use super::isq::ImatrixDataSource;
use super::llg::build_llg_factory;
use super::{
get_model_paths, get_xlora_paths, text_models_inputs_processor::ModelInputs, AdapterKind,
CacheManager, GeneralMetadata, Loader, ModelKind, ModelPaths, NormalModel, NormalModelLoader,
TokenSource,
};
use super::{
AnyMoePipelineMixin, CacheManagerMixin, EitherCache, ForwardInputsResult, IsqOrganization,
IsqPipelineMixin, MetadataMixin, ModelCategory, PreProcessingMixin,
};
use super::{
AutoNormalLoader, DeepSeekV2Loader, DeepSeekV3Loader, GLM4Loader, GLM4MoeLiteLoader,
GLM4MoeLoader, Gemma2Loader, GemmaLoader, GptOssLoader, GraniteMoeHybridLoader, LlamaLoader,
MistralLoader, MixtralLoader, NormalLoaderType, Phi2Loader, Phi3Loader, Phi3_5MoELoader,
Qwen2Loader, Qwen3Loader, Qwen3MoELoader, Qwen3NextLoader, SmolLm3Loader, Starcoder2Loader,
};
use crate::amoe::AnyMoeExpertType;
use crate::attention::ATTENTION_CHUNK_SIZE;
use crate::device_map::{self, DeviceMapper};
use crate::distributed::{self, WorkerTransferData};
use crate::kv_cache::{FullCacheManager, HybridCacheManager, NormalCacheManager};
use crate::lora::Ordering;
use crate::paged_attention::{calculate_cache_config, AttentionImplementation, CacheEngine};
use crate::pipeline::chat_template::{calculate_eos_tokens, GenerationConfig};
use crate::pipeline::isq::UqffFullSer;
use crate::pipeline::loaders::auto_device_map;
use crate::pipeline::loaders::QuantizationConfigShim;
use crate::pipeline::sampling::sample_and_add_toks;
use crate::pipeline::text_models_inputs_processor::make_prompt_chunk;
use crate::pipeline::{get_chat_template, Modalities, SupportedModality};
use crate::pipeline::{ChatTemplate, LocalModelPaths};
use crate::prefix_cacher::PrefixCacheManagerV2;
use crate::sequence::Sequence;
use crate::utils::tokenizer::get_tokenizer;
use crate::utils::varbuilder_utils::DeviceForLoadTensor;
use crate::utils::{
progress::{new_multi_progress, ProgressScopeGuard},
tokens::get_token,
varbuilder_utils::from_mmaped_safetensors,
};
use crate::xlora_models::NonGranularState;
use crate::{
api_dir_list, api_get_file, get_mut_arcmutex, get_paths, get_uqff_paths, lora_model_loader,
normal_model_loader, normal_model_loader_sharded, xlora_model_loader, DeviceMapSetting,
PagedAttentionConfig, Pipeline, Topology, TryIntoDType, GLOBAL_HF_CACHE,
};
use anyhow::Result;
use candle_core::{Device, Tensor, Var};
use hf_hub::Cache;
use hf_hub::{api::sync::ApiBuilder, Repo, RepoType};
use mistralrs_quant::log::once_log_info;
use mistralrs_quant::{
AfqLayer, GgufMatMul, HqqLayer, ImmediateIsqOverride, IsqType, QuantizedSerdeType,
};
use rand_isaac::Isaac64Rng;
use regex_automata::meta::Regex;
use std::any::Any;
use std::borrow::Cow;
use std::path::{Path, PathBuf};
use std::str::FromStr;
use std::sync::{Arc, RwLock};
use std::time::Instant;
use std::{env, fs};
use tokenizers::Tokenizer;
use tokio::sync::Mutex;
use tracing::{info, warn};
pub struct NormalPipeline {
model: Box<dyn NormalModel + Send + Sync>,
tokenizer: Arc<Tokenizer>,
no_kv_cache: bool,
chat_template: Arc<ChatTemplate>,
