mistralrs_core/xlora_models/
gemma.rs

1#![allow(clippy::cast_possible_truncation, clippy::cast_precision_loss)]
2
3use std::{collections::HashMap, sync::Arc};
4
5use crate::{
6    amoe::AnyMoeBaseModelMixin,
7    attention::SdpaParams,
8    layers::{self, Activation, RmsNorm, RotaryEmbedding, Sdpa},
9    lora::{linear_b as linear, LinearLayerLike, LoraConfig, Ordering},
10    paged_attention::ModelConfigMetadata,
11    pipeline::{
12        text_models_inputs_processor::{FlashParams, PagedAttentionInputMetadata},
13        EitherCache, IsqModel, NormalLoadingMetadata,
14    },
15    utils::progress::NiceProgressBar,
16};
17use candle_core::{DType, Device, Module, Result, Tensor};
18use mistralrs_quant::{QuantMethod, ShardedVarBuilder};
19use tqdm::Iter;
20use tracing::info;
21
22use crate::{
23    device_map::DeviceMapper,
24    layers::CausalMasker,
25    models::gemma::Config,
26    pipeline::{extract_logits, Cache, NormalModel},
27};
28
29use super::{classifier::XLoraClassifier, NonGranularState, ScalingsMaker, XLoraConfig};
30
31fn default_max_position_embeddings() -> usize {
32    4096
33}
34
35#[derive(Clone)]
36#[allow(clippy::upper_case_acronyms)]
37struct MLP {
38    gate_proj: Arc<dyn LinearLayerLike + Send + Sync>,
39    up_proj: Arc<dyn LinearLayerLike + Send + Sync>,
40    down_proj: Arc<dyn LinearLayerLike + Send + Sync>,
41    act_fn: Activation,
42}
43
44impl MLP {
45    #[allow(clippy::too_many_arguments)]
46    fn new(
47        cfg: &Config,
48        vb: ShardedVarBuilder,
49        lora_config: &[((String, String), LoraConfig)],
50        count: &mut usize,
51        ord: &Ordering,
52        mapper: &dyn DeviceMapper,
53        layer_idx: usize,
54        loading_isq: bool,
55        preload_adapters: &Option<HashMap<String, (ShardedVarBuilder, LoraConfig)>>,
56    ) -> Result<Self> {
57        let hidden_sz = cfg.hidden_size;
58        let intermediate_sz = cfg.intermediate_size;
59        let gate_proj = linear(
60            hidden_sz,
61            intermediate_sz,
62            false,
63            mapper.set_device(layer_idx, vb.pp("gate_proj"), loading_isq),
64            mapper.set_device(layer_idx, vb.pp("gate_proj"), false),
65            lora_config,
66            count,
67            ord,
68            preload_adapters,
69        )?;
70        let up_proj = linear(
71            hidden_sz,
72            intermediate_sz,
73            false,
74            mapper.set_device(layer_idx, vb.pp("up_proj"), loading_isq),
75            mapper.set_device(layer_idx, vb.pp("up_proj"), false),
76            lora_config,
77            count,
78            ord,
79            preload_adapters,
80        )?;
81        let down_proj = linear(
82            intermediate_sz,
83            hidden_sz,
84            false,
85            mapper.set_device(layer_idx, vb.pp("down_proj"), loading_isq),
86            mapper.set_device(layer_idx, vb.pp("down_proj"), false),
87            lora_config,
88            count,
89            ord,
90            preload_adapters,
91        )?;
92        Ok(Self {
93            gate_proj,
94            up_proj,
95            down_proj,
96            act_fn: cfg.hidden_act()?,
97        })
98    }
99
100    fn forward(
101        &self,
102        xs: &Tensor,
103        scalings: Option<Tensor>,
104        global_scaling_weight: f64,
105        is_scaling_pass: Option<f64>,
106    ) -> Result<Tensor> {
107        let original_dtype = xs.dtype();
108        let mut xs = xs.clone();
109        if let Some(t) = self.gate_proj.quantized_act_type() {
110            xs = xs.to_dtype(t)?;
111        }
112        let lhs = self
113            .gate_proj
114            .lora_forward(
115                &xs,
116                scalings.clone(),
117                global_scaling_weight,
118                is_scaling_pass,
119            )?
