mistralrs_core/models/
mistral.rs

1#![allow(clippy::cast_possible_truncation, clippy::cast_precision_loss)]
2
3/// Mistral LLM, https://github.com/mistralai/mistral-src
4use candle_core::{Device, Module, Result, Tensor};
5use mistralrs_quant::{
6    ColumnParallelLayer, QuantMethod, QuantizedConfig, ReplicatedLayer, RowParallelLayer,
7    ShardedVarBuilder,
8};
9use serde::{Deserialize, Serialize};
10use std::{collections::HashMap, sync::Arc};
11
12use crate::{
13    amoe::{AnyMoeBaseModelMixin, AnyMoeConfig, AnyMoeExpertType, MlpLayer, MoeMlp},
14    attention::SdpaParams,
15    device_map::DeviceMapper,
16    get_delta_from_lora_ab,
17    layers::{embedding, Activation, CausalMasker, MatMul, Mlp, RmsNorm, RotaryEmbedding, Sdpa},
18    layers_masker::PastKvLenCache,
19    paged_attention::{AttentionImplementation, ModelConfigMetadata, PagedAttention},
20    pipeline::{
21        extract_logits,
22        text_models_inputs_processor::{FlashParams, PagedAttentionInputMetadata},
23        EitherCache, IsqModel, KvCache, NormalCache, NormalLoadingMetadata, NormalModel,
24    },
25    serde_default_fn,
26    utils::{progress::NiceProgressBar, unvarbuilder::UnVarBuilder},
27};
28
29serde_default_fn!(bool, use_flash_attn, false);
30serde_default_fn!(bool, tie_word_embeddings, false);
31
32#[derive(Debug, Clone, Default, Serialize, Deserialize)]
33pub struct Config {
34    pub(crate) vocab_size: usize,
35    pub(crate) hidden_size: usize,
36    pub(crate) intermediate_size: usize,
37    pub(crate) num_hidden_layers: usize,
38    pub(crate) num_attention_heads: usize,
39    pub(crate) num_key_value_heads: usize,
40    pub(crate) hidden_act: Activation,
41    pub(crate) max_position_embeddings: usize,
42    pub(crate) rms_norm_eps: f64,
43    pub(crate) rope_theta: f64,
44    pub(crate) sliding_window: Option<usize>,
45    #[serde(default = "use_flash_attn")]
46    pub(crate) use_flash_attn: bool,
47    pub(crate) head_dim: Option<usize>,
48    pub(crate) quantization_config: Option<QuantizedConfig>,
49    #[serde(default = "tie_word_embeddings")]
50    pub(crate) tie_word_embeddings: bool,
51}
52
53impl Config {
54    pub(crate) fn head_dim(&self) -> usize {
55        self.head_dim
56            .unwrap_or(self.hidden_size / self.num_attention_heads)
57    }
58}
59
60struct Attention {
61    q_proj: Arc<dyn QuantMethod>,
62    k_proj: Arc<dyn QuantMethod>,
63    v_proj: Arc<dyn QuantMethod>,
64    o_proj: Arc<dyn QuantMethod>,
65    num_heads: usize,
66    num_kv_heads: usize,
67    head_dim: usize,
68    rotary_emb: Arc<RotaryEmbedding>,
69    paged_attn: Option<PagedAttention>,
70    sdpa_params: SdpaParams,
71}
72
73impl Attention {
74    fn new(
75        rotary_emb: Arc<RotaryEmbedding>,
76        cfg: &Config,
77        vb: ShardedVarBuilder,
78        paged_attn: Option<PagedAttention>,
79        comm: &Arc<mistralrs_quant::Comm>,
80    ) -> Result<Self> {
81        let hidden_sz = cfg.hidden_size;
82        let num_heads = cfg.num_attention_heads;
83        let num_kv_heads = cfg.num_key_value_heads;
84        let head_dim = cfg.head_dim();
85        let q_proj = ColumnParallelLayer::new(
86            hidden_sz,
87            num_heads * head_dim,
88            &cfg.quantization_config,
89            false,
90            comm,
91            vb.pp("q_proj"),
92        )?;
93        let kv_shard = mistralrs_quant::compute_kv_shard(
94            cfg.num_key_value_heads,
95            cfg.hidden_size / cfg.num_attention_heads,
96            comm,
97        );
98        let k_proj = ColumnParallelLayer::new_with_shard(
99            hidden_sz,
100            num_kv_heads * head_dim,
101            &cfg.quantization_config,
102            false,
103            comm,
104            kv_shard,
105            vb.pp("k_proj"),
106        )?;
107        let v_proj = ColumnParallelLayer::new_with_shard(
108            hidden_sz,
109            num_kv_heads * head_dim,
110            &cfg.quantization_config,
