mistralrs_core/vision_models/llava/llava_llm/
mistral.rs

1#![allow(clippy::cast_possible_truncation, clippy::cast_precision_loss)]
2
3use std::sync::Arc;
4
5/// Mistral LLM, https://github.com/mistralai/mistral-src
6use candle_core::{DType, Device, Module, Result, Tensor};
7use mistralrs_quant::{
8    ColumnParallelLayer, QuantMethod, QuantMethodConfig, RowParallelLayer, ShardedVarBuilder,
9    UnquantLinear,
10};
11
12use crate::{
13    amoe::{AnyMoeBaseModelMixin, AnyMoeTrainableLayer, MlpLayer, MoeMlp},
14    attention::SdpaParams,
15    device_map::DeviceMapper,
16    get_delta_from_lora_ab,
17    layers::{self, linear_no_bias, Activation, CausalMasker, MatMul, RmsNorm, Sdpa},
18    layers_masker::PastKvLenCache,
19    paged_attention::{AttentionImplementation, ModelConfigMetadata, PagedAttention},
20    pipeline::{
21        extract_logits,
22        text_models_inputs_processor::{FlashParams, PagedAttentionInputMetadata},
23        Cache, EitherCache, IsqModel, NormalLoadingMetadata, NormalModel,
24    },
25    utils::progress::NiceProgressBar,
26    AnyMoeConfig, AnyMoeExpertType,
27};
28
29use super::{LLaVALLM, OrdinaryRoPE};
30use crate::models::mistral::Config;
31
32#[derive(Clone)]
33#[allow(clippy::upper_case_acronyms)]
34struct MLP {
35    gate_proj: Arc<dyn QuantMethod>,
36    up_proj: Arc<dyn QuantMethod>,
37    down_proj: Arc<dyn QuantMethod>,
38    act_fn: Activation,
39    params: Vec<usize>,
40}
41
42impl MLP {
43    fn new(cfg: &Config, vb: ShardedVarBuilder, comm: &Arc<mistralrs_quant::Comm>) -> Result<Self> {
44        let hidden_sz = cfg.hidden_size;
45        let intermediate_sz = cfg.intermediate_size;
46        let gate_proj = ColumnParallelLayer::new(
47            hidden_sz,
48            intermediate_sz,
49            &cfg.quantization_config,
50            false,
51            comm,
52            vb.pp("gate_proj"),
53        )?;
54        let up_proj = ColumnParallelLayer::new(
55            hidden_sz,
56            intermediate_sz,
57            &cfg.quantization_config,
58            false,
59            comm,
60            vb.pp("up_proj"),
61        )?;
62        let down_proj = RowParallelLayer::new(
63            intermediate_sz,
64            hidden_sz,
65            &cfg.quantization_config,
66            false,
67            comm,
68            vb.pp("down_proj"),
69        )?;
70        Ok(Self {
71            gate_proj,
72            up_proj,
73            down_proj,
74            act_fn: cfg.hidden_act,
75            params: vec![hidden_sz, intermediate_sz],
76        })
77    }
78}
79
80impl AnyMoeTrainableLayer for MLP {}
81
82impl MlpLayer for MLP {
83    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
84        let original_dtype = xs.dtype();
85        let mut xs = xs.clone();
86        if let Some(t) = self.gate_proj.quantized_act_type() {
87            xs = xs.to_dtype(t)?;
88        }
89        let lhs = MatMul
90            .qmethod_matmul(&xs, &*self.gate_proj)?
91            .apply(&self.act_fn)?;
92        let rhs = MatMul.qmethod_matmul(&xs, &*self.up_proj)?;
93        let mut res = MatMul.qmethod_matmul(&(lhs * rhs)?, &*self.down_proj)?;
94        if self.gate_proj.quantized_act_type().is_some() {
95            res = res.to_dtype(original_dtype)?;
96        }
97        Ok(res)
98    }
99    fn get_isq_layers(&mut self) -> Vec<&mut Arc<dyn QuantMethod>> {
100        vec![&mut self.gate_proj, &mut self.up_proj, &mut self.down_proj]
101    }
102    fn clone(&self) -> Box<dyn MlpLayer> {
103        Box::new(Clone::clone(self))
104    }
105    fn get_params(&self) -> &[usize] {
106        &self.params
107    }
108    fn hidden_act(&self) -> Activation {
109        self.act_fn
110    }
111    // gate, up, down
112    fn new_added_delta(&self, deltas: Vec<Option<Tensor>>) -> Result<Box<dyn MlpLayer>> {
113        let gate_proj = if let Some(ref delta) = deltas[0] {
114            self.gate_proj.add_delta_w(delta)?
