mistralrs_core/pipeline/
normal.rs

1use super::isq::ImatrixDataSource;
2use super::llg::build_llg_factory;
3use super::{
4    get_model_paths, get_xlora_paths, text_models_inputs_processor::ModelInputs, AdapterKind,
5    CacheManager, GeneralMetadata, Loader, ModelKind, ModelPaths, NormalModel, NormalModelLoader,
6    TokenSource,
7};
8use super::{
9    AnyMoePipelineMixin, CacheManagerMixin, EitherCache, ForwardInputsResult, IsqOrganization,
10    IsqPipelineMixin, MetadataMixin, ModelCategory, PreProcessingMixin,
11};
12use super::{
13    AutoNormalLoader, DeepSeekV2Loader, DeepSeekV3Loader, GLM4Loader, GLM4MoeLiteLoader,
14    GLM4MoeLoader, Gemma2Loader, GemmaLoader, GptOssLoader, GraniteMoeHybridLoader, LlamaLoader,
15    MistralLoader, MixtralLoader, NormalLoaderType, Phi2Loader, Phi3Loader, Phi3_5MoELoader,
16    Qwen2Loader, Qwen3Loader, Qwen3MoELoader, Qwen3NextLoader, SmolLm3Loader, Starcoder2Loader,
17};
18use crate::amoe::AnyMoeExpertType;
19use crate::attention::ATTENTION_CHUNK_SIZE;
20use crate::device_map::{self, DeviceMapper};
21use crate::distributed::{self, WorkerTransferData};
22use crate::kv_cache::{FullCacheManager, HybridCacheManager, NormalCacheManager};
23use crate::lora::Ordering;
24use crate::paged_attention::{calculate_cache_config, AttentionImplementation, CacheEngine};
25use crate::pipeline::chat_template::{calculate_eos_tokens, GenerationConfig};
26use crate::pipeline::isq::UqffFullSer;
27use crate::pipeline::loaders::auto_device_map;
28use crate::pipeline::loaders::QuantizationConfigShim;
29use crate::pipeline::sampling::sample_and_add_toks;
30use crate::pipeline::text_models_inputs_processor::make_prompt_chunk;
31use crate::pipeline::{get_chat_template, Modalities, SupportedModality};
32use crate::pipeline::{ChatTemplate, LocalModelPaths};
33use crate::prefix_cacher::PrefixCacheManagerV2;
34use crate::sequence::Sequence;
35use crate::utils::tokenizer::get_tokenizer;
36use crate::utils::varbuilder_utils::DeviceForLoadTensor;
37use crate::utils::{
38    progress::{new_multi_progress, ProgressScopeGuard},
39    tokens::get_token,
40    varbuilder_utils::from_mmaped_safetensors,
41};
42use crate::xlora_models::NonGranularState;
43use crate::{
44    api_dir_list, api_get_file, get_mut_arcmutex, get_paths, get_uqff_paths, lora_model_loader,
45    normal_model_loader, normal_model_loader_sharded, xlora_model_loader, DeviceMapSetting,
46    PagedAttentionConfig, Pipeline, Topology, TryIntoDType, GLOBAL_HF_CACHE,
47};
48use anyhow::Result;
49use candle_core::{Device, Tensor, Var};
50use hf_hub::Cache;
51use hf_hub::{api::sync::ApiBuilder, Repo, RepoType};
52use mistralrs_quant::log::once_log_info;
53use mistralrs_quant::{
54    AfqLayer, GgufMatMul, HqqLayer, ImmediateIsqOverride, IsqType, QuantizedSerdeType,
55};
56use rand_isaac::Isaac64Rng;
57use regex_automata::meta::Regex;
58use std::any::Any;
59use std::borrow::Cow;
60use std::path::{Path, PathBuf};
61use std::str::FromStr;
62use std::sync::{Arc, RwLock};
63use std::time::Instant;
64use std::{env, fs};
65use tokenizers::Tokenizer;
66use tokio::sync::Mutex;
67use tracing::{info, warn};
68
69pub struct NormalPipeline {
70    model: Box<dyn NormalModel + Send + Sync>,
71    tokenizer: Arc<Tokenizer>,
72    no_kv_cache: bool,
73    chat_template: Arc<ChatTemplate>,
74    non_granular_state: Option<NonGranularState>,
75    model_id: String,
76    metadata: Arc<GeneralMetadata>,
77    topology: Option<Topology>,
78    silent: bool,
79    organization: IsqOrganization,
80    // For full UQFF serialization
81    template_filename: Option<PathBuf>,
82    generation_config: Option<PathBuf>,
83    generation_defaults: Option<crate::ModelGenerationDefaults>,
84    config: String,
85    imatrix: Option<PathBuf>,
86    mapper: Box<dyn DeviceMapper + Send + Sync>,
87}
88
89/// A loader for a "normal" (non-quantized) model.
90pub struct NormalLoader {
91    inner: Box<dyn NormalModelLoader>,
92    model_id: String,
93    config: NormalSpecificConfig,
94    xlora_model_id: Option<String>,
95    lora_adapter_ids: Option<Vec<String>>,
96    kind: ModelKind,
97    xlora_order: Option<Ordering>,
98    no_kv_cache: bool,
99    chat_template: Option<String>,
100    tokenizer_json: Option<String>,
101    tgt_non_granular_index: Option<usize>,
102    token_source: RwLock<Option<TokenSource>>,
103    revision: RwLock<Option<String>>,
104    from_uqff: RwLock<Option<Vec<PathBuf>>>,
105    jinja_explicit: Option<String>,
106    hf_cache_path: Option<PathBuf>,
107}
108
109#[derive(Default)]
110/// A builder for a loader for a "normal" (non-quantized) model.
111pub struct NormalLoaderBuilder {
112    model_id: Option<String>,
113    config: NormalSpecificConfig,
114    xlora_model_id: Option<String>,
115    lora_adapter_ids: Option<Vec<String>>,
116    kind: ModelKind,
117    xlora_order: Option<Ordering>,
118    no_kv_cache: bool,
119    chat_template: Option<String>,
120    tokenizer_json: Option<String>,
121    tgt_non_granular_index: Option<usize>,
122    jinja_explicit: Option<String>,
123    hf_cache_path: Option<PathBuf>,
124}
125
126#[derive(Clone, Default)]
127/// Config specific to loading a normal model.
