(vs: ShardedVarBuilder, c: &ClipTextConfig) -> Result<Self> {
171        let fc1 = layers::linear(c.projection_dim, c.intermediate_size, vs.pp("fc1"))?;
172        let fc2 = layers::linear(c.intermediate_size, c.projection_dim, vs.pp("fc2"))?;
173
174        Ok(ClipMlp {
175            fc1,
176            fc2,
177            activation: c.hidden_act,
178        })
179    }
180}
181
182impl ClipMlp {
183    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
184        let xs = self.fc1.forward(xs)?;
185        self.fc2.forward(&self.activation.forward(&xs)?)
186    }
187}
188
189#[derive(Clone, Debug)]
190struct ClipEncoderLayer {
191    self_attn: ClipAttention,
192    layer_norm1: candle_nn::LayerNorm,
193    mlp: ClipMlp,
194    layer_norm2: candle_nn::LayerNorm,
195}
196
197impl ClipEncoderLayer {
198    fn new(vs: ShardedVarBuilder, c: &ClipTextConfig) -> Result<Self> {
199        let self_attn = ClipAttention::new(vs.pp("self_attn"), c)?;
200        let layer_norm1 = layers::layer_norm(c.projection_dim, 1e-5, vs.pp("layer_norm1"))?;
201        let mlp = ClipMlp::new(vs.pp("mlp"), c)?;
202        let layer_norm2 = layers::layer_norm(c.projection_dim, 1e-5, vs.pp("layer_norm2"))?;
203
204        Ok(ClipEncoderLayer {
205            self_attn,
206            layer_norm1,
207            mlp,
208            layer_norm2,
209        })
210    }
211
212    fn forward(&self, xs: &Tensor, causal_attention_mask: Option<&Tensor>) -> Result<Tensor> {
213        let residual = xs;
214        let xs = self.layer_norm1.forward(xs)?;
215        let xs = self.self_attn.forward(&xs, causal_attention_mask)?;
216        let xs = (xs + residual)?;
217
218        let residual = &xs;
219        let xs = self.layer_norm2.forward(&xs)?;
220        let xs = self.mlp.forward(&xs)?;
221        xs + residual
222    }
223}
224
225#[derive(Clone, Debug)]
226pub struct ClipEncoder {
227    layers: Vec<ClipEncoderLayer>,
228}
229
230impl ClipEncoder {
231    pub fn new(vs: ShardedVarBuilder, c: &ClipTextConfig) -> Result<Self> {
232        let vs = vs.pp("layers");
233        let mut layers: Vec<ClipEncoderLayer> = Vec::new();
234        for index in 0..c.num_hidden_layers {
235            let layer = ClipEncoderLayer::new(vs.pp(index.to_string()), c)?;
236            layers.push(layer)
237        }
238        Ok(ClipEncoder { layers })
239    }
240
241    pub fn forward(&self, xs: &Tensor, causal_attention_mask: Option<&Tensor>) -> Result<Tensor> {
242        let mut xs = xs.clone();
243        for layer in self.layers.iter() {
244            xs = layer.forward(&xs, causal_attention_mask)?;
245        }
246        Ok(xs)
247    }
248}
249
250/// A CLIP transformer based model.
251#[derive(Clone, Debug)]
252pub struct ClipTextTransformer {
253    embeddings: ClipTextEmbeddings,
254    encoder: ClipEncoder,
255    final_layer_norm: candle_nn::LayerNorm,
256}
257
258impl ClipTextTransformer {
259    pub fn new(vs: ShardedVarBuilder, c: &ClipTextConfig) -> Result<Self> {
260        let embeddings = ClipTextEmbeddings::new(vs.pp("embeddings"), c)?;
261        let encoder = ClipEncoder::new(vs.pp("encoder"), c)?;
262        let final_layer_norm =
263            layers::layer_norm(c.projection_dim, 1e-5, vs.pp("final_layer_norm"))?;
264        Ok(ClipTextTransformer {
265            embeddings,
266            encoder,
267            final_layer_norm,
268        })
269    }
270
271    // TODO: rewrrite to newer version
272    fn build_causal_attention_mask(
273        bsz: usize,
274        seq_len: usize,
275        mask_after: usize,
276        device: &Device,
277    ) -> Result<Tensor> {
278        let mask: Vec<_> = (0..seq_len)
279            .flat_map(|i| {
280                (0..seq_len).map(move |j| {
281                    if j > i || j > mask_after {
282                        f32::MIN
283                    } else {
284                        0.
285                    }
286                })
287            })
288            .collect();
289        let mask = Tensor::from_slice(&mask, (seq_len, seq_len), device)?;
290        mask.broadcast_as((bsz, 1, seq_len, seq_len))
291    }
292
293    pub fn forward_with_mask(&self, input_ids: &Tensor, mask_after: usize) -> Result<Tensor> {
294        let (bsz, seq_len) = input_ids.dims2()?;
295        let input_ids = self.embeddings.forward(input_ids)?;
296        let causal_attention_mask =
297            Self::build_causal_attention_mask(bsz, seq_len, mask_after, input_ids.device())?;
298        let input_ids = self
299            .encoder
300            .forward(&input_ids, Some(&causal_attention_mask))?;
301        self.final_layer_norm.forward(&input_ids)
302    }
303}
304
305impl Module for ClipTextTransformer {
306    fn forward(&self, input_ids: &Tensor) -> Result<Tensor> {
307        let output = self.forward_with_mask(input_ids, usize::MAX)?;
308        let sequence_max_indices = input_ids.argmax(D::Minus1)?.to_dtype(DType::I64)?;
309
310        let mut indices = Vec::new();
311        for (batch_idx, &seq_idx) in sequence_max_indices.to_vec1::<i64>()?.iter().enumerate() {
312            let index = output.i((batch_idx, seq_idx as usize))?.unsqueeze(0)?;
313            indices.push(index);
314        }
315        Tensor::cat(&indices, 0)
316    }
317}
mistralrs_core/diffusion_models/clip/text.rs

mistralrs_core/diffusion_models/clip/
text.rs