115        } else {
116            self.gate_proj.clone()
117        };
118        let up_proj = if let Some(ref delta) = deltas[1] {
119            self.up_proj.add_delta_w(delta)?
120        } else {
121            self.up_proj.clone()
122        };
123        let down_proj = if let Some(ref delta) = deltas[2] {
124            self.down_proj.add_delta_w(delta)?
125        } else {
126            self.down_proj.clone()
127        };
128
129        Ok(Box::new(Self {
130            gate_proj,
131            up_proj,
132            down_proj,
133            act_fn: self.act_fn,
134            params: self.params.clone(),
135        }))
136    }
137
138    fn dtype_device(&self) -> (DType, Device) {
139        self.gate_proj.dtype_and_device()
140    }
141}
142
143struct Attention {
144    q_proj: Arc<dyn QuantMethod>,
145    k_proj: Arc<dyn QuantMethod>,
146    v_proj: Arc<dyn QuantMethod>,
147    o_proj: Arc<dyn QuantMethod>,
148    num_heads: usize,
149    num_kv_heads: usize,
150    head_dim: usize,
151    sliding_window: Option<usize>,
152    paged_attn: Option<PagedAttention>,
153    sdpa_params: SdpaParams,
154}
155
156impl Attention {
157    fn new(
158        cfg: &Config,
159        vb: ShardedVarBuilder,
160        paged_attn: Option<PagedAttention>,
161        comm: &Arc<mistralrs_quant::Comm>,
162    ) -> Result<Self> {
163        let hidden_sz = cfg.hidden_size;
164        let num_heads = cfg.num_attention_heads;
165        let num_kv_heads = cfg.num_key_value_heads;
166        let head_dim = cfg.head_dim();
167        let q_proj = ColumnParallelLayer::new(
168            hidden_sz,
169            num_heads * head_dim,
170            &cfg.quantization_config,
171            false,
172            comm,
173            vb.pp("q_proj"),
174        )?;
175        let kv_shard = mistralrs_quant::compute_kv_shard(
176            cfg.num_key_value_heads,
177            cfg.hidden_size / cfg.num_attention_heads,
178            comm,
179        );
180        let k_proj = ColumnParallelLayer::new_with_shard(
181            hidden_sz,
182            num_kv_heads * head_dim,
183            &cfg.quantization_config,
184            false,
185            comm,
186            kv_shard,
187            vb.pp("k_proj"),
188        )?;
189        let v_proj = ColumnParallelLayer::new_with_shard(
190            hidden_sz,
191            num_kv_heads * head_dim,
192            &cfg.quantization_config,
193            false,
194            comm,
195            kv_shard,
196            vb.pp("v_proj"),
197        )?;
198        let o_proj = RowParallelLayer::new(
199            num_heads * head_dim,
200            hidden_sz,
201            &cfg.quantization_config,
202            false,
203            comm,
204            vb.pp("o_proj"),
205        )?;
206        Ok(Self {
207            q_proj,
208            k_proj,
209            v_proj,
210            o_proj,
211            num_heads: num_heads / comm.world_size(),
212            num_kv_heads: (num_kv_heads / comm.world_size()).max(1),
213            head_dim,
214            sliding_window: cfg.sliding_window,
215            paged_attn,
216            sdpa_params: SdpaParams {
217                n_kv_groups: mistralrs_quant::compute_n_kv_groups(
218                    cfg.num_key_value_heads,
219                    cfg.num_attention_heads,
220                    comm,
221                ),
222                softcap: None,
223                softmax_scale: 1.0 / (head_dim as f32).sqrt(),
224                sliding_window: cfg.sliding_window,
225            },
226        })
227    }
228
229    #[allow(clippy::too_many_arguments)]
230    fn forward(
231        &self,
232        xs: &Tensor,
233        attention_mask: Option<&Tensor>,
234        seqlen_offsets: &[usize],
235        kv_cache: &mut Option<(Tensor, Tensor)>,
236        rope_parameter: (&Tensor, &Tensor),
237        metadata: Option<((Tensor, Tensor), &PagedAttentionInputMetadata)>,
238        flash_params: &FlashParams,
239    ) -> Result<Tensor> {
240        let (b_sz, q_len, _) = xs.dims3()?;
241
242        let original_dtype = xs.dtype();
243        let mut xs = xs.clone();
244        if let Some(t) = self.q_proj.quantized_act_type() {
245            xs = xs.to_dtype(t)?;
246        }
247        let mut q = MatMul.qmethod_matmul(&xs, &*self.q_proj)?;
248        let mut k = MatMul.qmethod_matmul(&xs, &*self.k_proj)?;
249        let mut v = MatMul.qmethod_matmul(&xs, &*self.v_proj)?;
250        if self.q_proj.quantized_act_type().is_some() {
251            q = q.to_dtype(original_dtype)?;
252            k = k.to_dtype(original_dtype)?;
253            v = v.to_dtype(original_dtype)?;
254        }
255
256        let mut q = q
257            .reshape((b_sz, q_len, self.num_heads, self.head_dim))?