128pub struct NormalSpecificConfig {
129    pub topology: Option<Topology>,
130    pub organization: IsqOrganization,
131    pub write_uqff: Option<PathBuf>,
132    pub from_uqff: Option<Vec<PathBuf>>,
133    pub imatrix: Option<PathBuf>,
134    pub calibration_file: Option<PathBuf>,
135    pub hf_cache_path: Option<PathBuf>,
136    pub matformer_config_path: Option<PathBuf>,
137    pub matformer_slice_name: Option<String>,
138}
139
140impl NormalLoaderBuilder {
141    pub fn new(
142        config: NormalSpecificConfig,
143        chat_template: Option<String>,
144        tokenizer_json: Option<String>,
145        model_id: Option<String>,
146        no_kv_cache: bool,
147        jinja_explicit: Option<String>,
148    ) -> Self {
149        Self {
150            config,
151            chat_template,
152            tokenizer_json,
153            model_id,
154            kind: ModelKind::Normal,
155            jinja_explicit,
156            no_kv_cache,
157            ..Default::default()
158        }
159    }
160
161    fn with_adapter(
162        mut self,
163        xlora_model_id: String,
164        xlora_order: Ordering,
165        no_kv_cache: bool,
166        tgt_non_granular_index: Option<usize>,
167    ) -> Self {
168        self.xlora_model_id = Some(xlora_model_id);
169        self.xlora_order = Some(xlora_order);
170        self.no_kv_cache = no_kv_cache;
171        self.tgt_non_granular_index = tgt_non_granular_index;
172        self.model_id = if let Some(id) = self.model_id {
173            Some(id)
174        } else {
175            info!(
176                "Using adapter base model ID: `{}`",
177                self.xlora_order.as_ref().unwrap().base_model_id
178            );
179            Some(self.xlora_order.as_ref().unwrap().base_model_id.clone())
180        };
181        self
182    }
183
184    pub fn with_xlora(
185        mut self,
186        xlora_model_id: String,
187        xlora_order: Ordering,
188        no_kv_cache: bool,
189        tgt_non_granular_index: Option<usize>,
190    ) -> Self {
191        self.kind = ModelKind::Adapter {
192            adapter: AdapterKind::XLora,
193        };
194        self.with_adapter(
195            xlora_model_id,
196            xlora_order,
197            no_kv_cache,
198            tgt_non_granular_index,
199        )
200    }
201
202    pub fn with_lora(mut self, lora_adapter_ids: Vec<String>) -> Self {
203        self.kind = ModelKind::Adapter {
204            adapter: AdapterKind::Lora,
205        };
206        self.lora_adapter_ids = Some(lora_adapter_ids);
207        self
208    }
209
210    pub fn hf_cache_path(mut self, hf_cache_path: PathBuf) -> Self {
211        self.hf_cache_path = Some(hf_cache_path);
212        self
213    }
214
215    /// If the loader type is not specified, loader type is automatically determined from the
216    /// `architectures` array in the config.
217    pub fn build(self, loader_tp: Option<NormalLoaderType>) -> anyhow::Result<Box<dyn Loader>> {
218        let loader: Box<dyn NormalModelLoader> = match loader_tp {
219            Some(NormalLoaderType::Mistral) => Box::new(MistralLoader),
220            Some(NormalLoaderType::Gemma) => Box::new(GemmaLoader),
221            Some(NormalLoaderType::Llama) => Box::new(LlamaLoader),
222            Some(NormalLoaderType::Mixtral) => Box::new(MixtralLoader),
223            Some(NormalLoaderType::Phi2) => Box::new(Phi2Loader),
224            Some(NormalLoaderType::Phi3) => Box::new(Phi3Loader),
225            Some(NormalLoaderType::Qwen2) => Box::new(Qwen2Loader),
226            Some(NormalLoaderType::Gemma2) => Box::new(Gemma2Loader),
227            Some(NormalLoaderType::Starcoder2) => Box::new(Starcoder2Loader),
228            Some(NormalLoaderType::Phi3_5MoE) => Box::new(Phi3_5MoELoader),
229            Some(NormalLoaderType::DeepSeekV2) => Box::new(DeepSeekV2Loader),
230            Some(NormalLoaderType::DeepSeekV3) => Box::new(DeepSeekV3Loader),
231            Some(NormalLoaderType::Qwen3) => Box::new(Qwen3Loader),
232            Some(NormalLoaderType::GLM4) => Box::new(GLM4Loader),
233            Some(NormalLoaderType::GLM4MoeLite) => Box::new(GLM4MoeLiteLoader),
234            Some(NormalLoaderType::GLM4Moe) => Box::new(GLM4MoeLoader),
235            Some(NormalLoaderType::Qwen3Moe) => Box::new(Qwen3MoELoader),
236            Some(NormalLoaderType::SmolLm3) => Box::new(SmolLm3Loader),
237            Some(NormalLoaderType::GraniteMoeHybrid) => Box::new(GraniteMoeHybridLoader),
238            Some(NormalLoaderType::GptOss) => Box::new(GptOssLoader),
239            Some(NormalLoaderType::Qwen3Next) => Box::new(Qwen3NextLoader),
240            None => Box::new(AutoNormalLoader),
241        };
242        Ok(Box::new(NormalLoader {
243            inner: loader,
244            model_id: self.model_id.unwrap(),
245            config: self.config,
246            xlora_model_id: self.xlora_model_id,
247            lora_adapter_ids: self.lora_adapter_ids,
248            kind: self.kind,
249            xlora_order: self.xlora_order,
250            no_kv_cache: self.no_kv_cache,
251            chat_template: self.chat_template,
252            tokenizer_json: self.tokenizer_json,
253            tgt_non_granular_index: self.tgt_non_granular_index,
254            jinja_explicit: self.jinja_explicit,
255            token_source: RwLock::new(None),
256            revision: RwLock::new(None),
257            from_uqff: RwLock::new(None),
258            hf_cache_path: self.hf_cache_path,
259        }))
260    }
261}
262
263impl Loader for NormalLoader {
264    #[allow(clippy::type_complexity, clippy::too_many_arguments)]
265    fn load_model_from_hf(
266        &self,
267        revision: Option<String>,
268        token_source: TokenSource,
269        dtype: &dyn TryIntoDType,
270        device: &Device,
271        silent: bool,
272        mapper: DeviceMapSetting,
273        in_situ_quant: Option<IsqType>,
274        paged_attn_config: Option<PagedAttentionConfig>,
275    ) -> Result<Arc<Mutex<dyn Pipeline + Send + Sync>>> {
276        let _progress_guard = ProgressScopeGuard::new(silent);
277        let cache = self
278            .hf_cache_path
279            .clone()
280            .map(Cache::new)
281            .unwrap_or_default();
282        GLOBAL_HF_CACHE.get_or_init(|| cache);
283
284        let paths: anyhow::Result<Box<dyn ModelPaths>> = get_paths!(
285            LocalModelPaths,
286            &token_source,
287            revision.clone(),
288            self,
289            None,
290            None,
291            silent,
292            self.config.from_uqff.is_some()
293        );
294        *self
295            .token_source
296            .write()
297            .expect("Failed to write to token source") = Some(token_source);
298        *self.revision.write().expect("Failed to write to revision") = revision.clone();
299        if let Some(from_uqff) = self.config.from_uqff.clone() {
300            *self.from_uqff.write().unwrap() = Some(get_uqff_paths!(&from_uqff, self, silent));
301        }
302        self.load_model_from_path(
303            &paths?,
304            dtype,
305            device,
306            silent,
307            mapper,
308            in_situ_quant,
309            paged_attn_config,
310        )
311    }
312
313    #[allow(clippy::type_complexity, clippy::too_many_arguments)]
314    fn load_model_from_path(
315        &self,
316        paths: &Box<dyn ModelPaths>,
317        dtype: &dyn TryIntoDType,
318        device: &Device,
319        silent: bool,
320        mut mapper: DeviceMapSetting,
321        in_situ_quant: Option<IsqType>,
322        mut paged_attn_config: Option<PagedAttentionConfig>,
323    ) -> Result<Arc<Mutex<dyn Pipeline + Send + Sync>>> {
324        let _progress_guard = ProgressScopeGuard::new(silent);
325        let config = std::fs::read_to_string(paths.get_config_filename())?;
326
327        if !self.inner.supports_paged_attention(&config)? {
328            paged_attn_config = None;
329        }
330
331        info!("Prompt chunk size is {ATTENTION_CHUNK_SIZE}.");
332
333        let use_nccl = mistralrs_quant::distributed::use_nccl();
334
335        let available_devices = if let Ok(payload) = env::var(distributed::IS_DAEMON_FLAG) {
336            let payload: WorkerTransferData = serde_json::from_str(&payload)?;
337            let WorkerTransferData::Init { id: _, worker_rank } = payload;
338            vec![candle_core::Device::new_cuda(worker_rank + 1)?]