111            false,
112            comm,
113            kv_shard,
114            vb.pp("v_proj"),
115        )?;
116        let o_proj = RowParallelLayer::new(
117            num_heads * head_dim,
118            hidden_sz,
119            &cfg.quantization_config,
120            false,
121            comm,
122            vb.pp("o_proj"),
123        )?;
124        Ok(Self {
125            q_proj,
126            k_proj,
127            v_proj,
128            o_proj,
129            num_heads: num_heads / comm.world_size(),
130            num_kv_heads: (num_kv_heads / comm.world_size()).max(1),
131            head_dim,
132            rotary_emb,
133            paged_attn,
134            sdpa_params: SdpaParams {
135                n_kv_groups: mistralrs_quant::compute_n_kv_groups(
136                    cfg.num_key_value_heads,
137                    cfg.num_attention_heads,
138                    comm,
139                ),
140                use_flash_attn: cfg.use_flash_attn,
141                softcap: None,
142                softmax_scale: 1.0 / (head_dim as f32).sqrt(),
143                sliding_window: cfg.sliding_window,
144            },
145        })
146    }
147
148    #[allow(clippy::too_many_arguments)]
149    fn forward(
150        &self,
151        xs: &Tensor,
152        attention_mask: Option<&Tensor>,
153        seqlen_offsets: &[usize],
154        kv_cache: &mut KvCache,
155        metadata: Option<((Tensor, Tensor), &PagedAttentionInputMetadata)>,
156        flash_params: &FlashParams,
157    ) -> Result<Tensor> {
158        let (b_sz, q_len, _) = xs.dims3()?;
159
160        let original_dtype = xs.dtype();
161        let mut xs = xs.clone();
162        if let Some(t) = self.q_proj.quantized_act_type() {
163            xs = xs.to_dtype(t)?;
164        }
165        let mut q = MatMul.qmethod_matmul(&xs, &*self.q_proj)?;
166        let mut k = MatMul.qmethod_matmul(&xs, &*self.k_proj)?;
167        let mut v = MatMul.qmethod_matmul(&xs, &*self.v_proj)?;
168        if self.q_proj.quantized_act_type().is_some() {
169            q = q.to_dtype(original_dtype)?;
170            k = k.to_dtype(original_dtype)?;
171            v = v.to_dtype(original_dtype)?;
172        }
173
174        let (q, k, v) = if q_len != 1 {
175            let q = q
176                .reshape((b_sz, q_len, self.num_heads, self.head_dim))?
177                .transpose(1, 2)?;
178            let k = k
179                .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
180                .transpose(1, 2)?;
181            let v = v
182                .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
183                .transpose(1, 2)?;
184            (q, k, v)
185        } else {
186            let q = q.reshape((b_sz, self.num_heads, q_len, self.head_dim))?;
187            let k = k.reshape((b_sz, self.num_kv_heads, q_len, self.head_dim))?;
188            let v = v.reshape((b_sz, self.num_kv_heads, q_len, self.head_dim))?;
189            (q, k, v)
190        };
191
192        let (q, k) = self.rotary_emb.forward(&q, &k, seqlen_offsets)?;
193
194        let mut attn_output = match &self.paged_attn {
195            Some(paged_attn) => match metadata {
196                Some(((key_cache, value_cache), input_metadata)) => paged_attn.forward(
197                    &q,
198                    &k,
199                    &v,
200                    attention_mask,
201                    Some(key_cache),
202                    Some(value_cache),
203                    input_metadata,
204                    &self.sdpa_params,
205                    Some(flash_params),
206                )?,
207                None => {
208                    // If we don't have metadata, we are most likely generating an imatrix so we don't want to populate that.
209                    // Generating the dummy metadata with the assumption that we are not generating text (only processing prompts).