258            .transpose(1, 2)?
259            .contiguous()?;
260        let mut k = k
261            .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
262            .transpose(1, 2)?
263            .contiguous()?;
264        q = OrdinaryRoPE::forward(&q, seqlen_offsets[0], rope_parameter.0, rope_parameter.1)?;
265        k = OrdinaryRoPE::forward(&k, seqlen_offsets[0], rope_parameter.0, rope_parameter.1)?;
266        let v = v
267            .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
268            .transpose(1, 2)?;
269
270        let mut attn_output = match &self.paged_attn {
271            Some(paged_attn) => match metadata {
272                Some(((key_cache, value_cache), input_metadata)) => paged_attn.forward(
273                    &q,
274                    &k,
275                    &v,
276                    attention_mask,
277                    Some(key_cache),
278                    Some(value_cache),
279                    input_metadata,
280                    &self.sdpa_params,
281                    Some(flash_params),
282                )?,
283                None => {
284                    // If we don't have metadata, we are most likely generating an imatrix so we don't want to populate that.
285                    // Generating the dummy metadata with the assumption that we are not generating text (only processing prompts).
286                    let input_metadata = PagedAttentionInputMetadata::dummy(q.device())?;
287                    // Sanity check.
288                    assert!(attention_mask.is_some());
289                    paged_attn.forward(
290                        &q,
291                        &k,
292                        &v,
293                        attention_mask,
294                        None,
295                        None,
296                        &input_metadata,
297                        &self.sdpa_params,
298                        Some(flash_params),
299                    )?
300                }
301            },
302            None => {
303                let (k, v, attn_mask) = Cache::update_kv_cache_sliding_window(
304                    kv_cache,
305                    k,
306                    v,
307                    attention_mask,
308                    self.sliding_window,
309                )?;
310
311                Sdpa.run_attention(
312                    &q,
313                    &k,
314                    &v,
315                    attn_mask.as_ref(),
316                    Some(flash_params),
317                    &self.sdpa_params,
318                )?
319            }
320        };
321
322        if let Some(t) = self.q_proj.quantized_act_type() {
323            attn_output = attn_output.to_dtype(t)?;
324        }
325        attn_output = if attention_mask.is_some() {
326            attn_output.transpose(1, 2)?.reshape((b_sz, q_len, ()))?
327        } else {
328            attn_output.reshape((b_sz, q_len, ()))?