339        } else if use_nccl {
340            vec![candle_core::Device::new_cuda(0)?]
341        } else {
342            device_map::get_all_similar_devices(device)?
343        };
344        #[cfg(feature = "cuda")]
345        for device in &available_devices {
346            if let Device::Cuda(dev) = device {
347                unsafe { dev.disable_event_tracking() };
348            }
349        }
350        let device = if use_nccl || cfg!(feature = "ring") {
351            available_devices[0].clone()
352        } else {
353            device.clone()
354        };
355
356        // If auto, convert to Map if not using nccl
357        let mut max_kv_tokens: Option<usize> = None;
358        if use_nccl || cfg!(feature = "ring") {
359            mapper = DeviceMapSetting::DummyNccl {
360                nm_device: available_devices[0].clone(),
361            };
362        } else if let DeviceMapSetting::Auto(params) = mapper.clone() {
363            max_kv_tokens = Some(params.max_seq_len() * params.max_batch_size());
364            // Initial dtype
365            let dtype = dtype.try_into_dtype(&available_devices.iter().collect::<Vec<_>>())?;
366
367            // ISQ or UQFF: quantized path
368            // Match logic below where UQFF has priority
369            let (layer_sizes_in_bytes, non_mapped_size_in_bytes, total_model_size_in_bytes) =
370                if let Some(serialized) = &*self.from_uqff.read().unwrap() {
371                    let weight_pack_factor = {
372                        let ser_artifacts = unsafe {
373                            candle_core::safetensors::MmapedSafetensors::multi(serialized)?
374                        };
375                        let mut total_pack_factors = 0;
376                        let total_tensors = ser_artifacts.tensors().len();
377                        for (_, artifact) in ser_artifacts.tensors() {
378                            let artifact = artifact.data();
379                            // NOTE(EricLBuehler): isq type is ALWAYS byte 4 (5th) of the tensor.
380                            let isq_type = artifact[mistralrs_quant::UQFF_QUANT_TYPE_OFFSET];
381                            let pack_factor = match QuantizedSerdeType::try_from(isq_type as usize)?
382                            {
383                                QuantizedSerdeType::Hqq => {
384                                    HqqLayer::get_isq_type_from_uqff(Cow::Borrowed(artifact))?
385                                        .pack_factor(dtype)
386                                }
387                                QuantizedSerdeType::Gguf => {
388                                    GgufMatMul::get_isq_type_from_uqff(Cow::Borrowed(artifact))?
389                                        .pack_factor(dtype)
390                                }
391                                QuantizedSerdeType::Fp8 => IsqType::F8E4M3.pack_factor(dtype),
392                                QuantizedSerdeType::Unquant => 1,
393                                QuantizedSerdeType::Afq => {
394                                    AfqLayer::get_isq_type_from_uqff(Cow::Borrowed(artifact))?
395                                        .pack_factor(dtype)
396                                }
397                                QuantizedSerdeType::F8Q8 => IsqType::F8Q8.pack_factor(dtype),
398                                QuantizedSerdeType::Mxfp4 => IsqType::MXFP4.pack_factor(dtype),
399                            };
400                            total_pack_factors += pack_factor;
401                        }
402
403                        total_pack_factors / total_tensors
404                    };
405
406                    let layer_sizes_in_bytes = self.inner.layer_sizes_in_bytes(
407                        &config,
408                        dtype,
409                        weight_pack_factor,
410                        None,
411                    )?;
412                    let non_mapped_size_in_bytes = self.inner.non_mapped_size_in_bytes(
413                        &config,
414                        dtype,
415                        weight_pack_factor,
416                        None,
417                    )?;
418                    let layer_sizes_sum = layer_sizes_in_bytes.iter().sum::<usize>();
419                    (
420                        layer_sizes_in_bytes,
421                        non_mapped_size_in_bytes,
422                        layer_sizes_sum + non_mapped_size_in_bytes,
423                    )
424                } else if let Some(isq) = in_situ_quant {
425                    let weight_pack_factor = isq.pack_factor(dtype);
426                    let layer_sizes_in_bytes = self.inner.layer_sizes_in_bytes(
427                        &config,
428                        dtype,
429                        weight_pack_factor,
430                        None,
431                    )?;
432                    let non_mapped_size_in_bytes = self.inner.non_mapped_size_in_bytes(
433                        &config,
434                        dtype,
435                        weight_pack_factor,
436                        None,
437                    )?;
438                    let layer_sizes_sum = layer_sizes_in_bytes.iter().sum::<usize>();
439                    (
440                        layer_sizes_in_bytes,
441                        non_mapped_size_in_bytes,
442                        layer_sizes_sum + non_mapped_size_in_bytes,
443                    )
444                } else {
445                    // Be sure to get the weight pack factor here; we might be loading a prequantized model.
