mistralrs/
text_model.rs

1use candle_core::Device;
2use mistralrs_core::*;
3use mistralrs_core::{SearchCallback, Tool, ToolCallback};
4use std::collections::HashMap;
5use std::{
6    ops::{Deref, DerefMut},
7    path::PathBuf,
8    sync::Arc,
9};
10
11use crate::{best_device, Model};
12
13/// A tool callback with its associated Tool definition.
14#[derive(Clone)]
15pub struct ToolCallbackWithTool {
16    pub callback: Arc<ToolCallback>,
17    pub tool: Tool,
18}
19
20#[derive(Clone)]
21/// Configure a text model with the various parameters for loading, running, and other inference behaviors.
22pub struct TextModelBuilder {
23    // Loading model
24    pub(crate) model_id: String,
25    pub(crate) token_source: TokenSource,
26    pub(crate) hf_revision: Option<String>,
27    pub(crate) write_uqff: Option<PathBuf>,
28    pub(crate) from_uqff: Option<Vec<PathBuf>>,
29    pub(crate) imatrix: Option<PathBuf>,
30    pub(crate) calibration_file: Option<PathBuf>,
31    pub(crate) chat_template: Option<String>,
32    pub(crate) jinja_explicit: Option<String>,
33    pub(crate) tokenizer_json: Option<String>,
34    pub(crate) device_mapping: Option<DeviceMapSetting>,
35    pub(crate) hf_cache_path: Option<PathBuf>,
36    pub(crate) search_bert_model: Option<BertEmbeddingModel>,
37    pub(crate) search_callback: Option<Arc<SearchCallback>>,
38    pub(crate) tool_callbacks: HashMap<String, Arc<ToolCallback>>,
39    pub(crate) tool_callbacks_with_tools: HashMap<String, ToolCallbackWithTool>,
40    pub(crate) mcp_client_config: Option<McpClientConfig>,
41    pub(crate) device: Option<Device>,
42    pub(crate) matformer_config_path: Option<PathBuf>,
43    pub(crate) matformer_slice_name: Option<String>,
44
45    // Model running
46    pub(crate) topology: Option<Topology>,
47    pub(crate) organization: IsqOrganization,
48    pub(crate) loader_type: Option<NormalLoaderType>,
49    pub(crate) dtype: ModelDType,
50    pub(crate) force_cpu: bool,
51    pub(crate) isq: Option<IsqType>,
52    pub(crate) throughput_logging: bool,
53
54    // Other things
55    pub(crate) paged_attn_cfg: Option<PagedAttentionConfig>,
56    pub(crate) max_num_seqs: usize,
57    pub(crate) no_kv_cache: bool,
58    pub(crate) with_logging: bool,
59    pub(crate) prefix_cache_n: Option<usize>,
60}
61
62/// Builder for PagedAttention metadata.
63pub struct PagedAttentionMetaBuilder {
64    block_size: Option<usize>,
65    mem_cpu: usize,
66    mem_gpu: MemoryGpuConfig,
67    cache_type: PagedCacheType,
68}
69
70impl Default for PagedAttentionMetaBuilder {
71    fn default() -> Self {
72        Self {
73            block_size: None,
74            mem_cpu: 64,
75            mem_gpu: MemoryGpuConfig::ContextSize(4096),
76            cache_type: PagedCacheType::Auto,
77        }
78    }
79}
80
81impl PagedAttentionMetaBuilder {
82    pub fn with_block_size(mut self, block_size: usize) -> Self {
83        self.block_size = Some(block_size);
84        self
85    }
86
87    pub fn with_gpu_memory(mut self, mem_gpu: MemoryGpuConfig) -> Self {
88        self.mem_gpu = mem_gpu;
89        self
90    }
91
92    pub fn with_paged_cache_type(mut self, cache_type: PagedCacheType) -> Self {
93        self.cache_type = cache_type;
94        self
95    }
96
97    pub fn build(self) -> anyhow::Result<PagedAttentionConfig> {
98        PagedAttentionConfig::new(self.block_size, self.mem_cpu, self.mem_gpu, self.cache_type)
99    }
100}
101
102impl TextModelBuilder {
103    /// A few defaults are applied here:
104    /// - MoQE ISQ organization
105    /// - Token source is from the cache (.cache/huggingface/token)
106    /// - Maximum number of sequences running is 32
107    /// - Number of sequences to hold in prefix cache is 16.
