mistralrs_core/pipeline/
normal.rs

1use super::inputs_processor::DEFAULT_PROMPT_CHUNK_SIZE;
2use super::isq::ImatrixDataSource;
3use super::llg::build_llg_factory;
4use super::{
5    get_model_paths, get_xlora_paths, text_models_inputs_processor::ModelInputs, AdapterKind,
6    CacheManager, GeneralMetadata, Loader, ModelKind, ModelPaths, NormalModel, NormalModelLoader,
7    TokenSource,
8};
9use super::{
10    AnyMoePipelineMixin, CacheManagerMixin, EitherCache, ForwardInputsResult, IsqOrganization,
11    IsqPipelineMixin, MetadataMixin, ModelCategory, PreProcessingMixin,
12};
13use super::{
14    AutoNormalLoader, DeepSeekV2Loader, DeepSeekV3Loader, GLM4Loader, Gemma2Loader, GemmaLoader,
15    LlamaLoader, MistralLoader, MixtralLoader, NormalLoaderType, Phi2Loader, Phi3Loader,
16    Phi3_5MoELoader, Qwen2Loader, Qwen3Loader, Qwen3MoELoader, Starcoder2Loader,
17};
18use crate::amoe::AnyMoeExpertType;
19use crate::device_map::{self, DeviceMapper};
20use crate::distributed::{self, WorkerTransferData};
21use crate::kv_cache::{FullCacheManager, NormalCacheManager};
22use crate::lora::Ordering;
23use crate::paged_attention::{calculate_cache_config, AttentionImplementation, CacheEngine};
24use crate::pipeline::chat_template::{calculate_eos_tokens, GenerationConfig};
25use crate::pipeline::isq::UqffFullSer;
26use crate::pipeline::loaders::auto_device_map;
27use crate::pipeline::loaders::QuantizationConfigShim;
28use crate::pipeline::sampling::sample_and_add_toks;
29use crate::pipeline::text_models_inputs_processor::make_prompt_chunk;
30use crate::pipeline::{get_chat_template, Modalities, SupportedModality};
31use crate::pipeline::{ChatTemplate, LocalModelPaths};
32use crate::prefix_cacher::PrefixCacheManagerV2;
33use crate::sequence::Sequence;
34use crate::utils::tokenizer::get_tokenizer;
35use crate::utils::varbuilder_utils::DeviceForLoadTensor;
36use crate::utils::{tokens::get_token, varbuilder_utils::from_mmaped_safetensors};
37use crate::xlora_models::NonGranularState;
38use crate::{
39    api_dir_list, api_get_file, get_mut_arcmutex, get_paths, get_uqff_paths, lora_model_loader,
40    normal_model_loader, normal_model_loader_sharded, xlora_model_loader, DeviceMapSetting,
41    PagedAttentionConfig, Pipeline, Topology, TryIntoDType, GLOBAL_HF_CACHE,
42};
43use anyhow::Result;
44use candle_core::{Device, Tensor, Var};
45use hf_hub::Cache;
46use hf_hub::{api::sync::ApiBuilder, Repo, RepoType};
47use indicatif::MultiProgress;
48use mistralrs_quant::log::once_log_info;
49use mistralrs_quant::{AfqLayer, GgufMatMul, HqqLayer, IsqType, QuantizedSerdeType};
50use rand_isaac::Isaac64Rng;
51use regex_automata::meta::Regex;
52use std::any::Any;
53use std::borrow::Cow;
54use std::num::{NonZero, NonZeroUsize};
55use std::path::{Path, PathBuf};
56use std::str::FromStr;
57use std::sync::{Arc, RwLock};
58use std::time::Instant;
59use std::{env, fs};
60use tokenizers::Tokenizer;
61use tokio::sync::Mutex;
62use tracing::{info, warn};
63
64pub struct NormalPipeline {
65    model: Box<dyn NormalModel + Send + Sync>,
66    tokenizer: Arc<Tokenizer>,
67    no_kv_cache: bool,
68    chat_template: Arc<ChatTemplate>,
69    non_granular_state: Option<NonGranularState>,
70    model_id: String,
71    metadata: Arc<GeneralMetadata>,
72    topology: Option<Topology>,
73    silent: bool,
74    organization: IsqOrganization,
75    // For full UQFF serialization
76    template_filename: Option<PathBuf>,
77    generation_config: Option<PathBuf>,
78    config: String,
79    imatrix: Option<PathBuf>,
80    mapper: Box<dyn DeviceMapper + Send + Sync>,
81}
82
83/// A loader for a "normal" (non-quantized) model.
84pub struct NormalLoader {
85    inner: Box<dyn NormalModelLoader>,
86    model_id: String,
87    config: NormalSpecificConfig,
88    xlora_model_id: Option<String>,
89    lora_adapter_ids: Option<Vec<String>>,
90    kind: ModelKind,
91    xlora_order: Option<Ordering>,
92    no_kv_cache: bool,
93    chat_template: Option<String>,
94    tokenizer_json: Option<String>,
95    tgt_non_granular_index: Option<usize>,
96    token_source: RwLock<Option<TokenSource>>,
97    revision: RwLock<Option<String>>,
98    from_uqff: RwLock<Option<Vec<PathBuf>>>,
99    jinja_explicit: Option<String>,
100    hf_cache_path: Option<PathBuf>,
101}
102
103#[derive(Default)]
104/// A builder for a loader for a "normal" (non-quantized) model.
105pub struct NormalLoaderBuilder {
106    model_id: Option<String>,
107    config: NormalSpecificConfig,
108    xlora_model_id: Option<String>,
109    lora_adapter_ids: Option<Vec<String>>,
110    kind: ModelKind,
111    xlora_order: Option<Ordering>,
112    no_kv_cache: bool,
113    chat_template: Option<String>,
114    tokenizer_json: Option<String>,
115    tgt_non_granular_index: Option<usize>,
116    jinja_explicit: Option<String>,
117    hf_cache_path: Option<PathBuf>,
118}
119
120#[derive(Clone, Default)]
121/// Config specific to loading a normal model.