446                    let weight_pack_factor =
447                        QuantizationConfigShim::get_quant_config_pack_factor(&config, dtype)?;
448                    let layer_sizes_in_bytes = self.inner.layer_sizes_in_bytes(
449                        &config,
450                        dtype,
451                        weight_pack_factor,
452                        None,
453                    )?;
454                    let non_mapped_size_in_bytes = self.inner.non_mapped_size_in_bytes(
455                        &config,
456                        dtype,
457                        weight_pack_factor,
458                        None,
459                    )?;
460                    let layer_sizes_sum = layer_sizes_in_bytes.iter().sum::<usize>();
461                    (
462                        layer_sizes_in_bytes,
463                        non_mapped_size_in_bytes,
464                        layer_sizes_sum + non_mapped_size_in_bytes,
465                    )
466                };
467
468            let new = auto_device_map::get_device_layers(
469                &*self.inner,
470                &config,
471                self.inner.num_layers(&config)?,
472                layer_sizes_in_bytes,
473                non_mapped_size_in_bytes,
474                total_model_size_in_bytes,
475                &available_devices,
476                dtype,
477                &params,
478                paged_attn_config.as_ref(),
479            )?;
480            mapper = DeviceMapSetting::Map(new);
481        }
482
483        let pipeline_mapper = mapper.into_mapper(
484            self.inner.num_layers(&config)?,
485            &device,
486            self.config.topology.as_ref(),
487            &available_devices,
488        )?;
489        let mapper = mapper.into_mapper(
490            self.inner.num_layers(&config)?,
491            &device,
492            self.config.topology.as_ref(),
493            &available_devices,
494        )?;
495        let mut layer_devices = Vec::new();
496        for layer in 0..self.inner.num_layers(&config)? {
497            let device = mapper.device_for(layer, false).cloned();
498            layer_devices.push(device);
499        }
500        let dtype = mapper.get_min_dtype(dtype)?;
501
502        // TODO: PagedAttention is not supported with CPU for now.
503        // This check is not really necessary because `get_device_layers` should prevent it.
504        let mapping_uses_cpu = mapper.get_unique_devices().iter().any(Device::is_cpu);
505        if mapping_uses_cpu && paged_attn_config.is_some() {
506            warn!("Device mapping contains a mix of GPU and CPU. There is no CPU support for PagedAttention, disabling PagedAttention.");
507            paged_attn_config = None;
508        }
509
510        info!("Model config: {:?}", self.inner.get_config_repr(&config)?);
511        if crate::using_flash_attn() {
512            once_log_info("FlashAttention is enabled.");
513        }
514
515        let topology_overrides = self
516            .config
517            .topology
518            .as_ref()
519            .map(|topology| {
520                topology
521                    .pattern_overrides()
522                    .into_iter()
523                    .map(|(regex, layer)| ImmediateIsqOverride {
524                        predicate: regex,
525                        ty: layer.isq,
526                        device: layer.device.clone(),
527                    })
528                    .collect::<Vec<_>>()
529            })
530            .unwrap_or_default();
531        let has_override_isq = topology_overrides
532            .iter()
533            .any(|override_entry| override_entry.ty.is_some());
534        let topology_requires_post_quant = self
535            .config
536            .topology
537            .as_ref()
538            .is_some_and(|topology| topology.requires_post_quantization());
539
540        let allow_immediate_cli = self.config.imatrix.is_none()
541            && self.config.calibration_file.is_none()
542            && in_situ_quant.is_some();
543
544        let mut immediate_ty = None;
545        let mut immediate_predicates = Vec::new();
546        if allow_immediate_cli {
547            immediate_ty = in_situ_quant;
548            immediate_predicates =
549                if matches!(self.config.organization, IsqOrganization::MoeExpertsOnly) {
550                    self.inner.immediate_isq_predicates_moqe(&config)?
551                } else {
552                    self.inner.immediate_isq_predicates(&config)?
553                };
554            info!("Applying ISQ to {in_situ_quant:?}");
555            if immediate_predicates.is_empty() {
556                warn!("No predicates for this model and ISQ setting detected. ISQ will not be applied to any weights!");
557            }
558        }
559
560        let use_immediate = allow_immediate_cli || has_override_isq;
561        if use_immediate {
562            let (pool, num_threads) = mistralrs_quant::create_isq_thread_pool(immediate_ty);
563            info!("Applying immediate ISQ in parallel on {num_threads} threads.");
564            mistralrs_quant::set_immediate_isq_with_pool(
565                immediate_ty,
566                immediate_predicates.clone(),
567                topology_overrides.clone(),
568                pool,
569            );
570        }
571
572        // Logic for ISQ here: if no calibration (i.e imatrix), then allow immediate ISQ. Otherwise, back to normal.
573        let mut loading_isq = if use_immediate {
574            false
575        } else {
576            in_situ_quant.is_some()
577        };
578        if self.config.imatrix.is_some() || self.config.calibration_file.is_some() {
579            loading_isq = true;
580        }
581        loading_isq |= topology_requires_post_quant;
582        loading_isq |= self.config.from_uqff.is_some();
583
584        if self.config.imatrix.is_some() && self.config.calibration_file.is_some() {
585            anyhow::bail!(
586                "`imatrix` and `calibration_file` were both specified, this is not allowed."
587            );
588        }
589
590        // Load onto the regular device if not using isq or if the calibration file is specified.
591        // For immediate ISQ on discrete GPUs, load to CPU: the mapper will set the correct target
592        // device per-layer, and linear constructors will override to CPU for ISQ-targeted weights.
593        // On integrated/unified memory systems (e.g. Grace Blackwell), CPU and GPU share memory,
594        // so we load directly to the device.
595        let load_device = if !loading_isq || self.config.calibration_file.is_some() {
596            loading_isq = false;
597            if use_immediate && !crate::utils::normal::is_integrated_gpu(&device) {
598                Device::Cpu
599            } else {
600                device.clone()
601            }
602        } else {
603            Device::Cpu
604        };
605
606        let is_xlora = self.kind.is_adapted_and(|a| a.is_x_lora());
607
608        let attention_mechanism = if paged_attn_config.is_some() {
609            AttentionImplementation::PagedAttention
610        } else {
611            AttentionImplementation::Eager
612        };
613
614        let multi_progress = Arc::new(new_multi_progress());
615
616        // Load matformer slicing config if provided
617        let matformer_slicing_config = if let Some(matformer_path) =
618            &self.config.matformer_config_path
619        {
620            use crate::matformer::{MatformerConfig, MatformerSliceConfig};
621            info!("Loading Matformer config from {:?}", matformer_path);
622            let config = Arc::new(MatformerConfig::from_file(matformer_path)?);
623
624            if let Some(slice_name) = &self.config.matformer_slice_name {
625                info!("Using Matformer slice: {}", slice_name);
626                Some(MatformerSliceConfig::new(slice_name.clone(), config))
627            } else {
628                // If no slice name is provided but config exists, we'll need to handle this
629                // For now, return None and let the model handle the default slice selection
630                warn!("Matformer config loaded but no slice name specified. Models will use their default slice.");
631                None
632            }
633        } else {
634            None
635        };
636
637        let mut model = if use_nccl || cfg!(feature = "ring") {
638            let (mapper, sharded_vb) = distributed::prepare_distributed_mapper(
639                dtype,
640                &device,
641                &available_devices,
642                silent,
643                &config,
644                loading_isq,
645                self.config.from_uqff.is_some(),
646                self.config.organization,
647                &*self.inner,
648                paths.as_ref(),
649            )?;
650
651            // Special case for where things can be more optimially loaded.