120            .apply(&self.act_fn)?;
121        let rhs = self.up_proj.lora_forward(
122            &xs,
123            scalings.clone(),
124            global_scaling_weight,
125            is_scaling_pass,
126        )?;
127        let mut res = self.down_proj.lora_forward(
128            &(lhs * rhs)?,
129            scalings,
130            global_scaling_weight,
131            is_scaling_pass,
132        )?;
133        if self.gate_proj.quantized_act_type().is_some() {
134            res = res.to_dtype(original_dtype)?;
135        }
136        Ok(res)
137    }
138}
139
140struct Attention {
141    q_proj: Arc<dyn LinearLayerLike + Send + Sync>,
142    k_proj: Arc<dyn LinearLayerLike + Send + Sync>,
143    v_proj: Arc<dyn LinearLayerLike + Send + Sync>,
144    o_proj: Arc<dyn LinearLayerLike + Send + Sync>,
145    num_heads: usize,
146    num_kv_heads: usize,
147    head_dim: usize,
148    rotary_emb: Arc<RotaryEmbedding>,
149    sdpa_params: SdpaParams,
150}
151
152impl Attention {
153    #[allow(clippy::too_many_arguments)]
154    fn new(
155        rotary_emb: Arc<RotaryEmbedding>,
156        cfg: &Config,
157        vb: ShardedVarBuilder,
158        lora_config: &[((String, String), LoraConfig)],
159        count: &mut usize,
160        ord: &Ordering,
161        mapper: &dyn DeviceMapper,
162        layer_idx: usize,
163        loading_isq: bool,
164        preload_adapters: &Option<HashMap<String, (ShardedVarBuilder, LoraConfig)>>,
165    ) -> Result<Self> {
166        let hidden_sz = cfg.hidden_size;
167        let num_heads = cfg.num_attention_heads;
168        let num_kv_heads = cfg.num_key_value_heads;
169        let head_dim = cfg.head_dim;
170        let bias = cfg.attention_bias;
171        let q_proj = linear(
172            hidden_sz,
173            num_heads * head_dim,
174            bias,
175            mapper.set_device(layer_idx, vb.pp("q_proj"), loading_isq),
176            mapper.set_device(layer_idx, vb.pp("q_proj"), false),
177            lora_config,
178            count,
179            ord,
180            preload_adapters,
181        )?;
182        let k_proj = linear(
183            hidden_sz,
184            num_kv_heads * head_dim,
185            bias,
186            mapper.set_device(layer_idx, vb.pp("k_proj"), loading_isq),
187            mapper.set_device(layer_idx, vb.pp("k_proj"), false),
188            lora_config,
189            count,
190            ord,
191            preload_adapters,
192        )?;
193        let v_proj = linear(
194            hidden_sz,
195            num_kv_heads * head_dim,
196            bias,
197            mapper.set_device(layer_idx, vb.pp("v_proj"), loading_isq),
198            mapper.set_device(layer_idx, vb.pp("v_proj"), false),
199            lora_config,
200            count,
201            ord,
202            preload_adapters,
203        )?;
204        let o_proj = linear(
205            num_heads * head_dim,
206            hidden_sz,
207            bias,
208            mapper.set_device(layer_idx, vb.pp("o_proj"), loading_isq),
209            mapper.set_device(layer_idx, vb.pp("o_proj"), false),
210            lora_config,
211            count,
212            ord,
213            preload_adapters,
214        )?;
215        Ok(Self {
216            q_proj,
217            k_proj,
218            v_proj,
219            o_proj,
220            num_heads,
221            num_kv_heads,
222            head_dim,
223            rotary_emb,
224            sdpa_params: SdpaParams {
225                n_kv_groups: num_heads / num_kv_heads,
226                softcap: None,
227                softmax_scale: 1.0 / (head_dim as f32).sqrt(),
228                sliding_window: None,
229            },
230        })
231    }
232
233    #[allow(clippy::too_many_arguments)]
234    fn forward(
235        &self,
236        xs: &Tensor,
237        attention_mask: Option<&Tensor>,
238        seqlen_offsets: &[usize],
239        kv_cache: &mut Option<(Tensor, Tensor)>,
240        scalings: Option<Tensor>,
241        global_scaling_weight: f64,
242        is_scaling_pass: Option<f64>,
243        flash_params: &FlashParams,
244    ) -> Result<Tensor> {
245        let (b_sz, q_len, _) = xs.dims3()?;
246
247        let original_dtype = xs.dtype();
248        let mut xs = xs.clone();
249        if let Some(t) = self.q_proj.quantized_act_type() {
250            xs = xs.to_dtype(t)?;
251        }
252        let mut q = self.q_proj.lora_forward(
253            &xs,
254            scalings.clone(),
255            global_scaling_weight,
256            is_scaling_pass,
257        )?;
258        let mut k = self.k_proj.lora_forward(
259            &xs,
260            scalings.clone(),
261            global_scaling_weight,
262            is_scaling_pass,
263        )?;
264        let mut v = self.v_proj.lora_forward(
265            &xs,
266            scalings.clone(),
267            global_scaling_weight,
268            is_scaling_pass,
269        )?;
270        if self.q_proj.quantized_act_type().is_some() {
271            q = q.to_dtype(original_dtype)?;
272            k = k.to_dtype(original_dtype)?;
273            v = v.to_dtype(original_dtype)?;
274        }
275
276        let (q, k, v) = if q_len != 1 {
277            let q = q
278                .reshape((b_sz, q_len, self.num_heads, self.head_dim))?