122pub struct NormalSpecificConfig {
123    pub prompt_chunksize: Option<NonZeroUsize>,
124    pub topology: Option<Topology>,
125    pub organization: IsqOrganization,
126    pub write_uqff: Option<PathBuf>,
127    pub from_uqff: Option<Vec<PathBuf>>,
128    pub imatrix: Option<PathBuf>,
129    pub calibration_file: Option<PathBuf>,
130    pub hf_cache_path: Option<PathBuf>,
131}
132
133impl NormalLoaderBuilder {
134    pub fn new(
135        config: NormalSpecificConfig,
136        chat_template: Option<String>,
137        tokenizer_json: Option<String>,
138        model_id: Option<String>,
139        no_kv_cache: bool,
140        jinja_explicit: Option<String>,
141    ) -> Self {
142        Self {
143            config,
144            chat_template,
145            tokenizer_json,
146            model_id,
147            kind: ModelKind::Normal,
148            jinja_explicit,
149            no_kv_cache,
150            ..Default::default()
151        }
152    }
153
154    fn with_adapter(
155        mut self,
156        xlora_model_id: String,
157        xlora_order: Ordering,
158        no_kv_cache: bool,
159        tgt_non_granular_index: Option<usize>,
160    ) -> Self {
161        self.xlora_model_id = Some(xlora_model_id);
162        self.xlora_order = Some(xlora_order);
163        self.no_kv_cache = no_kv_cache;
164        self.tgt_non_granular_index = tgt_non_granular_index;
165        self.model_id = if let Some(id) = self.model_id {
166            Some(id)
167        } else {
168            info!(
169                "Using adapter base model ID: `{}`",
170                self.xlora_order.as_ref().unwrap().base_model_id
171            );
172            Some(self.xlora_order.as_ref().unwrap().base_model_id.clone())
173        };
174        self
175    }
176
177    pub fn with_xlora(
178        mut self,
179        xlora_model_id: String,
180        xlora_order: Ordering,
181        no_kv_cache: bool,
182        tgt_non_granular_index: Option<usize>,
183    ) -> Self {
184        self.kind = ModelKind::Adapter {
185            adapter: AdapterKind::XLora,
186        };
187        self.with_adapter(
188            xlora_model_id,
189            xlora_order,
190            no_kv_cache,
191            tgt_non_granular_index,
192        )
193    }
194
195    pub fn with_lora(mut self, lora_adapter_ids: Vec<String>) -> Self {
196        self.kind = ModelKind::Adapter {
197            adapter: AdapterKind::Lora,
198        };
199        self.lora_adapter_ids = Some(lora_adapter_ids);
200        self
201    }
202
203    pub fn hf_cache_path(mut self, hf_cache_path: PathBuf) -> Self {
204        self.hf_cache_path = Some(hf_cache_path);
205        self
206    }
207
208    /// If the loader type is not specified, loader type is automatically determined from the
209    /// `architectures` array in the config.
210    pub fn build(self, loader_tp: Option<NormalLoaderType>) -> anyhow::Result<Box<dyn Loader>> {
211        let loader: Box<dyn NormalModelLoader> = match loader_tp {
212            Some(NormalLoaderType::Mistral) => Box::new(MistralLoader),
213            Some(NormalLoaderType::Gemma) => Box::new(GemmaLoader),
214            Some(NormalLoaderType::Llama) => Box::new(LlamaLoader),
215            Some(NormalLoaderType::Mixtral) => Box::new(MixtralLoader),
216            Some(NormalLoaderType::Phi2) => Box::new(Phi2Loader),
217            Some(NormalLoaderType::Phi3) => Box::new(Phi3Loader),
218            Some(NormalLoaderType::Qwen2) => Box::new(Qwen2Loader),
219            Some(NormalLoaderType::Gemma2) => Box::new(Gemma2Loader),
220            Some(NormalLoaderType::Starcoder2) => Box::new(Starcoder2Loader),
221            Some(NormalLoaderType::Phi3_5MoE) => Box::new(Phi3_5MoELoader),
222            Some(NormalLoaderType::DeepSeekV2) => Box::new(DeepSeekV2Loader),
223            Some(NormalLoaderType::DeepSeekV3) => Box::new(DeepSeekV3Loader),
224            Some(NormalLoaderType::Qwen3) => Box::new(Qwen3Loader),
225            Some(NormalLoaderType::GLM4) => Box::new(GLM4Loader),
226            Some(NormalLoaderType::Qwen3Moe) => Box::new(Qwen3MoELoader),
227            None => Box::new(AutoNormalLoader),
228        };
229        Ok(Box::new(NormalLoader {
230            inner: loader,
231            model_id: self.model_id.unwrap(),
232            config: self.config,
233            xlora_model_id: self.xlora_model_id,
234            lora_adapter_ids: self.lora_adapter_ids,
235            kind: self.kind,
236            xlora_order: self.xlora_order,
237            no_kv_cache: self.no_kv_cache,
238            chat_template: self.chat_template,
239            tokenizer_json: self.tokenizer_json,
240            tgt_non_granular_index: self.tgt_non_granular_index,
241            jinja_explicit: self.jinja_explicit,
242            token_source: RwLock::new(None),
243            revision: RwLock::new(None),
244            from_uqff: RwLock::new(None),
245            hf_cache_path: self.hf_cache_path,
246        }))
247    }
248}
249
250impl Loader for NormalLoader {
251    #[allow(clippy::type_complexity, clippy::too_many_arguments)]
252    fn load_model_from_hf(
253        &self,
254        revision: Option<String>,
255        token_source: TokenSource,
256        dtype: &dyn TryIntoDType,
257        device: &Device,
258        silent: bool,
259        mapper: DeviceMapSetting,
260        in_situ_quant: Option<IsqType>,
261        paged_attn_config: Option<PagedAttentionConfig>,
262    ) -> Result<Arc<Mutex<dyn Pipeline + Send + Sync>>> {
263        let cache = self
264            .hf_cache_path
265            .clone()
266            .map(Cache::new)
267            .unwrap_or_default();
268        GLOBAL_HF_CACHE.get_or_init(|| cache);
269
270        let paths: anyhow::Result<Box<dyn ModelPaths>> = get_paths!(
271            LocalModelPaths,
272            &token_source,
273            revision.clone(),
274            self,
275            None,
276            None,
277            silent,
278            self.config.from_uqff.is_some()
279        );
280        if let Some(from_uqff) = self.config.from_uqff.clone() {
281            *self.from_uqff.write().unwrap() = Some(get_uqff_paths!(&from_uqff, self, silent));
282        }
283        *self
284            .token_source
285            .write()
286            .expect("Failed to write to token source") = Some(token_source);
287        *self.revision.write().expect("Failed to write to revision") = revision;
288        self.load_model_from_path(