652            match self.kind {
653                ModelKind::Normal => normal_model_loader_sharded!(
654                    sharded_vb,
655                    config,
656                    self.inner,
657                    mapper,
658                    loading_isq,
659                    device.clone(),
660                    attention_mechanism,
661                    multi_progress.clone(),
662                    matformer_slicing_config.clone(),
663                ),
664                ModelKind::Adapter {
665                    adapter: AdapterKind::XLora,
666                } => xlora_model_loader!(
667                    paths,
668                    Some(dtype),
669                    &load_device,
670                    layer_devices.clone(),
671                    config,
672                    self.inner,
673                    silent,
674                    mapper,
675                    loading_isq,
676                    device.clone(),
677                    multi_progress.clone(),
678                    matformer_slicing_config.clone(),
679                ),
680                ModelKind::Adapter {
681                    adapter: AdapterKind::Lora,
682                } => lora_model_loader!(
683                    paths,
684                    Some(dtype),
685                    &load_device,
686                    layer_devices.clone(),
687                    config,
688                    self.inner,
689                    silent,
690                    mapper,
691                    loading_isq,
692                    self.config.from_uqff.is_some(),
693                    device.clone(),
694                    attention_mechanism,
695                    matches!(self.config.organization, IsqOrganization::MoeExpertsOnly),
696                    multi_progress.clone(),
697                    matformer_slicing_config.clone(),
698                ),
699                _ => unreachable!(),
700            }
701        } else {
702            match self.kind {
703                ModelKind::Normal => normal_model_loader!(
704                    paths,
705                    Some(dtype),
706                    &load_device,
707                    layer_devices.clone(),
708                    config,
709                    self.inner,
710                    silent,
711                    mapper,
712                    loading_isq,
713                    self.config.from_uqff.is_some(),
714                    device.clone(),
715                    attention_mechanism,
716                    matches!(self.config.organization, IsqOrganization::MoeExpertsOnly),
717                    multi_progress.clone(),
718                    matformer_slicing_config.clone(),
719                ),
720                ModelKind::Adapter {
721                    adapter: AdapterKind::XLora,
722                } => xlora_model_loader!(
723                    paths,
724                    Some(dtype),
725                    &load_device,
726                    layer_devices.clone(),
727                    config,
728                    self.inner,
729                    silent,
730                    mapper,
731                    loading_isq,
732                    device.clone(),
733                    multi_progress.clone(),
734                    matformer_slicing_config.clone(),
735                ),
736                ModelKind::Adapter {
737                    adapter: AdapterKind::Lora,
738                } => lora_model_loader!(
739                    paths,
740                    Some(dtype),
741                    &load_device,
742                    layer_devices.clone(),
743                    config,
744                    self.inner,
745                    silent,
746                    mapper,
747                    loading_isq,
748                    self.config.from_uqff.is_some(),
749                    device.clone(),
750                    attention_mechanism,
751                    matches!(self.config.organization, IsqOrganization::MoeExpertsOnly),
752                    multi_progress.clone(),
753                    matformer_slicing_config.clone(),
754                ),
755                _ => unreachable!(),
756            }
757        };
758
759        let tokenizer = get_tokenizer(paths.get_tokenizer_filename(), None)?;
760        let gen_conf: Option<GenerationConfig> = paths.get_gen_conf_filename().and_then(|f| {
761            match serde_json::from_str::<GenerationConfig>(&fs::read_to_string(f).unwrap()) {
762                Ok(conf) => Some(conf),
763                Err(e) => {
764                    warn!("Failed to parse generation_config.json: {}", e);
765                    None
766                }
767            }
768        });
769
770        let chat_template_explicit = paths
771            .get_chat_template_explicit()
772            .as_ref()
773            .map(|x| x.to_string_lossy().to_string());
774        let chat_template = get_chat_template(
775            paths,
776            self.jinja_explicit.as_ref(),
777            chat_template_explicit.as_ref(),
778            self.chat_template.as_ref(),
779            None,
780        );
781
782        if let Some(calibration_file) = &self.config.calibration_file {
783            let calibration_data = std::fs::read_to_string(calibration_file)?;
784            // Tokenize, don't add bos yet
785            let tokens = tokenizer
786                .encode_fast(calibration_data, false)
787                .map_err(anyhow::Error::msg)?