329        };
330        let mut res = MatMul.qmethod_matmul(&attn_output, &*self.o_proj)?;
331        if self.q_proj.quantized_act_type().is_some() {
332            res = res.to_dtype(original_dtype)?;
333        }
334        Ok(res)
335    }
336}
337
338struct DecoderLayer {
339    self_attn: Attention,
340    mlp: Box<dyn MlpLayer>,
341    input_layernorm: RmsNorm,
342    post_attention_layernorm: RmsNorm,
343    rope_parameter: (Tensor, Tensor),
344}
345
346impl DecoderLayer {
347    #[allow(clippy::too_many_arguments)]
348    fn new(
349        cfg: &Config,
350        vb: ShardedVarBuilder,
351        mapper: &dyn DeviceMapper,
352        layer_idx: usize,
353        loading_isq: bool,
354        paged_attn: Option<PagedAttention>,
355        rope_parameter: (Tensor, Tensor),
356        comm: &Arc<mistralrs_quant::Comm>,
357    ) -> Result<Self> {
358        let self_attn = Attention::new(
359            cfg,
360            mapper.set_device(layer_idx, vb.pp("self_attn"), loading_isq),
361            paged_attn,
362            comm,
363        )?;
364        let mlp = MLP::new(
365            cfg,
366            mapper.set_device(layer_idx, vb.pp("mlp"), loading_isq),
367            comm,
368        )?;
369        let input_layernorm = RmsNorm::new(
370            cfg.hidden_size,
371            cfg.rms_norm_eps,
372            mapper.set_device(layer_idx, vb.pp("input_layernorm"), false),
373        )?;
374        let post_attention_layernorm = RmsNorm::new(
375            cfg.hidden_size,
376            cfg.rms_norm_eps,
377            mapper.set_device(layer_idx, vb.pp("post_attention_layernorm"), false),
378        )?;
379        Ok(Self {
380            self_attn,
381            mlp: Box::new(mlp),
382            input_layernorm,
383            post_attention_layernorm,
384            rope_parameter,
385        })
386    }
387
388    #[allow(clippy::too_many_arguments)]
389    fn forward(
390        &self,
391        xs: &Tensor,
392        attention_mask: Option<&Tensor>,
393        seqlen_offsets: &[usize],
394        kv_cache: &mut Option<(Tensor, Tensor)>,
395        metadata: Option<((Tensor, Tensor), &PagedAttentionInputMetadata)>,
396        flash_params: &FlashParams,
397    ) -> Result<Tensor> {
398        let residual = xs;
399        let mut xs = self.input_layernorm.forward(xs)?;
400        xs = self.self_attn.forward(
401            &xs,
402            attention_mask,
403            seqlen_offsets,
404            kv_cache,
405            (&self.rope_parameter.0, &self.rope_parameter.1),
406            metadata,
407            flash_params,
408        )?;
409        xs = (xs + residual)?;
410        let residual = &xs;
411        let xs = self
412            .mlp
413            .forward(&xs.apply(&self.post_attention_layernorm)?)?;
414        residual + xs
415    }
416}
417
418pub struct Model {
419    embed_tokens: candle_nn::Embedding,
420    layers: Vec<DecoderLayer>,
421    norm: RmsNorm,
422    lm_head: Arc<dyn QuantMethod>,
423    sliding_window: Option<usize>,
424    device: Device,
425    cache: EitherCache,
426    max_seq_len: usize,
427    mapper: Box<dyn DeviceMapper + Send + Sync>,
428    cfg: ModelConfigMetadata,
429}
430
431impl Model {
432    pub fn new(
433        cfg: &Config,
434        vb: ShardedVarBuilder,
435        is_gptx: bool,
436        normal_loading_metadata: NormalLoadingMetadata,
437        attention_mechanism: AttentionImplementation,
438    ) -> Result<Self> {
439        let vb_m = vb.pp("model");
440        let vb_lm_head = vb.pp("lm_head");
441        Self::new_inner(
442            cfg,
443            vb_m,
444            vb_lm_head,
445            is_gptx,
446            normal_loading_metadata,
447            attention_mechanism,
448        )
449    }
450
451    pub fn new_inner(
452        cfg: &Config,
453        vb_m: ShardedVarBuilder,
454        vb_lm_head: ShardedVarBuilder,
455        _is_gptx: bool,
456        normal_loading_metadata: NormalLoadingMetadata,
457        attention_mechanism: AttentionImplementation,
458    ) -> Result<Self> {
459        if let Some(ref quant_cfg) = &cfg.quantization_config {
460            tracing::info!(
461                "Using {} quantization: {}.",
462                quant_cfg.name(),
463                quant_cfg.get_bits_name(&vb_m)
464            );
465        }
466        let mapper = normal_loading_metadata.mapper;
467        let embed_tokens = layers::embedding(
468            cfg.vocab_size,
469            cfg.hidden_size,
470            mapper.set_nm_device(vb_m.pp("embed_tokens"), false),
471            &cfg.quantization_config,
472        )?;
473        let head_dim = cfg.hidden_size / cfg.num_attention_heads;
474        let vb_l = vb_m.pp("layers");
475        let layers = NiceProgressBar::<_, 'b'>(
476            0..cfg.num_hidden_layers,
477            "Loading repeating layers",
478            &normal_loading_metadata.multi_progress,
479        )
480        .par_iter_if_isq(|layer_idx| {
481            let device = mapper
482                .device_for(layer_idx, false)
483                .unwrap_or(&normal_loading_metadata.real_device);
484            let rope_parameters = OrdinaryRoPE::create_parameters(
485                head_dim,
486                cfg.max_position_embeddings,
487                cfg.rope_theta as f32,
488                vb_m.dtype(),
489                device,
490            )?;