289            &paths?,
290            dtype,
291            device,
292            silent,
293            mapper,
294            in_situ_quant,
295            paged_attn_config,
296        )
297    }
298
299    #[allow(clippy::type_complexity, clippy::too_many_arguments)]
300    fn load_model_from_path(
301        &self,
302        paths: &Box<dyn ModelPaths>,
303        dtype: &dyn TryIntoDType,
304        device: &Device,
305        silent: bool,
306        mut mapper: DeviceMapSetting,
307        mut in_situ_quant: Option<IsqType>,
308        mut paged_attn_config: Option<PagedAttentionConfig>,
309    ) -> Result<Arc<Mutex<dyn Pipeline + Send + Sync>>> {
310        let config = std::fs::read_to_string(paths.get_config_filename())?;
311
312        if !self.inner.supports_paged_attention(&config)? {
313            paged_attn_config = None;
314        }
315
316        // Apply default prompt size here
317        let prompt_chunksize = self
318            .config
319            .prompt_chunksize
320            .unwrap_or(DEFAULT_PROMPT_CHUNK_SIZE.try_into().unwrap())
321            .get();
322
323        info!("Prompt chunk size is {prompt_chunksize}.",);
324
325        let use_nccl = mistralrs_quant::distributed::use_nccl();
326
327        let available_devices = if let Ok(payload) = env::var(distributed::IS_DAEMON_FLAG) {
328            let payload: WorkerTransferData = serde_json::from_str(&payload)?;
329            let WorkerTransferData::Init { id: _, worker_rank } = payload;
330            vec![candle_core::Device::new_cuda(worker_rank + 1)?]
331        } else if use_nccl {
332            vec![candle_core::Device::new_cuda(0)?]
333        } else {
334            device_map::get_all_similar_devices(device)?
335        };
336        let device = if use_nccl || cfg!(feature = "ring") {
337            available_devices[0].clone()
338        } else {
339            device.clone()
340        };
341
342        // If auto, convert to Map if not using nccl
343        if use_nccl || cfg!(feature = "ring") {
344            mapper = DeviceMapSetting::DummyNccl {
345                nm_device: available_devices[0].clone(),
346            };
347        } else if let DeviceMapSetting::Auto(params) = mapper.clone() {
348            // Initial dtype
349            let dtype = dtype.try_into_dtype(&available_devices.iter().collect::<Vec<_>>())?;
350
351            // Disable ISQ if we are loading a prequantized model.
352            if QuantizationConfigShim::get_quant_config_pack_factor(&config, dtype)? != 1 {
353                in_situ_quant = None;
354            }
355
356            // ISQ or UQFF: quantized path
357            // Match logic below where UQFF has priority
358            let (layer_sizes_in_bytes, non_mapped_size_in_bytes, total_model_size_in_bytes) =
359                if let Some(serialized) = &*self.from_uqff.read().unwrap() {
360                    let weight_pack_factor = {
361                        let ser_artifacts = unsafe {
362                            candle_core::safetensors::MmapedSafetensors::multi(serialized)?
363                        };
364                        let mut total_pack_factors = 0;
365                        let total_tensors = ser_artifacts.tensors().len();
366                        for (_, artifact) in ser_artifacts.tensors() {
367                            let artifact = artifact.data();
368                            // NOTE(EricLBuehler): isq type is ALWAYS byte 4 (5th) of the tensor.
369                            let isq_type = artifact[mistralrs_quant::UQFF_QUANT_TYPE_OFFSET];
370                            let pack_factor = match QuantizedSerdeType::try_from(isq_type as usize)?
371                            {
372                                QuantizedSerdeType::Hqq => {
373                                    HqqLayer::get_isq_type_from_uqff(Cow::Borrowed(artifact))?
374                                        .pack_factor(dtype)
375                                }
376                                QuantizedSerdeType::Gguf => {
377                                    GgufMatMul::get_isq_type_from_uqff(Cow::Borrowed(artifact))?
378                                        .pack_factor(dtype)
379                                }
380                                QuantizedSerdeType::Fp8 => IsqType::F8E4M3.pack_factor(dtype),
381                                QuantizedSerdeType::Unquant => 1,
382                                QuantizedSerdeType::Afq => {
383                                    AfqLayer::get_isq_type_from_uqff(Cow::Borrowed(artifact))?
384                                        .pack_factor(dtype)
385                                }
386                            };
387                            total_pack_factors += pack_factor;
388                        }
389
390                        total_pack_factors / total_tensors
391                    };
392
393                    let layer_sizes_in_bytes =
394                        self.inner
395                            .layer_sizes_in_bytes(&config, dtype, weight_pack_factor)?;
396                    let non_mapped_size_in_bytes =
397                        self.inner
398                            .non_mapped_size_in_bytes(&config, dtype, weight_pack_factor)?;
399                    let layer_sizes_sum = layer_sizes_in_bytes.iter().sum::<usize>();
400                    (
401                        layer_sizes_in_bytes,
402                        non_mapped_size_in_bytes,
403                        layer_sizes_sum + non_mapped_size_in_bytes,
404                    )
405                } else if let Some(isq) = in_situ_quant {
406                    let weight_pack_factor = isq.pack_factor(dtype);
407                    let layer_sizes_in_bytes =
408                        self.inner
409                            .layer_sizes_in_bytes(&config, dtype, weight_pack_factor)?;
410                    let non_mapped_size_in_bytes =
411                        self.inner
412                            .non_mapped_size_in_bytes(&config, dtype, weight_pack_factor)?;
413                    let layer_sizes_sum = layer_sizes_in_bytes.iter().sum::<usize>();
414                    (
415                        layer_sizes_in_bytes,
416                        non_mapped_size_in_bytes,
417                        layer_sizes_sum + non_mapped_size_in_bytes,
418                    )
419                } else {
420                    // Be sure to get the weight pack factor here; we might be loading a prequantized model.