210                    let input_metadata = PagedAttentionInputMetadata::dummy(q.device())?;
211                    // Sanity check.
212                    assert!(attention_mask.is_some());
213                    paged_attn.forward(
214                        &q,
215                        &k,
216                        &v,
217                        attention_mask,
218                        None,
219                        None,
220                        &input_metadata,
221                        &self.sdpa_params,
222                        Some(flash_params),
223                    )?
224                }
225            },
226            None => {
227                let (k, v) = kv_cache.append(&k, &v)?;
228
229                Sdpa.run_attention(
230                    &q,
231                    &k,
232                    &v,
233                    attention_mask,
234                    Some(flash_params),
235                    &self.sdpa_params,
236                )?
237            }
238        };
239
240        if let Some(t) = self.q_proj.quantized_act_type() {
241            attn_output = attn_output.to_dtype(t)?;
242        }
243        attn_output = if attention_mask.is_some() {
244            attn_output.transpose(1, 2)?.reshape((b_sz, q_len, ()))?
245        } else {
246            attn_output.reshape((b_sz, q_len, ()))?
247        };
248        let mut res = MatMul.qmethod_matmul(&attn_output, &*self.o_proj)?;
249        if self.q_proj.quantized_act_type().is_some() {
250            res = res.to_dtype(original_dtype)?;
251        }
252        Ok(res)
253    }
254}
255
256struct DecoderLayer {
257    self_attn: Attention,
258    mlp: Box<dyn MlpLayer>,
259    input_layernorm: RmsNorm,
260    post_attention_layernorm: RmsNorm,
261}
262
263impl DecoderLayer {
264    #[allow(clippy::too_many_arguments)]
265    fn new(
266        rotary_emb: Arc<RotaryEmbedding>,
267        cfg: &Config,
268        vb: ShardedVarBuilder,
269        mapper: &dyn DeviceMapper,
270        layer_idx: usize,
271        loading_isq: bool,
272        paged_attn: Option<PagedAttention>,
273        comm: &Arc<mistralrs_quant::Comm>,
274    ) -> Result<Self> {
275        let self_attn = Attention::new(
276            rotary_emb,
277            cfg,
278            mapper.set_device(layer_idx, vb.pp("self_attn"), loading_isq),
279            paged_attn,
280            comm,
281        )?;
282        let mlp = Mlp::new(
283            mapper.set_device(layer_idx, vb.pp("mlp"), loading_isq),
284            cfg.hidden_size,
285            cfg.intermediate_size,
286            &cfg.quantization_config,
287            cfg.hidden_act,
288            comm,
289        )?;
290        let input_layernorm = RmsNorm::new(
291            cfg.hidden_size,
292            cfg.rms_norm_eps,
293            mapper.set_device(layer_idx, vb.pp("input_layernorm"), false),
294        )?;
295        let post_attention_layernorm = RmsNorm::new(
296            cfg.hidden_size,
297            cfg.rms_norm_eps,
298            mapper.set_device(layer_idx, vb.pp("post_attention_layernorm"), false),
299        )?;
300        Ok(Self {
301            self_attn,
302            mlp: Box::new(mlp),
303            input_layernorm,
304            post_attention_layernorm,
305        })
306    }
307
308    #[allow(clippy::too_many_arguments)]
309    fn forward(
310        &self,
311        xs: &Tensor,
312        attention_mask: Option<&Tensor>,
313        seqlen_offsets: &[usize],
314        kv_cache: &mut KvCache,
315        metadata: Option<((Tensor, Tensor), &PagedAttentionInputMetadata)>,
316        flash_params: &FlashParams,
317    ) -> Result<Tensor> {
318        let residual = xs;
319        let xs = self.input_layernorm.forward(xs)?;
320        let xs = self.self_attn.forward(
321            &xs,
322            attention_mask,
323            seqlen_offsets,
324            kv_cache,
325            metadata,
326            flash_params,
327        )?;
328        let xs = (xs + residual)?;
329        let residual = &xs;
330        let xs = self
331            .mlp
332            .forward(&xs.apply(&self.post_attention_layernorm)?)?;
333        residual + xs
334    }
335}
336
337pub struct Model {
338    embed_tokens: candle_nn::Embedding,