279                .transpose(1, 2)?;
280            let k = k
281                .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
282                .transpose(1, 2)?;
283            let v = v
284                .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
285                .transpose(1, 2)?;
286            (q, k, v)
287        } else {
288            let q = q.reshape((b_sz, self.num_heads, q_len, self.head_dim))?;
289            let k = k.reshape((b_sz, self.num_kv_heads, q_len, self.head_dim))?;
290            let v = v.reshape((b_sz, self.num_kv_heads, q_len, self.head_dim))?;
291            (q, k, v)
292        };
293
294        let (q, k) = self.rotary_emb.forward(&q, &k, seqlen_offsets)?;
295
296        let (k, v) = Cache::update_kv_cache(kv_cache, k, v)?;
297
298        let mut attn_output = Sdpa.run_attention(
299            &q,
300            &k,
301            &v,
302            attention_mask,
303            Some(flash_params),
304            &self.sdpa_params,
305        )?;
306
307        if let Some(t) = self.q_proj.quantized_act_type() {
308            attn_output = attn_output.to_dtype(t)?;
309        }
310        let mut res = self.o_proj.lora_forward(
311            &attn_output.transpose(1, 2)?.reshape((b_sz, q_len, ()))?,
312            scalings.clone(),
313            global_scaling_weight,
314            is_scaling_pass,
315        )?;
316        if self.q_proj.quantized_act_type().is_some() {
317            res = res.to_dtype(original_dtype)?;
318        }
319        Ok(res)
320    }
321}
322
323struct DecoderLayer {
324    self_attn: Attention,
325    mlp: MLP,
326    input_layernorm: RmsNorm,
327    post_attention_layernorm: RmsNorm,
328}
329
330impl DecoderLayer {
331    #[allow(clippy::too_many_arguments)]
332    fn new(
333        rotary_emb: Arc<RotaryEmbedding>,
334        cfg: &Config,
335        vb: ShardedVarBuilder,
336        lora_config: &[((String, String), LoraConfig)],
337        count: &mut usize,
338        ord: &Ordering,
339        mapper: &dyn DeviceMapper,
340        layer_idx: usize,
341        loading_isq: bool,
342        preload_adapters: &Option<HashMap<String, (ShardedVarBuilder, LoraConfig)>>,
343    ) -> Result<Self> {
344        let self_attn = Attention::new(
345            rotary_emb,
346            cfg,
347            vb.pp("self_attn"),
348            lora_config,
349            count,
350            ord,
351            mapper,
352            layer_idx,
353            loading_isq,
354            preload_adapters,
355        )?;
356        let mlp = MLP::new(
357            cfg,
358            vb.pp("mlp"),
359            lora_config,
360            count,
361            ord,
362            mapper,
363            layer_idx,
364            loading_isq,
365            preload_adapters,
366        )?;
367        let input_layernorm = RmsNorm::new_gemma(
368            cfg.hidden_size,
369            cfg.rms_norm_eps,
370            mapper.set_device(layer_idx, vb.pp("input_layernorm"), false),
371        )?;
372        let post_attention_layernorm = RmsNorm::new_gemma(
373            cfg.hidden_size,
374            cfg.rms_norm_eps,
375            mapper.set_device(layer_idx, vb.pp("post_attention_layernorm"), false),
376        )?;
377        Ok(Self {
378            self_attn,
379            mlp,
380            input_layernorm,
381            post_attention_layernorm,
382        })
383    }
384
385    #[allow(clippy::too_many_arguments)]
386    fn forward(
387        &self,
388        xs: &Tensor,
389        attention_mask: Option<&Tensor>,
390        seqlen_offsets: &[usize],
391        kv_cache: &mut Option<(Tensor, Tensor)>,
392        scalings: Option<Tensor>,