491            let paged_attn = match &attention_mechanism {
492                AttentionImplementation::Eager => None,
493                AttentionImplementation::PagedAttention => {
494                    Some(PagedAttention::new(head_dim, device, None)?)
495                }
496            };
497            let comm = mapper.get_comm_for(layer_idx)?;
498            DecoderLayer::new(
499                cfg,
500                vb_l.pp(layer_idx),
501                &*mapper,
502                layer_idx,
503                normal_loading_metadata.loading_isq,
504                paged_attn,
505                rope_parameters,
506                &comm,
507            )
508        })?;
509        let norm = RmsNorm::new(
510            cfg.hidden_size,
511            cfg.rms_norm_eps,
512            mapper.set_nm_device(vb_m.pp("norm"), false),
513        )?;
514        let lm_head = linear_no_bias(
515            cfg.hidden_size,
516            cfg.vocab_size,
517            mapper.set_nm_device(vb_lm_head, normal_loading_metadata.loading_isq),
518        )?;
519        Ok(Self {
520            embed_tokens,
521            layers,
522            norm,
523            lm_head: Arc::new(UnquantLinear::new(QuantMethodConfig::Unquantized(lm_head))?),
524            sliding_window: cfg.sliding_window,
525            device: normal_loading_metadata.real_device,
526            cache: EitherCache::Full(Cache::new(cfg.num_hidden_layers, false)),
527            max_seq_len: cfg.max_position_embeddings,
528            cfg: ModelConfigMetadata {
529                max_seq_len: cfg.max_position_embeddings,
530                num_layers: cfg.num_hidden_layers,
531                hidden_size: cfg.hidden_size,
532                num_attn_heads: cfg.num_attention_heads / mapper.get_comm_for(0)?.world_size(),
533                num_kv_heads: (cfg.num_key_value_heads / mapper.get_comm_for(0)?.world_size())
534                    .max(1),
535                sliding_window: cfg.sliding_window,
536                k_head_dim: cfg.head_dim(),
537                v_head_dim: cfg.head_dim(),
538            },
539            mapper,
540        })
541    }
542
543    pub fn get_input_embeddings(&self, input_ids: &Tensor) -> Result<Tensor> {
544        self.embed_tokens.forward(input_ids)
545    }
546
547    pub fn forward(
548        &self,
549        input_ids: &Tensor,
550        seqlen_offsets: &[usize],
551        context_lens: Vec<(usize, usize)>,
552        metadata: Option<(Vec<(Tensor, Tensor)>, &PagedAttentionInputMetadata)>,
553        flash_params: &FlashParams,
554    ) -> Result<Tensor> {
555        self.forward_embeds(
556            input_ids,
557            self.embed_tokens.forward(input_ids)?,
558            seqlen_offsets,
559            context_lens,
560            metadata,
561            flash_params,
562        )
563    }
564
565    #[allow(clippy::too_many_arguments)]
566    pub fn forward_embeds(
567        &self,
568        input_ids: &Tensor,
569        input_embeds: Tensor,
570        seqlen_offsets: &[usize],
571        context_lens: Vec<(usize, usize)>,
572        metadata: Option<(Vec<(Tensor, Tensor)>, &PagedAttentionInputMetadata)>,
573        flash_params: &FlashParams,
574    ) -> Result<Tensor> {
575        let mut xs = input_embeds;
576        let mut cache = self.cache.full().lock();
577        let attention_mask = CausalMasker.make_sliding_window_causal_mask_matrix(
578            input_ids,
579            metadata
580                .as_ref()
581                .map(|(_, _)| &seqlen_offsets as &dyn PastKvLenCache)
582                .unwrap_or(&*cache as &dyn PastKvLenCache),
583            self.sliding_window,
584            xs.dtype(),
585            self.cfg.num_attn_heads,
586        )?;
587        let attention_mask = attention_mask.filter(|_| {
588            metadata
589                .as_ref()
590                .map(|(_, meta)| meta.is_first_prompt_chunk)
591                .unwrap_or(true)
592        });
593        for (i, layer) in self.layers.iter().enumerate() {
594            xs = self.mapper.map(xs, i)?;
595            xs = layer.forward(
596                &xs,
597                attention_mask
598                    .as_ref()
599                    .map(|m| m.to_device(xs.device()).unwrap())
600                    .as_ref(),
601                seqlen_offsets,
602                &mut cache[i],
603                metadata
604                    .as_ref()