108    /// - Automatic device mapping with model defaults according to `AutoDeviceMapParams`
109    /// - By default, web searching compatible with the OpenAI `web_search_options` setting is disabled.
110    pub fn new(model_id: impl ToString) -> Self {
111        Self {
112            model_id: model_id.to_string(),
113            topology: None,
114            organization: IsqOrganization::Default,
115            write_uqff: None,
116            from_uqff: None,
117            chat_template: None,
118            tokenizer_json: None,
119            loader_type: None,
120            dtype: ModelDType::Auto,
121            force_cpu: false,
122            token_source: TokenSource::CacheToken,
123            hf_revision: None,
124            isq: None,
125            paged_attn_cfg: None,
126            max_num_seqs: 32,
127            no_kv_cache: false,
128            prefix_cache_n: Some(16),
129            with_logging: false,
130            device_mapping: None,
131            imatrix: None,
132            calibration_file: None,
133            jinja_explicit: None,
134            throughput_logging: false,
135            hf_cache_path: None,
136            search_bert_model: None,
137            search_callback: None,
138            tool_callbacks: HashMap::new(),
139            tool_callbacks_with_tools: HashMap::new(),
140            mcp_client_config: None,
141            device: None,
142            matformer_config_path: None,
143            matformer_slice_name: None,
144        }
145    }
146
147    /// Enable searching compatible with the OpenAI `web_search_options` setting. This uses the BERT model specified or the default.
148    pub fn with_search(mut self, search_bert_model: BertEmbeddingModel) -> Self {
149        self.search_bert_model = Some(search_bert_model);
150        self
151    }
152
153    /// Override the search function used when `web_search_options` is enabled.
154    pub fn with_search_callback(mut self, callback: Arc<SearchCallback>) -> Self {
155        self.search_callback = Some(callback);
156        self
157    }
158
159    /// Register a callback for a specific tool name.
160    pub fn with_tool_callback(
161        mut self,
162        name: impl Into<String>,
163        callback: Arc<ToolCallback>,
164    ) -> Self {
165        self.tool_callbacks.insert(name.into(), callback);
166        self
167    }
168
169    /// Register a callback with an associated Tool definition that will be automatically
170    /// added to requests when tool callbacks are active.
171    pub fn with_tool_callback_and_tool(
172        mut self,
173        name: impl Into<String>,
174        callback: Arc<ToolCallback>,
175        tool: Tool,
176    ) -> Self {
177        let name = name.into();
178        self.tool_callbacks_with_tools
179            .insert(name, ToolCallbackWithTool { callback, tool });
180        self
181    }
182
183    /// Configure MCP client to connect to external MCP servers and automatically
184    /// register their tools for use in automatic tool calling.
185    pub fn with_mcp_client(mut self, config: McpClientConfig) -> Self {
186        self.mcp_client_config = Some(config);
187        self
188    }
189
190    /// Enable runner throughput logging.
191    pub fn with_throughput_logging(mut self) -> Self {
192        self.throughput_logging = true;
193        self
194    }
195
196    /// Explicit JINJA chat template file (.jinja) to be used. If specified, this overrides all other chat templates.
197    pub fn with_jinja_explicit(mut self, jinja_explicit: String) -> Self {
198        self.jinja_explicit = Some(jinja_explicit);
199        self
200    }
201
202    /// Set the model topology for use during loading. If there is an overlap, the topology type is used over the ISQ type.