393        global_scaling_weight: f64,
394        is_scaling_pass: Option<f64>,
395        flash_params: &FlashParams,
396    ) -> Result<Tensor> {
397        let residual = xs;
398        let xs = self.input_layernorm.forward(xs)?;
399        let xs = self.self_attn.forward(
400            &xs,
401            attention_mask,
402            seqlen_offsets,
403            kv_cache,
404            scalings.clone(),
405            global_scaling_weight,
406            is_scaling_pass,
407            flash_params,
408        )?;
409        let xs = (xs + residual)?;
410        let residual = &xs;
411        let xs = self.mlp.forward(
412            &xs.apply(&self.post_attention_layernorm)?,
413            scalings.clone(),
414            global_scaling_weight,
415            is_scaling_pass,
416        )?;
417        residual + xs
418    }
419}
420
421pub struct XLoraModel {
422    embed_tokens: candle_nn::Embedding,
423    layers: Vec<DecoderLayer>,
424    norm: RmsNorm,
425    lm_head: Arc<dyn LinearLayerLike + Send + Sync>,
426    dtype: DType,
427    hidden_size: usize,
428    device: Device,
429    cache: EitherCache,
430    max_seq_len: usize,
431    xlora_classifier: Option<XLoraClassifier>,
432    mapper: Box<dyn DeviceMapper + Send + Sync>,
433    cfg: ModelConfigMetadata,
434}
435
436impl XLoraModel {
437    #[allow(clippy::too_many_arguments)]
438    pub fn new(
439        cfg: &Config,
440        vb: ShardedVarBuilder,
441        lora_config: &[((String, String), LoraConfig)],
442        xlora_config: Option<XLoraConfig>,
443        xlora_ordering: Ordering,
444        is_gptx: bool,
445        normal_loading_metadata: NormalLoadingMetadata,
446        preload_adapters: &Option<HashMap<String, (ShardedVarBuilder, LoraConfig)>>,
447    ) -> Result<Self> {
448        if let Some(ref quant_cfg) = &cfg.quantization_config {
449            tracing::info!(
450                "Using {} quantization: {}.",
451                quant_cfg.name(),
452                quant_cfg.get_bits_name(&vb)
453            );
454        }
455        let mapper = normal_loading_metadata.mapper;
456        let vb_m = vb.pp("model");
457
458        let embed_tokens = layers::embedding(
459            cfg.vocab_size,
460            cfg.hidden_size,
461            mapper.set_nm_device(vb_m.pp("embed_tokens"), false),
462            &cfg.quantization_config,
463        )?;
464        let mut layers = Vec::with_capacity(cfg.num_hidden_layers);
465        let vb_l = vb_m.pp("layers");
466        let mut ropes = HashMap::new();
467        for layer_idx in 0..cfg.num_hidden_layers {
468            let device = mapper
469                .device_for(layer_idx, false)
470                .unwrap_or(&normal_loading_metadata.real_device);
471            ropes.insert(
472                device.location(),
473                Arc::new(RotaryEmbedding::new(
474                    cfg.rope_theta as f32,
475                    cfg.head_dim,
476                    cfg.max_position_embeddings,
477                    device,
478                    is_gptx,
479                    vb.dtype(),
480                )?),
481            );
482        }
483
484        let mut count = 0;
485        for layer_idx in NiceProgressBar::<_, 'b'>(
486            0..cfg.num_hidden_layers,
487            "Loading repeating layers",
488            &normal_loading_metadata.multi_progress,
489        ) {
490            let device = mapper
491                .device_for(layer_idx, false)
492                .unwrap_or(&normal_loading_metadata.real_device);
493            let rotary_emb = ropes
494                .get(&device.location())
495                .expect("No RoPE for device location!")