421                    let weight_pack_factor =
422                        QuantizationConfigShim::get_quant_config_pack_factor(&config, dtype)?;
423                    let layer_sizes_in_bytes =
424                        self.inner
425                            .layer_sizes_in_bytes(&config, dtype, weight_pack_factor)?;
426                    let non_mapped_size_in_bytes =
427                        self.inner
428                            .non_mapped_size_in_bytes(&config, dtype, weight_pack_factor)?;
429                    let layer_sizes_sum = layer_sizes_in_bytes.iter().sum::<usize>();
430                    (
431                        layer_sizes_in_bytes,
432                        non_mapped_size_in_bytes,
433                        layer_sizes_sum + non_mapped_size_in_bytes,
434                    )
435                };
436
437            let new = auto_device_map::get_device_layers(
438                &*self.inner,
439                &config,
440                self.inner.num_layers(&config)?,
441                layer_sizes_in_bytes,
442                non_mapped_size_in_bytes,
443                total_model_size_in_bytes,
444                &available_devices,
445                dtype,
446                &params,
447                prompt_chunksize,
448                paged_attn_config.as_ref(),
449            )?;
450            mapper = DeviceMapSetting::Map(new);
451        }
452
453        let pipeline_mapper = mapper.into_mapper(
454            self.inner.num_layers(&config)?,
455            &device,
456            self.config.topology.as_ref(),
457        )?;
458        let mapper = mapper.into_mapper(
459            self.inner.num_layers(&config)?,
460            &device,
461            self.config.topology.as_ref(),
462        )?;
463        let mut layer_devices = Vec::new();
464        for layer in 0..self.inner.num_layers(&config)? {
465            let device = mapper.device_for(layer, false).cloned();
466            layer_devices.push(device);
467        }
468        let dtype = mapper.get_min_dtype(dtype)?;
469
470        // TODO: PagedAttention is not supported with CPU for now.
471        // This check is not really necessary because `get_device_layers` should prevent it.
472        let mapping_uses_cpu = mapper.get_unique_devices().iter().any(Device::is_cpu);
473        if mapping_uses_cpu {
474            warn!("Device mapping contains a mix of GPU and CPU. There is no CPU support for PagedAttention, disabling PagedAttention.");
475            paged_attn_config = None;
476        }
477
478        info!("Model config: {:?}", self.inner.get_config_repr(&config)?);
479        if crate::using_flash_attn() {
480            once_log_info("FlashAttention is enabled.");
481        }
482
483        // Logic for ISQ here: if no calibration (i.e imatrix), then allow immediate ISQ. Otherwise, back to normal.
484        let mut loading_isq = if self.config.imatrix.is_none()
485            && self.config.calibration_file.is_none()
486            && !device.is_cuda()
487            && self.config.write_uqff.is_none()
488            && in_situ_quant.is_some()
489        {
490            let predicates = if matches!(self.config.organization, IsqOrganization::MoeExpertsOnly)
491            {
492                self.inner.immediate_isq_predicates_moqe(&config)?
493            } else {
494                self.inner.immediate_isq_predicates(&config)?
495            };
496            info!("Applying ISQ to {in_situ_quant:?}");
497            if predicates.is_empty() {
498                warn!("No predicates for this model and ISQ setting detected. ISQ will not be applied to any weights!");
499            }
500            mistralrs_quant::set_immediate_isq(in_situ_quant, predicates);
501            false
502        } else {
503            in_situ_quant.is_some()
504        };
505
506        if let Some(ref topology) = self.config.topology {
507            loading_isq |= topology
508                .0
509                .iter()
510                .any(|layer| layer.as_ref().is_some_and(|layer| layer.isq.is_some()));
511        }
512
513        if self.config.imatrix.is_some() && self.config.calibration_file.is_some() {
514            anyhow::bail!(
515                "`imatrix` and `calibration_file` were both specified, this is not allowed."
516            );
517        }
518
519        // Load onto the regular device if not using isq or if the calibration file is specified
520        let load_device = if !loading_isq || self.config.calibration_file.is_some() {
521            loading_isq = false;
522            device.clone()
523        } else {
524            Device::Cpu
525        };
526
527        let is_xlora = self.kind.is_adapted_and(|a| a.is_x_lora());
528
529        let attention_mechanism = if paged_attn_config.is_some() {
530            AttentionImplementation::PagedAttention
531        } else {
532            AttentionImplementation::Eager
533        };
534
535        let multi_progress = Arc::new(MultiProgress::new());
536
537        let mut model = if use_nccl || cfg!(feature = "ring") {
538            let (mapper, sharded_vb) = distributed::prepare_distributed_mapper(
539                dtype,
540                &device,
541                &available_devices,
542                silent,
543                &config,
544                loading_isq,
545                self.config.from_uqff.is_some(),
546                self.config.organization,
547                &*self.inner,
548                paths.as_ref(),
549            )?;
550
551            // Special case for where things can be more optimially loaded.