non_granular_state: Option<NonGranularState>,
model_id: String,
metadata: Arc<GeneralMetadata>,
topology: Option<Topology>,
silent: bool,
organization: IsqOrganization,
template_filename: Option<PathBuf>,
generation_config: Option<PathBuf>,
generation_defaults: Option<crate::ModelGenerationDefaults>,
config: String,
imatrix: Option<PathBuf>,
mapper: Box<dyn DeviceMapper + Send + Sync>,
}
pub struct NormalLoader {
inner: Box<dyn NormalModelLoader>,
model_id: String,
config: NormalSpecificConfig,
xlora_model_id: Option<String>,
lora_adapter_ids: Option<Vec<String>>,
kind: ModelKind,
xlora_order: Option<Ordering>,
no_kv_cache: bool,
chat_template: Option<String>,
tokenizer_json: Option<String>,
tgt_non_granular_index: Option<usize>,
token_source: RwLock<Option<TokenSource>>,
revision: RwLock<Option<String>>,
from_uqff: RwLock<Option<Vec<PathBuf>>>,
jinja_explicit: Option<String>,
hf_cache_path: Option<PathBuf>,
}
#[derive(Default)]
pub struct NormalLoaderBuilder {
model_id: Option<String>,
config: NormalSpecificConfig,
xlora_model_id: Option<String>,
lora_adapter_ids: Option<Vec<String>>,
kind: ModelKind,
xlora_order: Option<Ordering>,
no_kv_cache: bool,
chat_template: Option<String>,
tokenizer_json: Option<String>,
tgt_non_granular_index: Option<usize>,
jinja_explicit: Option<String>,
hf_cache_path: Option<PathBuf>,
}
#[derive(Clone, Default)]
pub struct NormalSpecificConfig {
pub topology: Option<Topology>,
pub organization: IsqOrganization,
pub write_uqff: Option<PathBuf>,
pub from_uqff: Option<Vec<PathBuf>>,
pub imatrix: Option<PathBuf>,
pub calibration_file: Option<PathBuf>,
pub hf_cache_path: Option<PathBuf>,
pub matformer_config_path: Option<PathBuf>,
pub matformer_slice_name: Option<String>,
}
impl NormalLoaderBuilder {
pub fn new(
config: NormalSpecificConfig,
chat_template: Option<String>,
tokenizer_json: Option<String>,
model_id: Option<String>,
no_kv_cache: bool,
jinja_explicit: Option<String>,
) -> Self {
Self {
config,
chat_template,
tokenizer_json,
model_id,
kind: ModelKind::Normal,
jinja_explicit,
no_kv_cache,
..Default::default()
}
}
fn with_adapter(
mut self,
xlora_model_id: String,
xlora_order: Ordering,
no_kv_cache: bool,
tgt_non_granular_index: Option<usize>,
) -> Self {
self.xlora_model_id = Some(xlora_model_id);
self.xlora_order = Some(xlora_order);
self.no_kv_cache = no_kv_cache;
self.tgt_non_granular_index = tgt_non_granular_index;
self.model_id = if let Some(id) = self.model_id {
Some(id)
} else {
info!(
"Using adapter base model ID: `{}`",
self.xlora_order.as_ref().unwrap().base_model_id
);
Some(self.xlora_order.as_ref().unwrap().base_model_id.clone())
};
self
}
pub fn with_xlora(
mut self,
xlora_model_id: String,
xlora_order: Ordering,
no_kv_cache: bool,
tgt_non_granular_index: Option<usize>,
) -> Self {
self.kind = ModelKind::Adapter {
adapter: AdapterKind::XLora,
};
self.with_adapter(
xlora_model_id,
xlora_order,
no_kv_cache,
tgt_non_granular_index,
)
}
pub fn with_lora(mut self, lora_adapter_ids: Vec<String>) -> Self {