496                .clone();
497            let layer = DecoderLayer::new(
498                rotary_emb.clone(),
499                cfg,
500                vb_l.pp(layer_idx),
501                lora_config,
502                &mut count,
503                &xlora_ordering,
504                &*mapper,
505                layer_idx,
506                normal_loading_metadata.loading_isq,
507                preload_adapters,
508            )?;
509            layers.push(layer)
510        }
511        if xlora_config.is_none() && preload_adapters.is_none() {
512            // We are now a LoRA model so we must merge the weights
513            info!("Merging LoRA adapters.");
514            for layer in layers.iter_mut().tqdm() {
515                Arc::get_mut(&mut layer.self_attn.k_proj)
516                    .unwrap()
517                    .merge_weights()?;
518                Arc::get_mut(&mut layer.self_attn.o_proj)
519                    .unwrap()
520                    .merge_weights()?;
521                Arc::get_mut(&mut layer.self_attn.q_proj)
522                    .unwrap()
523                    .merge_weights()?;
524                Arc::get_mut(&mut layer.self_attn.v_proj)
525                    .unwrap()
526                    .merge_weights()?;
527
528                Arc::get_mut(&mut layer.mlp.down_proj)
529                    .unwrap()
530                    .merge_weights()?;
531                Arc::get_mut(&mut layer.mlp.gate_proj)
532                    .unwrap()
533                    .merge_weights()?;
534                Arc::get_mut(&mut layer.mlp.up_proj)
535                    .unwrap()
536                    .merge_weights()?;
537            }
538        }
539        let norm = RmsNorm::new_gemma(
540            cfg.hidden_size,
541            cfg.rms_norm_eps,
542            mapper.set_nm_device(vb_m.pp("norm"), false),
543        )?;
544        let lm_head = linear(
545            embed_tokens.embeddings().dim(1)?,
546            embed_tokens.embeddings().dim(0)?,
547            false,
548            mapper.set_nm_device(vb_m.pp("embed_tokens"), normal_loading_metadata.loading_isq),
549            mapper.set_nm_device(vb_m.pp("embed_tokens"), false),
550            lora_config,
551            &mut count,
552            &xlora_ordering,
553            preload_adapters,
554        )?;
555        if xlora_config.is_some() && lm_head.is_lora() {
556            // This is why we can pass dummy values (..., None, 1.0, None)?
557            candle_core::bail!("Got an adapter `lm_head` layer, this is unsupported with X-LoRA.");
558        }
559
560        Ok(Self {
561            embed_tokens,
562            layers,
563            norm,
564            lm_head,
565            device: normal_loading_metadata.real_device,
566            dtype: vb.dtype(),
567            hidden_size: cfg.hidden_size,
568            cache: EitherCache::Full(Cache::new(cfg.num_hidden_layers, true)),
569            max_seq_len: default_max_position_embeddings(),
570            xlora_classifier: xlora_config.map(|xlora_config| {
571                XLoraClassifier::new(xlora_config, count, lora_config.len(), vb, false).unwrap()
572            }),
573            mapper,
574            cfg: ModelConfigMetadata {
575                max_seq_len: cfg.max_position_embeddings,
576                num_layers: cfg.num_hidden_layers,
577                hidden_size: cfg.hidden_size,
578                num_kv_heads: cfg.num_key_value_heads,
579                num_attn_heads: cfg.num_attention_heads,
580                sliding_window: None,
581                k_head_dim: cfg.head_dim,
582                v_head_dim: cfg.head_dim,
583                kv_cache_layout: crate::paged_attention::KvCacheLayout::Standard,
584            },
585        })
586    }
587
588    #[allow(clippy::too_many_arguments)]
589    fn inner_forward(
590        &self,
591        input_ids: &Tensor,
592        seqlen_offsets: &[usize],
593        scalings: Option<Tensor>,
594        is_full_pass: bool,
595        no_kv_cache: bool,
596        is_scaling_pass: Option<f64>,
597        flash_params: &FlashParams,
598    ) -> Result<Tensor> {
599        let mut cache = if is_full_pass {
600            if no_kv_cache {
601                let mut new_cache = Vec::new();
602                for _ in 0..self.cache.full().xlora_lock().len() {
603                    new_cache.push(None);
604                }
605
606                self.cache.full().xlora_lock().clone_from(&new_cache);
607            }
608            self.cache.full().xlora_lock()
609        } else {
610            self.cache.full().lock()
611        };
612        let xs = self.embed_tokens.forward(input_ids)?;
613        let attention_mask = CausalMasker.make_causal_mask_matrix(
614            input_ids,
615            &*cache,
616            xs.dtype(),
617            self.cfg.num_attn_heads,
618        )?;
619        let mut xs = (xs * (self.hidden_size as f64).sqrt())?;
620        for (i, layer) in self.layers.iter().enumerate() {
621            xs = self.mapper.map(xs, i)?;
622            xs = layer.forward(
623                &xs,
624                attention_mask
625                    .as_ref()
626                    .map(|m| m.to_device(xs.device()).unwrap())
627                    .as_ref(),
628                seqlen_offsets,
629                &mut cache[i],
630                scalings.clone(),
631                self.xlora_classifier
632                    .as_ref()
633                    .map(|classifier| classifier.get_global_scaling_weight())
634                    .unwrap_or(1.0),
635                is_scaling_pass,
636                flash_params,
637            )?