339    layers: Vec<DecoderLayer>,
340    norm: RmsNorm,
341    lm_head: Arc<dyn QuantMethod>,
342    sliding_window: Option<usize>,
343    device: Device,
344    pub(crate) cache: EitherCache,
345    max_seq_len: usize,
346    mapper: Box<dyn DeviceMapper + Send + Sync>,
347    cfg: ModelConfigMetadata,
348}
349
350impl Model {
351    pub fn new(
352        cfg: &Config,
353        vb: ShardedVarBuilder,
354        is_gptx: bool,
355        normal_loading_metadata: NormalLoadingMetadata,
356        attention_mechanism: AttentionImplementation,
357    ) -> Result<Self> {
358        let vb_m = vb.pp("model");
359        let vb_lm_head = vb.pp("lm_head");
360        Self::new_inner(
361            cfg,
362            vb_m,
363            vb_lm_head,
364            is_gptx,
365            normal_loading_metadata,
366            attention_mechanism,
367        )
368    }
369
370    pub fn new_inner(
371        cfg: &Config,
372        vb_m: ShardedVarBuilder,
373        vb_lm_head: ShardedVarBuilder,
374        is_gptx: bool,
375        normal_loading_metadata: NormalLoadingMetadata,
376        attention_mechanism: AttentionImplementation,
377    ) -> Result<Self> {
378        if let Some(ref quant_cfg) = &cfg.quantization_config {
379            tracing::info!(
380                "Using {} quantization: {}.",
381                quant_cfg.quant_method.to_string(),
382                quant_cfg.get_bits_name(&vb_m)
383            );
384        }
385        let mapper = normal_loading_metadata.mapper;
386
387        let embed_tokens = embedding(
388            cfg.vocab_size,
389            cfg.hidden_size,
390            mapper.set_nm_device(vb_m.pp("embed_tokens"), false),
391        )?;
392
393        let head_dim = cfg.head_dim();
394        let mut ropes = HashMap::new();
395        for layer_idx in 0..cfg.num_hidden_layers {
396            let device = mapper
397                .device_for(layer_idx, false)
398                .unwrap_or(&normal_loading_metadata.real_device);
399            ropes.insert(
400                device.location(),
401                Arc::new(RotaryEmbedding::new(
402                    cfg.rope_theta as f32,
403                    head_dim,
404                    cfg.max_position_embeddings,
405                    device,
406                    is_gptx,
407                    vb_m.dtype(),
408                )?),
409            );
410        }
411
412        let mut layers = Vec::with_capacity(cfg.num_hidden_layers);
413        let vb_l = vb_m.pp("layers");
414        for layer_idx in NiceProgressBar::<_, 'b'>(
415            0..cfg.num_hidden_layers,
416            "Loading repeating layers",
417            &normal_loading_metadata.multi_progress,
418        ) {
419            let device = mapper
420                .device_for(layer_idx, false)
421                .unwrap_or(&normal_loading_metadata.real_device);
422            let rotary_emb = ropes
423                .get(&device.location())
424                .expect("No RoPE for device location!")
425                .clone();
426            let paged_attn = match &attention_mechanism {
427                AttentionImplementation::Eager => None,
428                AttentionImplementation::PagedAttention => {
429                    Some(PagedAttention::new(head_dim, device, None)?)
430                }
431            };
432            let comm = mapper.get_comm_for(layer_idx)?;
433            let layer = DecoderLayer::new(
434                rotary_emb.clone(),
435                cfg,
436                vb_l.pp(layer_idx),
437                &*mapper,
438                layer_idx,
439                normal_loading_metadata.loading_isq,
440                paged_attn,
441                &comm,
442            )?;
443            layers.push(layer)
444        }
445        let norm = RmsNorm::new(
446            cfg.hidden_size,
447            cfg.rms_norm_eps,
448            mapper.set_nm_device(vb_m.pp("norm"), false),
449        )?;
450        let lm_head = if !cfg.tie_word_embeddings {
451            ReplicatedLayer::new(
452                cfg.hidden_size,
453                cfg.vocab_size,
454                &None,
455                false,
456                mapper.set_nm_device(vb_lm_head, normal_loading_metadata.loading_isq),
457            )?