203    pub fn with_topology(mut self, topology: Topology) -> Self {
204        self.topology = Some(topology);
205        self
206    }
207
208    /// Organize ISQ to enable MoQE (Mixture of Quantized Experts, <https://arxiv.org/abs/2310.02410>)
209    pub fn with_mixture_qexperts_isq(mut self) -> Self {
210        self.organization = IsqOrganization::MoeExpertsOnly;
211        self
212    }
213
214    /// Literal Jinja chat template OR Path (ending in `.json`) to one.
215    pub fn with_chat_template(mut self, chat_template: impl ToString) -> Self {
216        self.chat_template = Some(chat_template.to_string());
217        self
218    }
219
220    /// Path to a discrete `tokenizer.json` file.
221    pub fn with_tokenizer_json(mut self, tokenizer_json: impl ToString) -> Self {
222        self.tokenizer_json = Some(tokenizer_json.to_string());
223        self
224    }
225
226    /// Manually set the model loader type. Otherwise, it will attempt to automatically
227    /// determine the loader type.
228    pub fn with_loader_type(mut self, loader_type: NormalLoaderType) -> Self {
229        self.loader_type = Some(loader_type);
230        self
231    }
232
233    /// Load the model in a certain dtype.
234    pub fn with_dtype(mut self, dtype: ModelDType) -> Self {
235        self.dtype = dtype;
236        self
237    }
238
239    /// Force usage of the CPU device. Do not use PagedAttention with this.
240    pub fn with_force_cpu(mut self) -> Self {
241        self.force_cpu = true;
242        self
243    }
244
245    /// Source of the Hugging Face token.
246    pub fn with_token_source(mut self, token_source: TokenSource) -> Self {
247        self.token_source = token_source;
248        self
249    }
250
251    /// Set the revision to use for a Hugging Face remote model.
252    pub fn with_hf_revision(mut self, revision: impl ToString) -> Self {
253        self.hf_revision = Some(revision.to_string());
254        self
255    }
256
257    /// Use ISQ of a certain type. If there is an overlap, the topology type is used over the ISQ type.
258    pub fn with_isq(mut self, isq: IsqType) -> Self {
259        self.isq = Some(isq);
260        self
261    }
262
263    /// Utilise this imatrix file during ISQ. Incompatible with specifying a calibration file.
264    pub fn with_imatrix(mut self, path: PathBuf) -> Self {
265        self.imatrix = Some(path);
266        self
267    }
268
269    /// Utilise this calibration file to collcet an imatrix. Incompatible with specifying a calibration file.
270    pub fn with_calibration_file(mut self, path: PathBuf) -> Self {
271        self.calibration_file = Some(path);
272        self
273    }
274
275    /// Enable PagedAttention. Configure PagedAttention with a [`PagedAttentionConfig`] object, which
276    /// can be created with sensible values with a [`PagedAttentionMetaBuilder`].
277    ///
278    /// If PagedAttention is not supported (query with [`paged_attn_supported`]), this will do nothing.
279    pub fn with_paged_attn(
280        mut self,
281        paged_attn_cfg: impl FnOnce() -> anyhow::Result<PagedAttentionConfig>,
282    ) -> anyhow::Result<Self> {
283        if paged_attn_supported() {
284            self.paged_attn_cfg = Some(paged_attn_cfg()?);
285        } else {
286            self.paged_attn_cfg = None;
287        }
288        Ok(self)
289    }
290
291    /// Set the maximum number of sequences which can be run at once.
292    pub fn with_max_num_seqs(mut self, max_num_seqs: usize) -> Self {
293        self.max_num_seqs = max_num_seqs;
294        self
295    }
296
297    /// Disable KV cache. Trade performance for memory usage.
298    pub fn with_no_kv_cache(mut self) -> Self {
299        self.no_kv_cache = true;
300        self
301    }
302
303    /// Set the number of sequences to hold in the prefix cache. Set to `None` to disable the prefix cacher.