self.kind = ModelKind::Adapter {
adapter: AdapterKind::Lora,
};
self.lora_adapter_ids = Some(lora_adapter_ids);
self
}
pub fn hf_cache_path(mut self, hf_cache_path: PathBuf) -> Self {
self.hf_cache_path = Some(hf_cache_path);
self
}
pub fn build(self, loader_tp: Option<NormalLoaderType>) -> anyhow::Result<Box<dyn Loader>> {
let loader: Box<dyn NormalModelLoader> = match loader_tp {
Some(NormalLoaderType::Mistral) => Box::new(MistralLoader),
Some(NormalLoaderType::Gemma) => Box::new(GemmaLoader),
Some(NormalLoaderType::Llama) => Box::new(LlamaLoader),
Some(NormalLoaderType::Mixtral) => Box::new(MixtralLoader),
Some(NormalLoaderType::Phi2) => Box::new(Phi2Loader),
Some(NormalLoaderType::Phi3) => Box::new(Phi3Loader),
Some(NormalLoaderType::Qwen2) => Box::new(Qwen2Loader),
Some(NormalLoaderType::Gemma2) => Box::new(Gemma2Loader),
Some(NormalLoaderType::Starcoder2) => Box::new(Starcoder2Loader),
Some(NormalLoaderType::Phi3_5MoE) => Box::new(Phi3_5MoELoader),
Some(NormalLoaderType::DeepSeekV2) => Box::new(DeepSeekV2Loader),
Some(NormalLoaderType::DeepSeekV3) => Box::new(DeepSeekV3Loader),
Some(NormalLoaderType::Qwen3) => Box::new(Qwen3Loader),
Some(NormalLoaderType::GLM4) => Box::new(GLM4Loader),
Some(NormalLoaderType::GLM4MoeLite) => Box::new(GLM4MoeLiteLoader),
Some(NormalLoaderType::GLM4Moe) => Box::new(GLM4MoeLoader),
Some(NormalLoaderType::Qwen3Moe) => Box::new(Qwen3MoELoader),
Some(NormalLoaderType::SmolLm3) => Box::new(SmolLm3Loader),
Some(NormalLoaderType::GraniteMoeHybrid) => Box::new(GraniteMoeHybridLoader),
Some(NormalLoaderType::GptOss) => Box::new(GptOssLoader),
Some(NormalLoaderType::Qwen3Next) => Box::new(Qwen3NextLoader),
None => Box::new(AutoNormalLoader),
};
Ok(Box::new(NormalLoader {
inner: loader,
model_id: self.model_id.unwrap(),
config: self.config,
xlora_model_id: self.xlora_model_id,
lora_adapter_ids: self.lora_adapter_ids,
kind: self.kind,
xlora_order: self.xlora_order,
no_kv_cache: self.no_kv_cache,
chat_template: self.chat_template,
tokenizer_json: self.tokenizer_json,
tgt_non_granular_index: self.tgt_non_granular_index,
jinja_explicit: self.jinja_explicit,
token_source: RwLock::new(None),
revision: RwLock::new(None),
from_uqff: RwLock::new(None),
hf_cache_path: self.hf_cache_path,
}))
}
}
impl Loader for NormalLoader {
#[allow(clippy::type_complexity, clippy::too_many_arguments)]
fn load_model_from_hf(
&self,
revision: Option<String>,
token_source: TokenSource,
dtype: &dyn TryIntoDType,
device: &Device,
silent: bool,
mapper: DeviceMapSetting,
in_situ_quant: Option<IsqType>,
paged_attn_config: Option<PagedAttentionConfig>,
) -> Result<Arc<Mutex<dyn Pipeline + Send + Sync>>> {
let _progress_guard = ProgressScopeGuard::new(silent);
let cache = self
.hf_cache_path
.clone()
.map(Cache::new)
.unwrap_or_default();
GLOBAL_HF_CACHE.get_or_init(|| cache);
let paths: anyhow::Result<Box<dyn ModelPaths>> = get_paths!(
LocalModelPaths,
&token_source,
revision.clone(),
self,
None,
None,
silent,
self.config.from_uqff.is_some()
);
*self
.token_source
.write()