788                .get_ids()
789                .to_vec();
790            info!(
791                "Collecting imatrix from calibration file `{}` of {} tokens.",
792                calibration_file.display(),
793                tokens.len()
794            );
795            let bos_tok_id = chat_template
796                .bos_tok()
797                .as_deref()
798                .and_then(|tok| tokenizer.token_to_id(tok));
799
800            match self.config.organization {
801                IsqOrganization::Default => model.begin_track_stats()?,
802                IsqOrganization::MoeExpertsOnly => model.begin_track_stats_moe_experts_only()?,
803            }
804
805            const CHUNK_SIZE: usize = 1024;
806            let n_chunks = tokens.len().div_ceil(CHUNK_SIZE);
807            let start = Instant::now();
808            for (i, chunk) in tokens.chunks(CHUNK_SIZE).enumerate() {
809                let mut chunk = chunk.to_vec();
810                if let Some(bos_tok_id) = bos_tok_id {
811                    chunk.insert(0, bos_tok_id);
812                }
813                let chunk_len = chunk.len();
814
815                let start = Instant::now();
816                let inputs = make_prompt_chunk(
817                    0,
818                    vec![&chunk],
819                    &[0],
820                    &load_device,
821                    None,
822                    false,
823                    None,
824                    Some(pipeline_mapper.as_ref()),
825                    None,
826                    model.config().sliding_window,
827                )?;
828
829                model.forward(
830                    &inputs.input.to_device(model.device())?,
831                    &inputs.positions,
832                    inputs.context_lens.clone(),
833                    inputs.position_ids.clone(),
834                    None,
835                    &inputs.flash_meta.clone(),
836                )?;
837
838                match model.cache_mut() {
839                    EitherCache::Full(full) => {
840                        for layer in &mut *full.lock() {
841                            *layer = None
842                        }
843                    }
844                    EitherCache::Normal(normal) => {
845                        for layer in &mut *normal.lock().unwrap().0 {
846                            layer.reset();
847                        }
848                    }
849                    EitherCache::Hybrid(hybrid) => {
850                        hybrid.lock().unwrap().reset();
851                    }
852                }
853
854                let end = Instant::now();
855                info!(
856                    "Processed chunk {}/{n_chunks} ({chunk_len} tokens), {:.2}s",
857                    i + 1,
858                    end.duration_since(start).as_secs_f32()
859                );
860            }
861            load_device.synchronize()?;
862            let end = Instant::now();
863            info!(
864                "Finished collecting imatrix in {:.2}s",
865                end.duration_since(start).as_secs_f32()
866            );
867        }
868
869        // Only if loading from UQFF
870        let should_serialize = self.config.write_uqff.is_some();
871        let should_quantize_pass = loading_isq;
872
873        if (should_quantize_pass || should_serialize) && self.config.from_uqff.is_none() {
874            let imatrix_source = if should_quantize_pass {
875                match (
876                    self.config.imatrix.as_ref(),
877                    self.config.calibration_file.is_some(),
878                ) {
879                    (None, false) => None,
880                    (Some(file), false) => Some(ImatrixDataSource::File(file)),
881                    (None, true) => Some(ImatrixDataSource::Collected),
882                    (Some(_), true) => unreachable!(),
883                }
884            } else {
885                None
886            };
887
888            if should_quantize_pass {
889                info!("Applying ISQ to all ranks.");
890            } else {
891                info!("Serializing existing ISQ tensors without additional quantization.");
892            }
893
894            let multi_progress = Arc::new(new_multi_progress());
895
896            model.quantize(
897                in_situ_quant,
898                model.device().clone(),
899                self.config.topology.as_ref(),
900                silent,
901                imatrix_source,
902                self.config.organization,
903                should_quantize_pass,
904                self.config.write_uqff.as_ref(),
905                UqffFullSer {
906                    tokenizer: &tokenizer,
907                    template_filename: paths.get_template_filename(),
908                    generation_config: paths.get_gen_conf_filename(),
909                    config: config.clone(),
910                    processor_filename: &None,
911                    preprocessor_filename: &None,
912                    modules: None,
913                    module_paths: None,
914                },
915                multi_progress.clone(),
916            )?;
917        } else if let Some(from_uqff) = &*self.from_uqff.read().unwrap() {
918            model.load_from_artifacts(
919                device.clone(),
920                self.config.topology.as_ref(),
921                silent,
922                from_uqff,
923            )?;
924        }
925
926        let paged_attn_config = if matches!(
927            self.kind,
928            ModelKind::Adapter {
929                adapter: AdapterKind::XLora
930            }
931        ) {
932            warn!(
933                "Adapter parallel_models do not currently support PagedAttention, running without"
934            );
935            None
936        } else {
937            paged_attn_config
938        };
939
940        let model_metadata = model.model_config();
941        let (cache_config, cache_engine) = if let Some(paged_attn_config) = paged_attn_config {
942            let cache_config = calculate_cache_config(
943                paged_attn_config.mem_gpu,
944                paged_attn_config.block_size,
945                dtype,
946                paged_attn_config.cache_type,
947                model_metadata.as_ref(),
948                &device,
949                &pipeline_mapper
950                    .get_unique_devices()
951                    .into_iter()
952                    .map(Some)
953                    .collect::<Vec<_>>(),
954                silent,
955                None,
956                max_kv_tokens,
957            )?;
958
959            let mut layer_devices = Vec::new();
960            for layer in 0..self.inner.num_layers(&config)? {
961                let device = model.get_layers().1.device_for(layer, false).cloned();
962                layer_devices.push(device);
963            }
964            let cache_engine = CacheEngine::new(
965                model_metadata.as_ref(),
966                &cache_config,
967                dtype,
968                model.device(),
969                layer_devices.clone(),
970            )?;
971
972            (Some(cache_config), Some(cache_engine))
973        } else {
974            (None, None)
975        };
976
977        let max_seq_len = model.max_seq_len();
978        let llg_factory = build_llg_factory(tokenizer.clone())?;
979        let num_hidden_layers = match model.cache() {
980            EitherCache::Full(full) => full.lock().len(),
981            EitherCache::Normal(normal) => normal.lock().unwrap().0.len(),
982            EitherCache::Hybrid(hybrid) => hybrid.lock().unwrap().num_layers(),
983        };
984        let generation_defaults = gen_conf
985            .as_ref()
986            .and_then(GenerationConfig::generation_defaults);
987        let eos = calculate_eos_tokens(&chat_template, gen_conf.as_ref(), &tokenizer);
988        let sliding_window = model.config().sliding_window;
989        Ok(Arc::new(Mutex::new(NormalPipeline {
990            model,
991            tokenizer: tokenizer.into(),