458        } else {
459            ReplicatedLayer::from_linear(candle_nn::Linear::new(
460                mapper.cast_nm_device(
461                    embed_tokens.embeddings(),
462                    normal_loading_metadata.loading_isq,
463                )?,
464                None,
465            ))?
466        };
467        Ok(Self {
468            embed_tokens,
469            layers,
470            norm,
471            lm_head,
472            sliding_window: cfg.sliding_window,
473            device: normal_loading_metadata.real_device,
474            cache: EitherCache::Normal(NormalCache::new_sliding(
475                cfg.num_hidden_layers,
476                cfg.max_position_embeddings,
477                cfg.sliding_window,
478            )),
479            max_seq_len: cfg.max_position_embeddings,
480            cfg: ModelConfigMetadata {
481                max_seq_len: cfg.max_position_embeddings,
482                num_layers: cfg.num_hidden_layers,
483                hidden_size: cfg.hidden_size,
484                num_kv_heads: (cfg.num_key_value_heads / mapper.get_comm_for(0)?.world_size())
485                    .max(1),
486                num_attn_heads: cfg.num_attention_heads / mapper.get_comm_for(0)?.world_size(),
487                sliding_window: cfg.sliding_window,
488                k_head_dim: cfg.head_dim(),
489                v_head_dim: cfg.head_dim(),
490            },
491            mapper,
492        })
493    }
494
495    pub fn get_input_embeddings(&self, input_ids: &Tensor) -> Result<Tensor> {
496        self.embed_tokens.forward(input_ids)
497    }
498
499    pub fn forward(
500        &self,
501        input_ids: &Tensor,
502        seqlen_offsets: &[usize],
503        context_lens: Vec<(usize, usize)>,
504        metadata: Option<(Vec<(Tensor, Tensor)>, &PagedAttentionInputMetadata)>,
505        flash_params: &FlashParams,
506    ) -> Result<Tensor> {
507        self.forward_embeds(
508            input_ids,
509            self.embed_tokens.forward(input_ids)?,
510            seqlen_offsets,
511            context_lens,
512            metadata,
513            flash_params,
514        )
515    }
516
517    #[allow(clippy::too_many_arguments)]
518    pub fn forward_embeds(
519        &self,
520        input_ids: &Tensor,
521        input_embeds: Tensor,
522        seqlen_offsets: &[usize],
523        context_lens: Vec<(usize, usize)>,
524        metadata: Option<(Vec<(Tensor, Tensor)>, &PagedAttentionInputMetadata)>,
525        flash_params: &FlashParams,
526    ) -> Result<Tensor> {
527        let mut xs = input_embeds;
528        let cache = &mut self.cache.normal().0;
529        let attention_mask = CausalMasker.make_sliding_window_causal_mask_matrix(
530            input_ids,
531            metadata
532                .as_ref()
533                .map(|(_, _)| &seqlen_offsets as &dyn PastKvLenCache)
534                .unwrap_or(cache as &dyn PastKvLenCache),
535            self.sliding_window,
536            xs.dtype(),
537            self.cfg.num_attn_heads,
538        )?;
539        // PagedAttention prompt chunking
540        let attention_mask = attention_mask.filter(|_| {
541            metadata
542                .as_ref()
543                .map(|(_, meta)| meta.is_first_prompt_chunk)
544                .unwrap_or(true)
545        });
546        for (i, layer) in self.layers.iter().enumerate() {
547            xs = self.mapper.map(xs, i)?;
548            xs = layer.forward(
549                &xs,
550                attention_mask
551                    .as_ref()
552                    .map(|m| m.to_device(xs.device()).unwrap())
553                    .as_ref(),
554                seqlen_offsets,
555                &mut cache[i],
556                metadata
557                    .as_ref()
558                    .map(|(kv_cache, metadata)| (kv_cache[i].clone(), *metadata)),
559                flash_params,
560            )?;
561        }
562        let xs = xs.to_device(&self.device)?;
563        let mut xs = xs.apply(&self.norm)?;
564        if let Some(t) = self.lm_head.quantized_act_type() {
565            xs = xs.to_dtype(t)?;
566        }
567        extract_logits(&MatMul.qmethod_matmul(&xs, &*self.lm_head)?, context_lens)
568    }
569}
570
571impl IsqModel for Model {
572    fn get_layers(
573        &mut self,
574    ) -> (
575        Vec<(&mut Arc<dyn QuantMethod>, Option<usize>)>,