.expect("Failed to write to token source") = Some(token_source);
*self.revision.write().expect("Failed to write to revision") = revision.clone();
if let Some(from_uqff) = self.config.from_uqff.clone() {
*self.from_uqff.write().unwrap() = Some(get_uqff_paths!(&from_uqff, self, silent));
}
self.load_model_from_path(
&paths?,
dtype,
device,
silent,
mapper,
in_situ_quant,
paged_attn_config,
)
}
#[allow(clippy::type_complexity, clippy::too_many_arguments)]
fn load_model_from_path(
&self,
paths: &Box<dyn ModelPaths>,
dtype: &dyn TryIntoDType,
device: &Device,
silent: bool,
mut mapper: DeviceMapSetting,
in_situ_quant: Option<IsqType>,
mut paged_attn_config: Option<PagedAttentionConfig>,
) -> Result<Arc<Mutex<dyn Pipeline + Send + Sync>>> {
let _progress_guard = ProgressScopeGuard::new(silent);
let config = std::fs::read_to_string(paths.get_config_filename())?;
if !self.inner.supports_paged_attention(&config)? {
paged_attn_config = None;
}
info!("Prompt chunk size is {ATTENTION_CHUNK_SIZE}.");
let use_nccl = mistralrs_quant::distributed::use_nccl();
let available_devices = if let Ok(payload) = env::var(distributed::IS_DAEMON_FLAG) {
let payload: WorkerTransferData = serde_json::from_str(&payload)?;
let WorkerTransferData::Init { id: _, worker_rank } = payload;
vec![candle_core::Device::new_cuda(worker_rank + 1)?]
} else if use_nccl {
vec![candle_core::Device::new_cuda(0)?]
} else {
device_map::get_all_similar_devices(device)?
};
#[cfg(feature = "cuda")]
for device in &available_devices {
if let Device::Cuda(dev) = device {
unsafe { dev.disable_event_tracking() };
}
}
let device = if use_nccl || cfg!(feature = "ring") {
available_devices[0].clone()
} else {
device.clone()
};
let mut max_kv_tokens: Option<usize> = None;
if use_nccl || cfg!(feature = "ring") {
mapper = DeviceMapSetting::DummyNccl {
nm_device: available_devices[0].clone(),
};
} else if let DeviceMapSetting::Auto(params) = mapper.clone() {
max_kv_tokens = Some(params.max_seq_len() * params.max_batch_size());
let dtype = dtype.try_into_dtype(&available_devices.iter().collect::<Vec<_>>())?;
let (layer_sizes_in_bytes, non_mapped_size_in_bytes, total_model_size_in_bytes) =
if let Some(serialized) = &*self.from_uqff.read().unwrap() {
let weight_pack_factor = {
let ser_artifacts = unsafe {
candle_core::safetensors::MmapedSafetensors::multi(serialized)?
};
let mut total_pack_factors = 0;
let total_tensors = ser_artifacts.tensors().len();
for (_, artifact) in ser_artifacts.tensors() {
let artifact = artifact.data();
let isq_type = artifact[mistralrs_quant::UQFF_QUANT_TYPE_OFFSET];
let pack_factor = match QuantizedSerdeType::try_from(isq_type as usize)?
{
QuantizedSerdeType::Hqq => {
HqqLayer::get_isq_type_from_uqff(Cow::Borrowed(artifact))?
.pack_factor(dtype)
}
QuantizedSerdeType::Gguf => {
GgufMatMul::get_isq_type_from_uqff(Cow::Borrowed(artifact))?
.pack_factor(dtype)
}
QuantizedSerdeType::Fp8 => IsqType::F8E4M3.pack_factor(dtype),
QuantizedSerdeType::Unquant => 1,
QuantizedSerdeType::Afq => {
AfqLayer::get_isq_type_from_uqff(Cow::Borrowed(artifact))?