992            no_kv_cache: self.no_kv_cache,
993            chat_template: Arc::new(chat_template),
994            non_granular_state: self.tgt_non_granular_index.map(|tgt_non_granular_index| {
995                NonGranularState {
996                    non_granular_index: Arc::new(Mutex::new(0)),
997                    tgt_non_granular_index,
998                }
999            }),
1000            model_id: self.model_id.clone(),
1001            metadata: Arc::new(GeneralMetadata {
1002                max_seq_len,
1003                llg_factory: Some(llg_factory),
1004                no_kv_cache: self.no_kv_cache,
1005                no_prefix_cache: is_xlora,
1006                num_hidden_layers,
1007                eos_tok: eos,
1008                kind: self.kind.clone(),
1009                is_xlora,
1010                activation_dtype: dtype,
1011                sliding_window,
1012                cache_config,
1013                cache_engine,
1014                model_metadata: Some(model_metadata),
1015                modalities: Modalities {
1016                    input: vec![SupportedModality::Text],
1017                    output: vec![SupportedModality::Text],
1018                },
1019            }),
1020            topology: self.config.topology.clone(),
1021            silent,
1022            organization: self.config.organization,
1023            template_filename: paths.get_template_filename().clone(),
1024            generation_config: paths.get_gen_conf_filename().cloned(),
1025            generation_defaults,
1026            config,
1027            imatrix: self.config.imatrix.clone(),
1028            mapper: pipeline_mapper,
1029        })))
1030    }
1031
1032    fn get_id(&self) -> String {
1033        self.model_id.clone()
1034    }
1035
1036    fn get_kind(&self) -> ModelKind {
1037        self.kind.clone()
1038    }
1039}
1040
1041impl PreProcessingMixin for NormalPipeline {
1042    fn get_chat_template(&self) -> Option<Arc<ChatTemplate>> {
1043        Some(self.chat_template.clone())
1044    }
1045    fn get_input_processor_config(&self) -> Option<Arc<dyn Any>> {
1046        None
1047    }
1048}
1049
1050impl IsqPipelineMixin for NormalPipeline {
1051    fn re_isq_model(&mut self, dtype: IsqType) -> Result<()> {
1052        let device = self.device().clone();
1053        let multi_progress = Arc::new(new_multi_progress());
1054        self.model.quantize(
1055            Some(dtype),
1056            device.clone(),
1057            self.topology.as_ref(),
1058            self.silent,
1059            self.imatrix.as_ref().map(ImatrixDataSource::File),
1060            self.organization,
1061            true,
1062            None,
1063            UqffFullSer {
1064                tokenizer: &self.tokenizer,
1065                template_filename: &self.template_filename,
1066                generation_config: self.generation_config.as_ref(),
1067                config: self.config.clone(),
1068                processor_filename: &None,
1069                preprocessor_filename: &None,
1070                modules: None,
1071                module_paths: None,
1072            },
1073            multi_progress.clone(),
1074        )?;
1075        Ok(())
1076    }
1077}
1078
1079impl CacheManagerMixin for NormalPipeline {
1080    fn clone_in_cache(&self, seqs: &mut [&mut Sequence]) {
1081        match self.model.cache() {
1082            EitherCache::Full(_) => FullCacheManager.clone_in_cache(self, seqs, false),
1083            EitherCache::Normal(_) => NormalCacheManager.clone_in_cache(self, seqs, false),
1084            EitherCache::Hybrid(_) => HybridCacheManager.clone_in_cache(self, seqs, false),
1085        }
1086    }
1087    fn clone_out_cache(&self, seqs: &mut [&mut Sequence]) {
1088        match self.model.cache() {
1089            EitherCache::Full(_) => FullCacheManager.clone_out_cache(self, seqs, false),
1090            EitherCache::Normal(_) => NormalCacheManager.clone_out_cache(self, seqs, false),
1091            EitherCache::Hybrid(_) => HybridCacheManager.clone_out_cache(self, seqs, false),
1092        }
1093    }
1094    fn set_none_cache(
1095        &self,
1096        seqs: &mut [&mut Sequence],
1097        reset_non_granular: bool,
1098        modify_draft_cache: bool,
1099        load_preallocated_cache: bool,
1100    ) {
1101        match self.model.cache() {
1102            EitherCache::Full(_) => {
1103                FullCacheManager.set_none_cache(self, seqs, modify_draft_cache, false)
1104            }
1105            EitherCache::Normal(_) => NormalCacheManager.set_none_cache(
1106                self,
1107                seqs,
1108                modify_draft_cache,
1109                load_preallocated_cache,
1110            ),
1111            EitherCache::Hybrid(_) => HybridCacheManager.set_none_cache(
1112                self,
1113                seqs,
1114                modify_draft_cache,
1115                load_preallocated_cache,
1116            ),
1117        }
1118        if reset_non_granular {
1119            self.reset_non_granular_state()
1120        }
1121    }
1122    fn cache(&self) -> &EitherCache {
1123        self.model.cache()
1124    }
1125}
1126
1127impl MetadataMixin for NormalPipeline {
1128    fn device(&self) -> Device {
1129        self.model.device().clone()
1130    }
1131    fn tokenizer(&self) -> Option<Arc<Tokenizer>> {
1132        Some(self.tokenizer.clone())
1133    }
1134    fn name(&self) -> String {
1135        self.model_id.clone()
1136    }
1137    fn reset_non_granular_state(&self) {
1138        if let Some(s) = self.non_granular_state.as_ref() {
1139            *self.cache().full().get_scalings_cache() = None;
1140            *get_mut_arcmutex!(s.non_granular_index) = 0;
1141        }
1142    }
1143    fn get_metadata(&self) -> Arc<GeneralMetadata> {
1144        self.metadata.clone()
1145    }
1146    fn generation_defaults(&self) -> Option<crate::ModelGenerationDefaults> {
1147        self.generation_defaults.clone()
1148    }
1149    fn device_mapper(&self) -> Option<&dyn DeviceMapper> {
1150        Some(&*self.mapper)
1151    }
1152}
1153
1154#[async_trait::async_trait]
1155impl Pipeline for NormalPipeline {
1156    fn forward_inputs(
1157        &mut self,
1158        inputs: Box<dyn Any>,
1159        return_raw_logits: bool,
1160    ) -> Result<ForwardInputsResult, candle_core::Error> {
1161        let ModelInputs {
1162            input_ids,
1163            input_ids_full,
1164            seqlen_offsets,
1165            seqlen_offsets_full,
1166            context_lens,
1167            position_ids,
1168            paged_attn_meta,
1169            flash_meta,
1170            flash_meta_full,
1171        } = *inputs.downcast().expect("Downcast failed.");
1172        let metadata = self.get_metadata();
1173        let paged_attn_meta = match (&metadata.cache_engine, &paged_attn_meta) {
1174            (Some(cache_engine), Some(meta)) => Some((cache_engine, meta)),
1175            (Some(_), None) => {
1176                // This can happen if Rust-side user code is wrong
1177                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.")
1178            }
1179            (None, Some(_)) => {
1180                // This should never happen but we handle it anyway
1181                candle_core::bail!("Forward step got a PagedAttention input metadata but there is no cache engine. Please raise an issue.")
1182            }
1183            (None, None) => None,
1184        };
1185        let logits = match self.model.is_xlora() {
1186            false => {
1187                let paged_attn_meta = paged_attn_meta
1188                    .as_ref()
1189                    .map(|meta| (meta.0.get_kv_cache().clone(), meta.1.clone()));
1190
1191                self.model.forward(
1192                    &input_ids,
1193                    &seqlen_offsets,
1194                    context_lens,
1195                    position_ids,
1196                    paged_attn_meta.as_ref().map(|(a, b)| (a.clone(), b)),
1197                    &flash_meta,
1198                )?