638        }
639        let xs = xs.to_device(&self.device)?;
640        xs.apply(&self.norm)
641    }
642
643    #[allow(clippy::too_many_arguments)]
644    pub fn forward(
645        &self,
646        input_ids: &Tensor,
647        input_ids_full: &Tensor,
648        seqlen_offsets: &[usize],
649        seqlen_offsets_full: &[usize],
650        no_kv_cache: bool,
651        non_granular_state: &Option<NonGranularState>,
652        context_lens: Vec<(usize, usize)>,
653        flash_params: &FlashParams,
654        flash_params_full: &FlashParams,
655    ) -> Result<Tensor> {
656        if self.xlora_classifier.is_some() {
657            let scalings = self.get_scalings(
658                input_ids,
659                input_ids_full,
660                seqlen_offsets,
661                seqlen_offsets_full,
662                no_kv_cache,
663                non_granular_state,
664                &vec![usize::MAX; context_lens.len()],
665                flash_params,
666                flash_params_full,
667            )?;
668
669            if no_kv_cache {
670                let mut res = self
671                    .inner_forward(
672                        input_ids_full,
673                        seqlen_offsets_full,
674                        Some(scalings),
675                        true,
676                        no_kv_cache,
677                        None,
678                        flash_params_full,
679                    )?
680                    .contiguous()?;
681                if let Some(t) = self.lm_head.quantized_act_type() {
682                    res = res.to_dtype(t)?;
683                }
684                extract_logits(
685                    &self.lm_head.lora_forward(&res, None, 1.0, None)?,
686                    context_lens,
687                )
688            } else {
689                // is_full_pass=true is ok because no_kv_cache=false
690                let mut res = self
691                    .inner_forward(
692                        input_ids,
693                        seqlen_offsets,
694                        Some(scalings),
695                        true,
696                        no_kv_cache,
697                        None,
698                        flash_params,
699                    )?
700                    .contiguous()?;
701                if let Some(t) = self.lm_head.quantized_act_type() {
702                    res = res.to_dtype(t)?;
703                }
704                extract_logits(
705                    &self.lm_head.lora_forward(&res, None, 1.0, None)?,
706                    context_lens,
707                )
708            }
709        } else {
710            let mut res = self
711                .inner_forward(
712                    input_ids,
713                    seqlen_offsets,
714                    None,
715                    false,
716                    no_kv_cache,
717                    None,
718                    flash_params,
719                )?