552            match self.kind {
553                ModelKind::Normal => normal_model_loader_sharded!(
554                    sharded_vb,
555                    config,
556                    self.inner,
557                    mapper,
558                    loading_isq,
559                    device.clone(),
560                    attention_mechanism,
561                    multi_progress.clone(),
562                ),
563                ModelKind::Adapter {
564                    adapter: AdapterKind::XLora,
565                } => xlora_model_loader!(
566                    paths,
567                    Some(dtype),
568                    &load_device,
569                    layer_devices.clone(),
570                    config,
571                    self.inner,
572                    silent,
573                    mapper,
574                    loading_isq,
575                    device.clone(),
576                    multi_progress.clone(),
577                ),
578                ModelKind::Adapter {
579                    adapter: AdapterKind::Lora,
580                } => lora_model_loader!(
581                    paths,
582                    Some(dtype),
583                    &load_device,
584                    layer_devices.clone(),
585                    config,
586                    self.inner,
587                    silent,
588                    mapper,
589                    loading_isq,
590                    self.config.from_uqff.is_some(),
591                    device.clone(),
592                    attention_mechanism,
593                    matches!(self.config.organization, IsqOrganization::MoeExpertsOnly),
594                    multi_progress.clone(),
595                ),
596                _ => unreachable!(),
597            }
598        } else {
599            match self.kind {
600                ModelKind::Normal => normal_model_loader!(
601                    paths,
602                    Some(dtype),
603                    &load_device,
604                    layer_devices.clone(),
605                    config,
606                    self.inner,
607                    silent,
608                    mapper,
609                    loading_isq,
610                    self.config.from_uqff.is_some(),
611                    device.clone(),
612                    attention_mechanism,
613                    matches!(self.config.organization, IsqOrganization::MoeExpertsOnly),
614                    multi_progress.clone(),
615                ),
616                ModelKind::Adapter {
617                    adapter: AdapterKind::XLora,
618                } => xlora_model_loader!(
619                    paths,
620                    Some(dtype),
621                    &load_device,
622                    layer_devices.clone(),
623                    config,
624                    self.inner,
625                    silent,
626                    mapper,
627                    loading_isq,
628                    device.clone(),
629                    multi_progress.clone(),
630                ),
631                ModelKind::Adapter {
632                    adapter: AdapterKind::Lora,
633                } => lora_model_loader!(
634                    paths,
635                    Some(dtype),
636                    &load_device,
637                    layer_devices.clone(),
638                    config,
639                    self.inner,
640                    silent,
641                    mapper,
642                    loading_isq,
643                    self.config.from_uqff.is_some(),
644                    device.clone(),
645                    attention_mechanism,
646                    matches!(self.config.organization, IsqOrganization::MoeExpertsOnly),
647                    multi_progress.clone(),
648                ),
649                _ => unreachable!(),
650            }
651        };
652
653        let tokenizer = get_tokenizer(paths.get_tokenizer_filename(), None)?;
654        let gen_conf: Option<GenerationConfig> = paths.get_gen_conf_filename().and_then(|f| {
655            match serde_json::from_str::<GenerationConfig>(&fs::read_to_string(f).unwrap()) {
656                Ok(conf) => Some(conf),
657                Err(e) => {
658                    warn!("Failed to parse generation_config.json: {}", e);
659                    None
660                }
661            }
662        });
663
664        let chat_template_explicit = paths
665            .get_chat_template_explicit()
666            .as_ref()
667            .map(|x| x.to_string_lossy().to_string());
668        let chat_template = get_chat_template(
669            paths,
670            self.jinja_explicit.as_ref(),
671            chat_template_explicit.as_ref(),
672            self.chat_template.as_ref(),
673            None,
674        );
675
676        if let Some(calibration_file) = &self.config.calibration_file {
677            let calibration_data = std::fs::read_to_string(calibration_file)?;
678            // Tokenize, don't add bos yet
679            let tokens = tokenizer
680                .encode_fast(calibration_data, false)
681                .map_err(anyhow::Error::msg)?
682                .get_ids()
683                .to_vec();
684            info!(
685                "Collecting imatrix from calibration file `{}` of {} tokens.",
686                calibration_file.display(),
687                tokens.len()
688            );
689            let bos_toks = chat_template.bos_tok().map(|b| vec![b]).unwrap_or_default();
690            let bos_tok_id = tokenizer
691                .token_to_id(&bos_toks[0])
692                .expect("Somehow the bos token is not present.");
693
694            match self.config.organization {
695                IsqOrganization::Default => model.begin_track_stats()?,
696                IsqOrganization::MoeExpertsOnly => model.begin_track_stats_moe_experts_only()?,
697            }
698
699            const CHUNK_SIZE: usize = 1024;
700            let n_chunks = tokens.len().div_ceil(CHUNK_SIZE);
701            let start = Instant::now();
702            for (i, chunk) in tokens.chunks(CHUNK_SIZE).enumerate() {
703                let chunk = [vec![bos_tok_id], chunk.to_vec()].concat();
704                let chunk_len = chunk.len();
705
706                let start = Instant::now();
707                let inputs = make_prompt_chunk(
708                    0,
709                    vec![&chunk],
710                    &[0],
711                    &load_device,
712                    None,
713                    false,