304    pub fn with_prefix_cache_n(mut self, n_seqs: Option<usize>) -> Self {
305        self.prefix_cache_n = n_seqs;
306        self
307    }
308
309    /// Enable logging.
310    pub fn with_logging(mut self) -> Self {
311        self.with_logging = true;
312        self
313    }
314
315    /// Provide metadata to initialize the device mapper.
316    pub fn with_device_mapping(mut self, device_mapping: DeviceMapSetting) -> Self {
317        self.device_mapping = Some(device_mapping);
318        self
319    }
320
321    #[deprecated(
322        note = "Use `UqffTextModelBuilder` to load a UQFF model instead of the generic `from_uqff`"
323    )]
324    /// Path to read a `.uqff` file from. Other necessary configuration files must be present at this location.
325    ///
326    /// For example, these include:
327    /// - `residual.safetensors`
328    /// - `tokenizer.json`
329    /// - `config.json`
330    /// - More depending on the model
331    pub fn from_uqff(mut self, path: Vec<PathBuf>) -> Self {
332        self.from_uqff = Some(path);
333        self
334    }
335
336    /// Path to write a `.uqff` file to and serialize the other necessary files.
337    ///
338    /// The parent (part of the path excluding the filename) will determine where any other files
339    /// serialized are written to.
340    ///
341    /// For example, these include:
342    /// - `residual.safetensors`
343    /// - `tokenizer.json`
344    /// - `config.json`
345    /// - More depending on the model
346    pub fn write_uqff(mut self, path: PathBuf) -> Self {
347        self.write_uqff = Some(path);
348        self
349    }
350
351    /// Cache path for Hugging Face models downloaded locally
352    pub fn from_hf_cache_pathf(mut self, hf_cache_path: PathBuf) -> Self {
353        self.hf_cache_path = Some(hf_cache_path);
354        self
355    }
356
357    /// Set the main device to load this model onto. Automatic device mapping will be performed starting with this device.
358    pub fn with_device(mut self, device: Device) -> Self {
359        self.device = Some(device);
360        self
361    }
362
363    /// Path to a Matryoshka Transformer configuration CSV file.
364    pub fn with_matformer_config_path(mut self, path: PathBuf) -> Self {
365        self.matformer_config_path = Some(path);
366        self
367    }
368
369    /// Name of the slice to use from the Matryoshka Transformer configuration.
370    pub fn with_matformer_slice_name(mut self, name: String) -> Self {
371        self.matformer_slice_name = Some(name);
372        self
373    }
374
375    pub async fn build(self) -> anyhow::Result<Model> {
376        let config = NormalSpecificConfig {
377            topology: self.topology,
378            organization: self.organization,
379            write_uqff: self.write_uqff,
380            from_uqff: self.from_uqff,
381            imatrix: self.imatrix,
382            calibration_file: self.calibration_file,
383            hf_cache_path: self.hf_cache_path,
384            matformer_config_path: self.matformer_config_path,
385            matformer_slice_name: self.matformer_slice_name,
386        };
387
388        if self.with_logging {
389            initialize_logging();
390        }
391
392        let loader = NormalLoaderBuilder::new(
393            config,
394            self.chat_template,
395            self.tokenizer_json,
396            Some(self.model_id),
397            self.no_kv_cache,
398            self.jinja_explicit,
399        )
400        .build(self.loader_type)?;
401
402        // Load, into a Pipeline
403        let pipeline = loader.load_model_from_hf(
404            self.hf_revision,
405            self.token_source,
406            &self.dtype,
407            &self.device.unwrap_or(best_device(self.force_cpu).unwrap()),
408            !self.with_logging,
409            self.device_mapping
410                .unwrap_or(DeviceMapSetting::Auto(AutoDeviceMapParams::default_text())),
411            self.isq,
412            self.paged_attn_cfg,