576        &dyn DeviceMapper,
577    ) {
578        let mut tensors = Vec::new();
579        tensors.push((&mut self.lm_head, None));
580        for (i, layer) in self.layers.iter_mut().enumerate() {
581            tensors.push((&mut layer.self_attn.q_proj, Some(i)));
582            tensors.push((&mut layer.self_attn.k_proj, Some(i)));
583            tensors.push((&mut layer.self_attn.v_proj, Some(i)));
584            tensors.push((&mut layer.self_attn.o_proj, Some(i)));
585            tensors.extend(
586                layer
587                    .mlp
588                    .get_isq_layers()
589                    .into_iter()
590                    .map(|m| (m, Some(i)))
591                    .collect::<Vec<_>>(),
592            );
593        }
594        (tensors, &*self.mapper)
595    }
596
597    fn residual_tensors(&self) -> Vec<(String, Tensor)> {
598        let uvb = UnVarBuilder::new();
599
600        let uvb_m = uvb.pp("model");
601        uvb_m.pp("embed_tokens").add(&self.embed_tokens);
602        uvb_m.pp("norm").add(&self.norm);
603
604        for (layer_idx, layer) in self.layers.iter().enumerate() {
605            let uvb_l = uvb_m.pp("layers").pp(layer_idx);
606            uvb_l.pp("input_layernorm").add(&layer.input_layernorm);
607            uvb_l
608                .pp("post_attention_layernorm")
609                .add(&layer.post_attention_layernorm);
610        }
611
612        uvb.to_safetensors()
613    }
614
615    fn imatrix_names(&self) -> candle_core::Result<Vec<Option<String>>> {
616        // NOTE: dependant on the exact implementation in get_layers!
617        let mut names = Vec::new();
618        // lm_head
619        names.push(None);
620        for i in 0..self.layers.len() {
621            names.push(Some(format!("blk.{i}.attn_q.weight")));
622            names.push(Some(format!("blk.{i}.attn_k.weight")));
623            names.push(Some(format!("blk.{i}.attn_v.weight")));
624            names.push(Some(format!("blk.{i}.attn_output.weight")));
625            names.push(Some(format!("blk.{i}.ffn_gate.weight")));
626            names.push(Some(format!("blk.{i}.ffn_up.weight")));
627            names.push(Some(format!("blk.{i}.ffn_down.weight")));
628        }
629        Ok(names)
630    }
631}
632
633impl NormalModel for Model {
634    fn forward(
635        &self,
636        input_ids: &Tensor,
637        seqlen_offsets: &[usize],
638        context_lens: Vec<(usize, usize)>,
639        _position_ids: Vec<usize>,
640        metadata: Option<(Vec<(Tensor, Tensor)>, &PagedAttentionInputMetadata)>,
641        flash_params: &FlashParams,
642    ) -> Result<Tensor> {
643        self.forward(
644            input_ids,
645            seqlen_offsets,
646            context_lens,
647            metadata,
648            flash_params,
649        )
650    }
651    fn xlora_forward(
652        &self,
653        _input_ids: &Tensor,
654        _input_ids_full: &Tensor,
655        _seqlen_offsets: &[usize],
656        _seqlen_offsets_full: &[usize],
657        _no_kv_cache: bool,
658        _non_granular_state: &Option<crate::xlora_models::NonGranularState>,
659        _context_lens: Vec<(usize, usize)>,
660        _position_ids: Vec<usize>,
661        _flash_params: &FlashParams,
662        _flash_params_full: &FlashParams,
663    ) -> Result<Tensor> {
664        unimplemented!()
665    }
666    fn cache(&self) -> &EitherCache {
667        &self.cache
668    }
669    fn cache_mut(&mut self) -> &mut EitherCache {
670        &mut self.cache
671    }
672    fn device(&self) -> &Device {
673        &self.device
674    }
675    fn is_xlora(&self) -> bool {
676        false
677    }
678    fn max_seq_len(&self) -> usize {
679        self.max_seq_len
680    }
681    fn config(&self) -> &ModelConfigMetadata {
682        &self.cfg
683    }
684}
685
686impl AnyMoeBaseModelMixin for Model {
687    fn get_mlps(&self) -> Vec<&dyn MlpLayer> {
688        let mut mlps = Vec::new();
689        for layer in &self.layers {
690            mlps.push(&*layer.mlp);
691        }
692        mlps
693    }
694    fn get_mlps_mut(&mut self) -> Vec<&mut Box<dyn MlpLayer>> {
695        let mut mlps = Vec::new();
696        for layer in &mut self.layers {
697            mlps.push(&mut layer.mlp);
698        }
699        mlps
700    }
701    fn create_anymoe_layers(
702        &mut self,
703        additional_vbs: Vec<ShardedVarBuilder>,
704        config: AnyMoeConfig,