605                    .map(|(kv_cache, metadata)| (kv_cache[i].clone(), *metadata)),
606                flash_params,
607            )?;
608        }
609        xs = xs.to_device(&self.device)?;
610        xs = xs.apply(&self.norm)?;
611        if let Some(t) = self.lm_head.quantized_act_type() {
612            xs = xs.to_dtype(t)?;
613        }
614        extract_logits(&MatMul.qmethod_matmul(&xs, &*self.lm_head)?, context_lens)
615    }
616}
617
618impl IsqModel for Model {
619    fn get_layers(
620        &mut self,
621    ) -> (
622        Vec<(&mut Arc<dyn QuantMethod>, Option<usize>)>,
623        &dyn DeviceMapper,
624    ) {
625        let mut tensors = Vec::new();
626        tensors.push((&mut self.lm_head, None));
627        for (i, layer) in self.layers.iter_mut().enumerate() {
628            tensors.push((&mut layer.self_attn.q_proj, Some(i)));
629            tensors.push((&mut layer.self_attn.k_proj, Some(i)));
630            tensors.push((&mut layer.self_attn.v_proj, Some(i)));
631            tensors.push((&mut layer.self_attn.o_proj, Some(i)));
632            tensors.extend(
633                layer
634                    .mlp
635                    .get_isq_layers()
636                    .into_iter()
637                    .map(|m| (m, Some(i)))
638                    .collect::<Vec<_>>(),
639            );
640        }
641        (tensors, &*self.mapper)
642    }
643
644    fn residual_tensors(&self) -> Vec<(String, Tensor)> {
645        Vec::new()
646    }
647}
648
649impl LLaVALLM for Model {
650    fn embed(&self, input_ids: &Tensor) -> Result<Tensor> {
651        self.get_input_embeddings(input_ids)
652    }
653
654    fn forward_input_embed(
655        &self,
656        input_ids: &Tensor,
657        input_embed: Tensor,
658        seqlen_offsets: &[usize],
659        context_lens: Vec<(usize, usize)>,
660        metadata: Option<(Vec<(Tensor, Tensor)>, &PagedAttentionInputMetadata)>,
661        flash_params: &FlashParams,
662    ) -> Result<Tensor> {
663        self.forward_embeds(
664            input_ids,
665            input_embed,
666            seqlen_offsets,
667            context_lens,
668            metadata,
669            flash_params,
670        )
671    }
672}
673
674impl NormalModel for Model {
675    fn forward(
676        &self,
677        input_ids: &Tensor,
678        seqlen_offsets: &[usize],
679        context_lens: Vec<(usize, usize)>,
680        _position_ids: Vec<usize>,
681        metadata: Option<(Vec<(Tensor, Tensor)>, &PagedAttentionInputMetadata)>,
682        flash_params: &FlashParams,
683    ) -> Result<Tensor> {
684        self.forward(
685            input_ids,
686            seqlen_offsets,
687            context_lens,
688            metadata,
689            flash_params,
690        )
691    }
692    fn xlora_forward(
693        &self,
694        _input_ids: &Tensor,
695        _input_ids_full: &Tensor,
696        _seqlen_offsets: &[usize],
697        _seqlen_offsets_full: &[usize],
698        _no_kv_cache: bool,
699        _non_granular_state: &Option<crate::xlora_models::NonGranularState>,
700        _context_lens: Vec<(usize, usize)>,
701        _position_ids: Vec<usize>,
702        _flash_params: &FlashParams,
703        _flash_params_full: &FlashParams,
704    ) -> Result<Tensor> {
705        unimplemented!()
706    }
707    fn cache(&self) -> &EitherCache {
708        &self.cache
709    }
710    fn cache_mut(&mut self) -> &mut EitherCache {
711        &mut self.cache
712    }
713    fn device(&self) -> &Device {
714        &self.device
715    }
716    fn is_xlora(&self) -> bool {
717        false
718    }
719    fn max_seq_len(&self) -> usize {
720        self.max_seq_len
721    }
722    fn config(&self) -> &ModelConfigMetadata {
723        &self.cfg
724    }
725}
726
727impl AnyMoeBaseModelMixin for Model {
728    fn get_mlps(&self) -> Vec<&dyn MlpLayer> {
729        let mut mlps = Vec::new();
730        for layer in &self.layers {
731            mlps.push(&*layer.mlp);
732        }
733        mlps
734    }
735    fn get_mlps_mut(&mut self) -> Vec<&mut Box<dyn MlpLayer>> {
736        let mut mlps = Vec::new();
737        for layer in &mut self.layers {
738            mlps.push(&mut layer.mlp);
739        }
740        mlps
741    }
742    fn create_anymoe_layers(
743        &mut self,
744        additional_vbs: Vec<ShardedVarBuilder>,
745        config: AnyMoeConfig,
746        (prefix, mlp): (String, String),
747        mut layers: Vec<usize>,
748        expert_type: AnyMoeExpertType,