714                    None,
715                    Some(pipeline_mapper.as_ref()),
716                )?;
717
718                model.forward(
719                    &inputs.input.to_device(model.device())?,
720                    &inputs.positions,
721                    inputs.context_lens.clone(),
722                    inputs.position_ids.clone(),
723                    None,
724                    &inputs.flash_meta.clone(),
725                )?;
726
727                match model.cache_mut() {
728                    EitherCache::Full(full) => {
729                        for layer in &mut *full.lock() {
730                            *layer = None
731                        }
732                    }
733                    EitherCache::Normal(normal) => {
734                        for layer in &mut *normal.lock().unwrap().0 {
735                            layer.reset();
736                        }
737                    }
738                }
739
740                let end = Instant::now();
741                info!(
742                    "Processed chunk {}/{n_chunks} ({chunk_len} tokens), {:.2}s",
743                    i + 1,
744                    end.duration_since(start).as_secs_f32()
745                );
746            }
747            load_device.synchronize()?;
748            let end = Instant::now();
749            info!(
750                "Finished collecting imatrix in {:.2}s",
751                end.duration_since(start).as_secs_f32()
752            );
753        }
754
755        // Only if loading from UQFF
756        if (loading_isq || self.config.topology.is_some()) && self.config.from_uqff.is_none() {
757            let imatrix_source = match (
758                self.config.imatrix.as_ref(),
759                self.config.calibration_file.is_some(),
760            ) {
761                (None, false) => None,
762                (Some(file), false) => Some(ImatrixDataSource::File(file)),
763                (None, true) => Some(ImatrixDataSource::Collected),
764                (Some(_), true) => unreachable!(),
765            };
766
767            info!("Applying ISQ to all ranks.");
768
769            let multi_progress = Arc::new(MultiProgress::new());
770
771            model.quantize(
772                in_situ_quant,
773                model.device().clone(),
774                self.config.topology.as_ref(),
775                silent,
776                imatrix_source,
777                self.config.organization,
778                self.config.write_uqff.as_ref(),
779                UqffFullSer {
780                    tokenizer: &tokenizer,
781                    template_filename: paths.get_template_filename(),
782                    generation_config: paths.get_gen_conf_filename(),
783                    config: config.clone(),
784                    processor_filename: &None,
785                    preprocessor_filename: &None,
786                },
787                multi_progress.clone(),
788            )?;
789        } else if let Some(from_uqff) = &*self.from_uqff.read().unwrap() {
790            model.load_from_artifacts(
791                device.clone(),
792                self.config.topology.as_ref(),
793                silent,
794                from_uqff,
795            )?;
796        }
797
798        let paged_attn_config = if matches!(
799            self.kind,
800            ModelKind::Adapter {
801                adapter: AdapterKind::XLora
802            }
803        ) {
804            warn!(
805                "Adapter parallel_models do not currently support PagedAttention, running without"
806            );
807            None
808        } else {
809            paged_attn_config
810        };
811
812        let (cache_config, cache_engine) = if let Some(paged_attn_config) = paged_attn_config {
813            let cache_config = calculate_cache_config(
814                paged_attn_config.mem_gpu,
815                paged_attn_config.mem_cpu,
816                paged_attn_config.block_size,
817                dtype,
818                paged_attn_config.cache_type,
819                model.config(),
820                &device,
821                &pipeline_mapper
822                    .get_unique_devices()
823                    .into_iter()
824                    .map(Some)
825                    .collect::<Vec<_>>(),
826                silent,
827            )?;
828
829            let mut layer_devices = Vec::new();
830            for layer in 0..self.inner.num_layers(&config)? {
831                let device = model.get_layers().1.device_for(layer, false).cloned();
832                layer_devices.push(device);
833            }
834            let cache_engine = CacheEngine::new(
835                model.config(),
836                &cache_config,
837                dtype,
838                model.device(),
839                layer_devices.clone(),
840            )?;
841
842            (Some(cache_config), Some(cache_engine))
843        } else {
844            (None, None)
845        };
846
847        let max_seq_len = model.max_seq_len();
848        let llg_factory = build_llg_factory(tokenizer.clone())?;
849        let num_hidden_layers = match model.cache() {
850            EitherCache::Full(full) => full.lock().len(),
851            EitherCache::Normal(normal) => normal.lock().unwrap().0.len(),
852        };
853        let eos = calculate_eos_tokens(&chat_template, gen_conf, &tokenizer);
854        let sliding_window = model.config().sliding_window;
855        let model_metadata = Arc::new(model.config().clone());
856
857        Ok(Arc::new(Mutex::new(NormalPipeline {
858            model,
859            tokenizer: tokenizer.into(),
860            no_kv_cache: self.no_kv_cache,
861            chat_template: Arc::new(chat_template),
862            non_granular_state: self.tgt_non_granular_index.map(|tgt_non_granular_index| {
863                NonGranularState {
864                    non_granular_index: Arc::new(Mutex::new(0)),
865                    tgt_non_granular_index,
866                }
867            }),
868            model_id: self.model_id.clone(),
869            metadata: Arc::new(GeneralMetadata {
870                max_seq_len,
871                llg_factory: Some(llg_factory),
872                no_kv_cache: self.no_kv_cache,
873                no_prefix_cache: is_xlora,
874                num_hidden_layers,
875                eos_tok: eos,
876                kind: self.kind.clone(),
877                is_xlora,
878                activation_dtype: dtype,
879                sliding_window,