413        )?;
414
415        let scheduler_method = match self.paged_attn_cfg {
416            Some(_) => {
417                let config = pipeline
418                    .lock()
419                    .await
420                    .get_metadata()
421                    .cache_config
422                    .as_ref()
423                    .cloned();
424
425                if let Some(config) = config {
426                    SchedulerConfig::PagedAttentionMeta {
427                        max_num_seqs: self.max_num_seqs,
428                        config,
429                    }
430                } else {
431                    SchedulerConfig::DefaultScheduler {
432                        method: DefaultSchedulerMethod::Fixed(self.max_num_seqs.try_into()?),
433                    }
434                }
435            }
436            None => SchedulerConfig::DefaultScheduler {
437                method: DefaultSchedulerMethod::Fixed(self.max_num_seqs.try_into()?),
438            },
439        };
440
441        let mut runner = MistralRsBuilder::new(
442            pipeline,
443            scheduler_method,
444            self.throughput_logging,
445            self.search_bert_model,
446        );
447        if let Some(cb) = self.search_callback.clone() {
448            runner = runner.with_search_callback(cb);
449        }
450        for (name, cb) in &self.tool_callbacks {
451            runner = runner.with_tool_callback(name.clone(), cb.clone());
452        }
453        for (name, callback_with_tool) in &self.tool_callbacks_with_tools {
454            runner = runner.with_tool_callback_and_tool(
455                name.clone(),
456                callback_with_tool.callback.clone(),
457                callback_with_tool.tool.clone(),
458            );
459        }
460        if let Some(mcp_config) = self.mcp_client_config {
461            runner = runner.with_mcp_client(mcp_config);
462        }
463        runner = runner
464            .with_no_kv_cache(self.no_kv_cache)
465            .with_no_prefix_cache(self.prefix_cache_n.is_none());
466
467        if let Some(n) = self.prefix_cache_n {
468            runner = runner.with_prefix_cache_n(n)
469        }
470
471        Ok(Model::new(runner.build().await))
472    }
473}
474
475#[derive(Clone)]
476/// Configure a UQFF text model with the various parameters for loading, running, and other inference behaviors.
477/// This wraps and implements `DerefMut` for the TextModelBuilder, so users should take care to not call UQFF-related methods.
478pub struct UqffTextModelBuilder(TextModelBuilder);
479
480impl UqffTextModelBuilder {
481    /// A few defaults are applied here:
482    /// - MoQE ISQ organization
483    /// - Token source is from the cache (.cache/huggingface/token)
484    /// - Maximum number of sequences running is 32
485    /// - Number of sequences to hold in prefix cache is 16.
486    /// - Automatic device mapping with model defaults according to `AutoDeviceMapParams`
487    pub fn new(model_id: impl ToString, uqff_file: Vec<PathBuf>) -> Self {
488        let mut inner = TextModelBuilder::new(model_id);
489        inner.from_uqff = Some(uqff_file);
490        Self(inner)
491    }
492
493    pub async fn build(self) -> anyhow::Result<Model> {
494        self.0.build().await
495    }
496
497    /// This wraps the VisionModelBuilder, so users should take care to not call UQFF-related methods.
498    pub fn into_inner(self) -> TextModelBuilder {
499        self.0
500    }
501}
502
503impl Deref for UqffTextModelBuilder {
504    type Target = TextModelBuilder;
505
506    fn deref(&self) -> &Self::Target {
507        &self.0
508    }
509}
510
511impl DerefMut for UqffTextModelBuilder {
512    fn deref_mut(&mut self) -> &mut Self::Target {
513        &mut self.0
514    }
515}
516
517impl From<UqffTextModelBuilder> for TextModelBuilder {
518    fn from(value: UqffTextModelBuilder) -> Self {
519        value.0
520    }
521}
mistralrs/text_model.rs

mistralrs/
text_model.rs