749        gate_vb: Option<ShardedVarBuilder>,
750    ) -> Result<()> {
751        let mut experts: Vec<Vec<Box<dyn MlpLayer>>> = Vec::new();
752        if layers.is_empty() {
753            layers = (0..self.layers.len()).collect::<Vec<_>>();
754        }
755        for _ in 0..layers.len() {
756            experts.push(Vec::new());
757        }
758        for vb in additional_vbs {
759            let vb = vb.pp(&prefix);
760            for (layer, row) in experts.iter_mut().enumerate() {
761                if !layers.contains(&layer) {
762                    continue;
763                }
764
765                let intermediate_size = self.layers[layer].mlp.get_params()[1];
766                let hidden_size = self.layers[layer].mlp.get_params()[0];
767                match expert_type {
768                    AnyMoeExpertType::FineTuned => {
769                        let (dtype, device) = self.layers[layer].mlp.dtype_device();
770                        row.push(Box::new(MLP::new(
771                            &Config {
772                                intermediate_size: self.layers[layer].mlp.get_params()[1],
773                                hidden_size: self.layers[layer].mlp.get_params()[0],
774                                ..Default::default()
775                            },
776                            vb.pp(layer).pp(&mlp).set_dtype(dtype).set_device(device),
777                            &self.mapper.get_comm_for(layer)?,
778                        )?));
779                    }
780                    AnyMoeExpertType::LoraAdapter {
781                        rank,
782                        alpha,
783                        ref target_modules,
784                    } => {
785                        let vb_mlp = vb.pp(layer).pp(&mlp);
786
787                        let gate_proj_delta = if target_modules.contains(&"gate_proj".to_string()) {
788                            Some(get_delta_from_lora_ab!(
789                                vb_mlp,
790                                rank,
791                                alpha,
792                                (hidden_size, intermediate_size),
793                                "gate_proj"
794                            ))
795                        } else {
796                            None
797                        };
798                        let up_proj_delta = if target_modules.contains(&"up_proj".to_string()) {
799                            Some(get_delta_from_lora_ab!(
800                                vb_mlp,
801                                rank,
802                                alpha,
803                                (hidden_size, intermediate_size),
804                                "up_proj"
805                            ))
806                        } else {
807                            None
808                        };
809                        let down_proj_delta = if target_modules.contains(&"down_proj".to_string()) {
810                            Some(get_delta_from_lora_ab!(
811                                vb_mlp,
812                                rank,
813                                alpha,
814                                (intermediate_size, hidden_size),
815                                "down_proj"
816                            ))
817                        } else {
818                            None
819                        };
820
821                        row.push(self.layers[layer].mlp.new_added_delta(vec![
822                            gate_proj_delta,
823                            up_proj_delta,
824                            down_proj_delta,
825                        ])?);
826                    }
827                }
828            }
829        }
830        for (layer, expert) in layers.into_iter().zip(experts) {
831            let mut experts_all = vec![self.layers[layer].mlp.clone()];
832            experts_all.extend(expert);
833            let (dtype, device) = self.layers[layer].mlp.dtype_device();
834            self.layers[layer].mlp = Box::new(MoeMlp::new(
835                experts_all,
836                config.clone(),
837                dtype,
838                &device,
839                layer,
840                gate_vb.as_ref(),
841            )?);
842        }
843        Ok(())
844    }
845    fn amoe_supported(&self) -> bool {
846        true
847    }
848}
mistralrs_core/vision_models/llava/llava_llm/mistral.rs

mistralrs_core/vision_models/llava/llava_llm/
mistral.rs