880                cache_config,
881                cache_engine,
882                prompt_chunksize: Some(NonZero::new(prompt_chunksize).unwrap()),
883                model_metadata: Some(model_metadata),
884                modalities: Modalities {
885                    input: vec![SupportedModality::Text],
886                    output: vec![SupportedModality::Text],
887                },
888            }),
889            topology: self.config.topology.clone(),
890            silent,
891            organization: self.config.organization,
892            template_filename: paths.get_template_filename().clone(),
893            generation_config: paths.get_gen_conf_filename().cloned(),
894            config,
895            imatrix: self.config.imatrix.clone(),
896            mapper: pipeline_mapper,
897        })))
898    }
899
900    fn get_id(&self) -> String {
901        self.model_id.clone()
902    }
903
904    fn get_kind(&self) -> ModelKind {
905        self.kind.clone()
906    }
907}
908
909impl PreProcessingMixin for NormalPipeline {
910    fn get_chat_template(&self) -> Option<Arc<ChatTemplate>> {
911        Some(self.chat_template.clone())
912    }
913    fn get_input_processor_config(&self) -> Option<Arc<dyn Any>> {
914        None
915    }
916}
917
918impl IsqPipelineMixin for NormalPipeline {
919    fn re_isq_model(&mut self, dtype: IsqType) -> Result<()> {
920        let device = self.device().clone();
921        let multi_progress = Arc::new(MultiProgress::new());
922        self.model.quantize(
923            Some(dtype),
924            device.clone(),
925            self.topology.as_ref(),
926            self.silent,
927            self.imatrix.as_ref().map(ImatrixDataSource::File),
928            self.organization,
929            None,
930            UqffFullSer {
931                tokenizer: &self.tokenizer,
932                template_filename: &self.template_filename,
933                generation_config: self.generation_config.as_ref(),
934                config: self.config.clone(),
935                processor_filename: &None,
936                preprocessor_filename: &None,
937            },
938            multi_progress.clone(),
939        )?;
940        Ok(())
941    }
942}
943
944impl CacheManagerMixin for NormalPipeline {
945    fn clone_in_cache(&self, seqs: &mut [&mut Sequence]) {
946        if matches!(self.model.cache(), EitherCache::Full(_)) {
947            FullCacheManager.clone_in_cache(self, seqs, false)
948        } else {
949            NormalCacheManager.clone_in_cache(self, seqs, false)
950        }
951    }
952    fn clone_out_cache(&self, seqs: &mut [&mut Sequence]) {
953        if matches!(self.model.cache(), EitherCache::Full(_)) {
954            FullCacheManager.clone_out_cache(self, seqs, false)
955        } else {
956            NormalCacheManager.clone_out_cache(self, seqs, false)
957        }
958    }
959    fn set_none_cache(
960        &self,
961        seqs: &mut [&mut Sequence],
962        reset_non_granular: bool,
963        modify_draft_cache: bool,
964        load_preallocated_cache: bool,
965    ) {
966        if matches!(self.model.cache(), EitherCache::Full(_)) {
967            FullCacheManager.set_none_cache(self, seqs, modify_draft_cache, false);
968        } else {
969            NormalCacheManager.set_none_cache(
970                self,
971                seqs,
972                modify_draft_cache,
973                load_preallocated_cache,
974            );
975        }
976        if reset_non_granular {
977            self.reset_non_granular_state()
978        }
979    }
980    fn cache(&self) -> &EitherCache {
981        self.model.cache()
982    }
983}
984
985impl MetadataMixin for NormalPipeline {
986    fn device(&self) -> Device {
987        self.model.device().clone()
988    }
989    fn tokenizer(&self) -> Option<Arc<Tokenizer>> {
990        Some(self.tokenizer.clone())
991    }
992    fn name(&self) -> String {
993        self.model_id.clone()
994    }
995    fn reset_non_granular_state(&self) {
996        if let Some(s) = self.non_granular_state.as_ref() {
997            *self.cache().full().get_scalings_cache() = None;
998            *get_mut_arcmutex!(s.non_granular_index) = 0;
999        }
1000    }
1001    fn get_metadata(&self) -> Arc<GeneralMetadata> {
1002        self.metadata.clone()
1003    }
1004    fn device_mapper(&self) -> Option<&dyn DeviceMapper> {
1005        Some(&*self.mapper)
1006    }
1007}
1008
1009#[async_trait::async_trait]
1010impl Pipeline for NormalPipeline {
1011    fn forward_inputs(
1012        &mut self,
1013        inputs: Box<dyn Any>,
1014        return_raw_logits: bool,
1015    ) -> Result<ForwardInputsResult, candle_core::Error> {
1016        let ModelInputs {
1017            input_ids,
1018            input_ids_full,
1019            seqlen_offsets,
1020            seqlen_offsets_full,
1021            context_lens,
1022            position_ids,
1023            paged_attn_meta,
1024            flash_meta,
1025            flash_meta_full,
1026        } = *inputs.downcast().expect("Downcast failed.");
1027        let metadata = self.get_metadata();
1028        let paged_attn_meta = match (&metadata.cache_engine, &paged_attn_meta) {
1029            (Some(cache_engine), Some(meta)) => Some((cache_engine, meta)),
1030            (Some(_), None) => {
1031                // This can happen if Rust-side user code is wrong
1032                candle_core::bail!("Forward step expected a PagedAttention input metadata. This was not provided, please ensure that the scheduler config is correctly configured for PagedAttention.")
1033            }
1034            (None, Some(_)) => {
1035                // This should never happen but we handle it anyway
1036                candle_core::bail!("Forward step got a PagedAttention input metadata but there is no cache engine. Please raise an issue.")
1037            }
1038            (None, None) => None,
1039        };
1040        let logits = match self.model.is_xlora() {
1041            false => {
1042                let paged_attn_meta = paged_attn_meta
1043                    .as_ref()
1044                    .map(|meta| (meta.0.get_kv_cache().clone(), meta.1.clone()));
1045
1046                self.model.forward(
1047                    &input_ids,
1048                    &seqlen_offsets,
1049                    context_lens,
1050                    position_ids,
1051                    paged_attn_meta.as_ref().map(|(a, b)| (a.clone(), b)),
1052                    &flash_meta,
1053                )?