705        (prefix, mlp): (String, String),
706        mut layers: Vec<usize>,
707        expert_type: AnyMoeExpertType,
708        gate_vb: Option<ShardedVarBuilder>,
709    ) -> Result<()> {
710        let mut experts: Vec<Vec<Box<dyn MlpLayer>>> = Vec::new();
711        if layers.is_empty() {
712            layers = (0..self.layers.len()).collect::<Vec<_>>();
713        }
714        for _ in 0..layers.len() {
715            experts.push(Vec::new());
716        }
717        for vb in additional_vbs {
718            let vb = vb.pp(&prefix);
719            for (layer, row) in experts.iter_mut().enumerate() {
720                if !layers.contains(&layer) {
721                    continue;
722                }
723
724                let intermediate_size = self.layers[layer].mlp.get_params()[1];
725                let hidden_size = self.layers[layer].mlp.get_params()[0];
726                match expert_type {
727                    AnyMoeExpertType::FineTuned => {
728                        let (dtype, device) = self.layers[layer].mlp.dtype_device();
729                        row.push(Box::new(Mlp::replicate(
730                            self.layers[layer].mlp.get_params(),
731                            vb.pp(layer).pp(&mlp).set_dtype(dtype).set_device(device),
732                            self.layers[layer].mlp.hidden_act(),
733                            &self.mapper.get_comm_for(layer)?,
734                        )?));
735                    }
736                    AnyMoeExpertType::LoraAdapter {
737                        rank,
738                        alpha,
739                        ref target_modules,
740                    } => {
741                        let vb_mlp = vb.pp(layer).pp(&mlp);
742
743                        let gate_proj_delta = if target_modules.contains(&"gate_proj".to_string()) {
744                            Some(get_delta_from_lora_ab!(
745                                vb_mlp,
746                                rank,
747                                alpha,
748                                (hidden_size, intermediate_size),
749                                "gate_proj"
750                            ))
751                        } else {
752                            None
753                        };
754                        let up_proj_delta = if target_modules.contains(&"up_proj".to_string()) {
755                            Some(get_delta_from_lora_ab!(
756                                vb_mlp,
757                                rank,
758                                alpha,
759                                (hidden_size, intermediate_size),
760                                "up_proj"
761                            ))
762                        } else {
763                            None
764                        };
765                        let down_proj_delta = if target_modules.contains(&"down_proj".to_string()) {
766                            Some(get_delta_from_lora_ab!(
767                                vb_mlp,
768                                rank,
769                                alpha,
770                                (intermediate_size, hidden_size),
771                                "down_proj"
772                            ))
773                        } else {
774                            None
775                        };
776
777                        row.push(self.layers[layer].mlp.new_added_delta(vec![
778                            gate_proj_delta,
779                            up_proj_delta,
780                            down_proj_delta,
781                        ])?);
782                    }
783                }
784            }
785        }
786        for (layer, expert) in layers.into_iter().zip(experts) {
787            let mut experts_all = vec![self.layers[layer].mlp.clone()];
788            experts_all.extend(expert);
789            let (dtype, device) = self.layers[layer].mlp.dtype_device();
790            self.layers[layer].mlp = Box::new(MoeMlp::new(
791                experts_all,
792                config.clone(),
793                dtype,
794                &device,
795                layer,
796                gate_vb.as_ref(),
797            )?);
798        }
799        Ok(())
800    }
801    fn amoe_supported(&self) -> bool {
802        true
803    }
804}
mistralrs_core/models/mistral.rs

mistralrs_core/models/
mistral.rs