1199            }
1200            true => self.model.xlora_forward(
1201                &input_ids,
1202                input_ids_full.as_ref().unwrap_or(&input_ids),
1203                &seqlen_offsets,
1204                seqlen_offsets_full.as_ref().unwrap_or(&seqlen_offsets),
1205                self.no_kv_cache,
1206                &self.non_granular_state,
1207                context_lens,
1208                position_ids,
1209                &flash_meta,
1210                flash_meta_full.as_ref().unwrap_or(&flash_meta),
1211            )?,
1212        };
1213        if return_raw_logits {
1214            Ok(ForwardInputsResult::RawLogits { logits })
1215        } else {
1216            Ok(ForwardInputsResult::CausalGeneration { logits })
1217        }
1218    }
1219    async fn sample_causal_gen(
1220        &self,
1221        seqs: &mut [&mut Sequence],
1222        logits: Vec<Tensor>,
1223        prefix_cacher: &mut PrefixCacheManagerV2,
1224        disable_eos_stop: bool,
1225        rng: Arc<std::sync::Mutex<Isaac64Rng>>,
1226    ) -> Result<(), candle_core::Error> {
1227        sample_and_add_toks(self, seqs, logits, prefix_cacher, disable_eos_stop, rng).await
1228    }
1229    fn category(&self) -> ModelCategory {
1230        ModelCategory::Text
1231    }
1232}
1233
1234impl AnyMoePipelineMixin for NormalPipeline {
1235    fn amoe_finish_training(&mut self, gate_model_id: Option<String>) -> candle_core::Result<()> {
1236        self.model.finish_training(gate_model_id)
1237    }
1238    fn amoe_layer_vars(&self) -> Vec<Vec<Var>> {
1239        self.model.get_vars()
1240    }
1241    fn amoe_base_model_trainable_params(&self) -> usize {
1242        self.model.trainable_params()
1243    }
1244    fn amoe_take_cached_gating_outputs(&mut self) -> Vec<Tensor> {
1245        self.model.take_cached_gating_outputs()
1246    }
1247    fn amoe_create_layers(
1248        &mut self,
1249        model_ids: Vec<String>,
1250        token: &TokenSource,
1251        revision: Option<String>,
1252        match_regex: &str,
1253        config: crate::amoe::AnyMoeConfig,
1254        dtype: candle_core::DType,
1255        dev: &Device,
1256        (prefix, mlp): (String, String),
1257        layers: Vec<usize>,
1258        expert_type: AnyMoeExpertType,
1259        silent: bool,
1260        gate_model_id: Option<String>,
1261    ) -> candle_core::Result<()> {
1262        let mut vbs = Vec::new();
1263        // Precompile regex here
1264        let regex = Regex::new(match_regex).map_err(candle_core::Error::msg)?;
1265        for model_id in model_ids {
1266            let model_id_str = &model_id;
1267            let model_id = Path::new(&model_id);
1268
1269            let api = {
1270                let cache = GLOBAL_HF_CACHE.get().cloned().unwrap_or_default();
1271                let mut api = ApiBuilder::from_cache(cache)
1272                    .with_progress(!silent)
1273                    .with_token(get_token(token).map_err(candle_core::Error::msg)?);
1274                if let Some(cache_dir) = crate::hf_hub_cache_dir() {
1275                    api = api.with_cache_dir(cache_dir);
1276                }
1277                api.build().map_err(candle_core::Error::msg)?
1278            };
1279            let revision = revision.clone().unwrap_or("main".to_string());
1280            let api = api.repo(Repo::with_revision(
1281                model_id_str.clone(),
1282                RepoType::Model,
1283                revision.clone(),
1284            ));
1285
1286            let mut filenames = vec![];
1287            for rfilename in
1288                api_dir_list!(api, model_id, true).filter(|x| x.ends_with(".safetensors"))
1289            {
1290                filenames.push(api_get_file!(api, &rfilename, model_id));
1291            }
1292
1293            let regex = regex.clone();
1294            let match_regex_clone = match_regex.to_string();
1295            let layers_clone = layers.clone();
1296            let vb = from_mmaped_safetensors(
1297                filenames,
1298                vec![],
1299                Some(dtype),
1300                dev,
1301                vec![None],
1302                silent,
1303                None,
1304                move |key| {
1305                    if regex.is_match(&key) {
1306                        // Idx of the last char of the layer id, +1
1307                        // Assumes N.MLP
1308                        let last_layer_idx = key.find(&match_regex_clone).unwrap() - 1;
1309                        let first_layer_idx = key[..last_layer_idx].rfind('.').unwrap();
1310                        let layer_n = key[first_layer_idx + 1..last_layer_idx]
1311                            .parse::<usize>()
1312                            .unwrap();
1313                        layers_clone.contains(&layer_n) || layers_clone.is_empty()
1314                    } else {
1315                        false
1316                    }
1317                },
1318                Arc::new(|_| DeviceForLoadTensor::Base),
1319            )?;
1320            vbs.push(vb);
1321        }
1322
1323        let gate_vb = if let Some(gate_model_id) = gate_model_id {
1324            let model_id_str = &gate_model_id;
1325            let model_id = Path::new(&gate_model_id);
1326
1327            let api = {
1328                let cache = GLOBAL_HF_CACHE.get().cloned().unwrap_or_default();
1329                let mut api = ApiBuilder::from_cache(cache)
1330                    .with_progress(!silent)
1331                    .with_token(get_token(token).map_err(candle_core::Error::msg)?);
1332                if let Some(cache_dir) = crate::hf_hub_cache_dir() {
1333                    api = api.with_cache_dir(cache_dir);
1334                }
1335                api.build().map_err(candle_core::Error::msg)?
1336            };
1337            let revision = revision.clone().unwrap_or("main".to_string());
1338            let api = api.repo(Repo::with_revision(
1339                model_id_str.clone(),
1340                RepoType::Model,
1341                revision.clone(),
1342            ));
1343
1344            let mut gate_filenames = vec![];
1345            for rfilename in
1346                api_dir_list!(api, model_id, true).filter(|x| x.ends_with(".safetensors"))
1347            {
1348                gate_filenames.push(api_get_file!(api, &rfilename, model_id));
1349            }
1350            assert_eq!(
1351                gate_filenames.len(),
1352                1,
1353                "Gate model ID must contain only one .safetensors file"
1354            );
1355
1356            let vb = from_mmaped_safetensors(
1357                gate_filenames.clone(),
1358                vec![],
1359                Some(dtype),
1360                dev,
1361                vec![None],
1362                silent,
1363                None,
1364                |_| true,
1365                Arc::new(|_| DeviceForLoadTensor::Base),
1366            )?;
1367            info!(
1368                "Loaded gating layers from `{}`",
1369                gate_filenames[0].display()
1370            );
1371            Some(vb)
1372        } else {
1373            None
1374        };
1375
1376        self.model.create_anymoe_layers(
1377            vbs.clone(),
1378            config.clone(),
1379            (prefix.clone(), mlp.clone()),
1380            layers.clone(),
1381            expert_type.clone(),
1382            gate_vb.clone(),
1383        )?;
1384
1385        Ok(())
1386    }
1387    fn amoe_supported(&self) -> bool {
1388        self.model.amoe_supported()
1389    }
1390}
mistralrs_core/pipeline/normal.rs

mistralrs_core/pipeline/
normal.rs