1054            }
1055            true => self.model.xlora_forward(
1056                &input_ids,
1057                input_ids_full.as_ref().unwrap_or(&input_ids),
1058                &seqlen_offsets,
1059                seqlen_offsets_full.as_ref().unwrap_or(&seqlen_offsets),
1060                self.no_kv_cache,
1061                &self.non_granular_state,
1062                context_lens,
1063                position_ids,
1064                &flash_meta,
1065                flash_meta_full.as_ref().unwrap_or(&flash_meta),
1066            )?,
1067        };
1068        if return_raw_logits {
1069            Ok(ForwardInputsResult::RawLogits { logits })
1070        } else {
1071            Ok(ForwardInputsResult::CausalGeneration { logits })
1072        }
1073    }
1074    async fn sample_causal_gen(
1075        &self,
1076        seqs: &mut [&mut Sequence],
1077        logits: Vec<Tensor>,
1078        prefix_cacher: &mut PrefixCacheManagerV2,
1079        disable_eos_stop: bool,
1080        rng: Arc<std::sync::Mutex<Isaac64Rng>>,
1081    ) -> Result<(), candle_core::Error> {
1082        sample_and_add_toks(self, seqs, logits, prefix_cacher, disable_eos_stop, rng).await
1083    }
1084    fn category(&self) -> ModelCategory {
1085        ModelCategory::Text
1086    }
1087}
1088
1089impl AnyMoePipelineMixin for NormalPipeline {
1090    fn amoe_finish_training(&mut self, gate_model_id: Option<String>) -> candle_core::Result<()> {
1091        self.model.finish_training(gate_model_id)
1092    }
1093    fn amoe_layer_vars(&self) -> Vec<Vec<Var>> {
1094        self.model.get_vars()
1095    }
1096    fn amoe_base_model_trainable_params(&self) -> usize {
1097        self.model.trainable_params()
1098    }
1099    fn amoe_take_cached_gating_outputs(&mut self) -> Vec<Tensor> {
1100        self.model.take_cached_gating_outputs()
1101    }
1102    fn amoe_create_layers(
1103        &mut self,
1104        model_ids: Vec<String>,
1105        token: &TokenSource,
1106        revision: Option<String>,
1107        match_regex: &str,
1108        config: crate::amoe::AnyMoeConfig,
1109        dtype: candle_core::DType,
1110        dev: &Device,
1111        (prefix, mlp): (String, String),
1112        layers: Vec<usize>,
1113        expert_type: AnyMoeExpertType,
1114        silent: bool,
1115        gate_model_id: Option<String>,
1116    ) -> candle_core::Result<()> {
1117        let mut vbs = Vec::new();
1118        // Precompile regex here
1119        let regex = Regex::new(match_regex).map_err(candle_core::Error::msg)?;
1120        for model_id in model_ids {
1121            let model_id_str = &model_id;
1122            let model_id = Path::new(&model_id);
1123
1124            let api = {
1125                let cache = GLOBAL_HF_CACHE.get().cloned().unwrap_or_default();
1126                let mut api = ApiBuilder::from_cache(cache)
1127                    .with_progress(!silent)
1128                    .with_token(get_token(token).map_err(candle_core::Error::msg)?);
1129                if let Ok(x) = std::env::var("HF_HUB_CACHE") {
1130                    api = api.with_cache_dir(x.into());
1131                }
1132                api.build().map_err(candle_core::Error::msg)?
1133            };
1134            let revision = revision.clone().unwrap_or("main".to_string());
1135            let api = api.repo(Repo::with_revision(
1136                model_id_str.clone(),
1137                RepoType::Model,
1138                revision.clone(),
1139            ));
1140
1141            let mut filenames = vec![];
1142            for rfilename in
1143                api_dir_list!(api, model_id, true).filter(|x| x.ends_with(".safetensors"))
1144            {
1145                filenames.push(api_get_file!(api, &rfilename, model_id));
1146            }
1147
1148            let regex = regex.clone();
1149            let match_regex_clone = match_regex.to_string();
1150            let layers_clone = layers.clone();
1151            let vb = from_mmaped_safetensors(
1152                filenames,
1153                vec![],
1154                Some(dtype),
1155                dev,
1156                vec![None],
1157                silent,
1158                None,
1159                move |key| {
1160                    if regex.is_match(&key) {
1161                        // Idx of the last char of the layer id, +1
1162                        // Assumes N.MLP
1163                        let last_layer_idx = key.find(&match_regex_clone).unwrap() - 1;
1164                        let first_layer_idx = key[..last_layer_idx].rfind('.').unwrap();
1165                        let layer_n = key[first_layer_idx + 1..last_layer_idx]
1166                            .parse::<usize>()
1167                            .unwrap();
1168                        layers_clone.contains(&layer_n) || layers_clone.is_empty()
1169                    } else {
1170                        false
1171                    }
1172                },
1173                Arc::new(|_| DeviceForLoadTensor::Base),
1174            )?;
1175            vbs.push(vb);
1176        }
1177
1178        let gate_vb = if let Some(gate_model_id) = gate_model_id {
1179            let model_id_str = &gate_model_id;
1180            let model_id = Path::new(&gate_model_id);
1181
1182            let api = {
1183                let cache = GLOBAL_HF_CACHE.get().cloned().unwrap_or_default();
1184                let mut api = ApiBuilder::from_cache(cache)
1185                    .with_progress(!silent)
1186                    .with_token(get_token(token).map_err(candle_core::Error::msg)?);
1187                if let Ok(x) = std::env::var("HF_HUB_CACHE") {
1188                    api = api.with_cache_dir(x.into());
1189                }
1190                api.build().map_err(candle_core::Error::msg)?
1191            };
1192            let revision = revision.clone().unwrap_or("main".to_string());
1193            let api = api.repo(Repo::with_revision(
1194                model_id_str.clone(),
1195                RepoType::Model,
1196                revision.clone(),
1197            ));
1198
1199            let mut gate_filenames = vec![];
1200            for rfilename in
1201                api_dir_list!(api, model_id, true).filter(|x| x.ends_with(".safetensors"))
1202            {
1203                gate_filenames.push(api_get_file!(api, &rfilename, model_id));
1204            }
1205            assert_eq!(
1206                gate_filenames.len(),
1207                1,
1208                "Gate model ID must contain only one .safetensors file"
1209            );
1210
1211            let vb = from_mmaped_safetensors(
1212                gate_filenames.clone(),
1213                vec![],
1214                Some(dtype),
1215                dev,
1216                vec![None],
1217                silent,
1218                None,
1219                |_| true,
1220                Arc::new(|_| DeviceForLoadTensor::Base),
1221            )?;
1222            info!(
1223                "Loaded gating layers from `{}`",
1224                gate_filenames[0].display()
1225            );
1226            Some(vb)
1227        } else {
1228            None
1229        };
1230
1231        self.model.create_anymoe_layers(
1232            vbs.clone(),
1233            config.clone(),
1234            (prefix.clone(), mlp.clone()),
1235            layers.clone(),
1236            expert_type.clone(),
1237            gate_vb.clone(),
1238        )?;
1239
1240        Ok(())
1241    }
1242    fn amoe_supported(&self) -> bool {
1243        self.model.amoe_supported()
1244    }
1245}
mistralrs_core/pipeline/normal.rs

mistralrs_core/pipeline/
normal.rs