720                .contiguous()?;
721            if let Some(t) = self.lm_head.quantized_act_type() {
722                res = res.to_dtype(t)?;
723            }
724            extract_logits(
725                &self.lm_head.lora_forward(&res, None, 1.0, None)?,
726                context_lens,
727            )
728        }
729    }
730}
731
732impl IsqModel for XLoraModel {
733    fn get_layers(
734        &mut self,
735    ) -> (
736        Vec<(&mut Arc<dyn QuantMethod>, Option<usize>)>,
737        &dyn DeviceMapper,
738    ) {
739        let mut tensors = Vec::new();
740        tensors.push((Arc::get_mut(&mut self.lm_head).unwrap().quant_inner(), None));
741        for (i, layer) in self.layers.iter_mut().enumerate() {
742            tensors.push((
743                Arc::get_mut(&mut layer.self_attn.q_proj)
744                    .unwrap()
745                    .quant_inner(),
746                Some(i),
747            ));
748            tensors.push((
749                Arc::get_mut(&mut layer.self_attn.k_proj)
750                    .unwrap()
751                    .quant_inner(),
752                Some(i),
753            ));
754            tensors.push((
755                Arc::get_mut(&mut layer.self_attn.v_proj)
756                    .unwrap()
757                    .quant_inner(),
758                Some(i),
759            ));
760            tensors.push((
761                Arc::get_mut(&mut layer.self_attn.o_proj)
762                    .unwrap()
763                    .quant_inner(),
764                Some(i),
765            ));
766            tensors.push((
767                Arc::get_mut(&mut layer.mlp.gate_proj)
768                    .unwrap()
769                    .quant_inner(),
770                Some(i),
771            ));
772            tensors.push((
773                Arc::get_mut(&mut layer.mlp.up_proj).unwrap().quant_inner(),
774                Some(i),
775            ));
776            tensors.push((
777                Arc::get_mut(&mut layer.mlp.down_proj)
778                    .unwrap()
779                    .quant_inner(),
780                Some(i),
781            ));
782        }
783        (tensors, &*self.mapper)
784    }
785
786    fn residual_tensors(&self) -> Vec<(String, Tensor)> {
787        panic!("Cannot generate UQFF for an adapter model.")
788    }
789}
790
791impl NormalModel for XLoraModel {
792    fn forward(
793        &self,
794        _input_ids: &Tensor,
795        _seqlen_offsets: &[usize],
796        _context_lens: Vec<(usize, usize)>,
797        _position_ids: Vec<usize>,
798        _metadata: Option<(Vec<(Tensor, Tensor)>, &PagedAttentionInputMetadata)>,
799        _flash_params: &FlashParams,
800    ) -> Result<Tensor> {
801        unreachable!()
802    }
803    fn xlora_forward(
804        &self,
805        input_ids: &Tensor,
806        input_ids_full: &Tensor,
807        seqlen_offsets: &[usize],
808        seqlen_offsets_full: &[usize],
809        no_kv_cache: bool,
810        non_granular_state: &Option<crate::xlora_models::NonGranularState>,
811        context_lens: Vec<(usize, usize)>,
812        _position_ids: Vec<usize>,
813        flash_params: &FlashParams,
814        flash_params_full: &FlashParams,
815    ) -> Result<Tensor> {
816        self.forward(
817            input_ids,
818            input_ids_full,
819            seqlen_offsets,
820            seqlen_offsets_full,
821            no_kv_cache,
822            non_granular_state,
823            context_lens,
824            flash_params,
825            flash_params_full,
826        )
827    }
828    fn cache(&self) -> &EitherCache {
829        &self.cache
830    }
831    fn cache_mut(&mut self) -> &mut EitherCache {
832        &mut self.cache
833    }
834    fn device(&self) -> &Device {
835        &self.device
836    }
837    fn is_xlora(&self) -> bool {
838        true
839    }
840    fn max_seq_len(&self) -> usize {
841        self.max_seq_len
842    }
843    fn config(&self) -> &ModelConfigMetadata {
844        &self.cfg
845    }
846}
847
848impl ScalingsMaker for XLoraModel {
849    fn dtype(&self) -> DType {
850        self.dtype
851    }
852    fn get_cache(&self) -> &EitherCache {
853        &self.cache
854    }
855    fn get_classifier(&self) -> &XLoraClassifier {
856        self.xlora_classifier.as_ref().unwrap()
857    }
858    fn forward(
859        &self,
860        input_ids: &Tensor,
861        seqlen_offsets: &[usize],
862        scalings: Tensor,
863        is_full_pass: bool,
864        no_kv_cache: bool,
865        is_scaling_pass: Option<f64>,
866        _context_lens: &[usize],
867        flash_params: &FlashParams,
868    ) -> Result<Tensor> {
869        self.inner_forward(
870            input_ids,
871            seqlen_offsets,
872            Some(scalings),
873            is_full_pass,
874            no_kv_cache,
875            is_scaling_pass,
876            flash_params,
877        )
878    }
879}
880
881impl AnyMoeBaseModelMixin for XLoraModel {}
mistralrs_core/xlora_models/gemma.rs

mistralrs_core/xlora_models/
gemma.rs