.pack_factor(dtype)
}
QuantizedSerdeType::F8Q8 => IsqType::F8Q8.pack_factor(dtype),
QuantizedSerdeType::Mxfp4 => IsqType::MXFP4.pack_factor(dtype),
};
total_pack_factors += pack_factor;
}
total_pack_factors / total_tensors
};
let layer_sizes_in_bytes = self.inner.layer_sizes_in_bytes(
&config,
dtype,
weight_pack_factor,
None,
)?;
let non_mapped_size_in_bytes = self.inner.non_mapped_size_in_bytes(
&config,
dtype,
weight_pack_factor,
None,
)?;
let layer_sizes_sum = layer_sizes_in_bytes.iter().sum::<usize>();
(
layer_sizes_in_bytes,
non_mapped_size_in_bytes,
layer_sizes_sum + non_mapped_size_in_bytes,
)
} else if let Some(isq) = in_situ_quant {
let weight_pack_factor = isq.pack_factor(dtype);
let layer_sizes_in_bytes = self.inner.layer_sizes_in_bytes(
&config,
dtype,
weight_pack_factor,
None,
)?;
let non_mapped_size_in_bytes = self.inner.non_mapped_size_in_bytes(
&config,
dtype,
weight_pack_factor,
None,
)?;
let layer_sizes_sum = layer_sizes_in_bytes.iter().sum::<usize>();
(
layer_sizes_in_bytes,
non_mapped_size_in_bytes,
layer_sizes_sum + non_mapped_size_in_bytes,
)
} else {
let weight_pack_factor =
QuantizationConfigShim::get_quant_config_pack_factor(&config, dtype)?;
let layer_sizes_in_bytes = self.inner.layer_sizes_in_bytes(
&config,
dtype,
weight_pack_factor,
None,
)?;
let non_mapped_size_in_bytes = self.inner.non_mapped_size_in_bytes(
&config,
dtype,
weight_pack_factor,
None,
)?;
let layer_sizes_sum = layer_sizes_in_bytes.iter().sum::<usize>();
(
layer_sizes_in_bytes,
non_mapped_size_in_bytes,
layer_sizes_sum + non_mapped_size_in_bytes,
)
};
let new = auto_device_map::get_device_layers(
&*self.inner,
&config,
self.inner.num_layers(&config)?,
layer_sizes_in_bytes,
non_mapped_size_in_bytes,
total_model_size_in_bytes,
&available_devices,
dtype,
¶ms,
paged_attn_config.as_ref(),
)?;
mapper = DeviceMapSetting::Map(new);
}
let pipeline_mapper = mapper.into_mapper(
self.inner.num_layers(&config)?,
&device,
self.config.topology.as_ref(),
&available_devices,
)?;
let mapper = mapper.into_mapper(
self.inner.num_layers(&config)?,
&device,
self.config.topology.as_ref(),
&available_devices,
)?;
let mut layer_devices = Vec::new();
for layer in 0..self.inner.num_layers(&config)? {
let device = mapper.device_for(layer, false).cloned();
layer_devices.push(device);
}
let dtype = mapper.get_min_dtype(dtype)?;
let mapping_uses_cpu = mapper.get_unique_devices().iter().any(Device::is_cpu);
if mapping_uses_cpu && paged_attn_config.is_some() {
warn!("Device mapping contains a mix of GPU and CPU. There is no CPU support for PagedAttention, disabling PagedAttention.");
paged_attn_config = None;
}
info!("Model config: {:?}", self.inner.get_config_repr(&config)?);
if crate::using_flash_attn() {
once_log_info("FlashAttention is enabled.");
}
let topology_overrides = self
.config
.topology
.as_ref()
.map(|topology| {
topology
.pattern_overrides()
.into_iter()
.map(|(regex, layer)| ImmediateIsqOverride {
predicate: regex,
ty: layer.isq,
device: layer.device.clone(),
})
.collect::<Vec<_>>()
})
.unwrap_or_default();
let has_override_isq = topology_overrides
.iter()
.any(|override_entry| override_entry.ty.is_some());
let topology_requires_post_quant = self
.config
.topology
.as_ref()
.is_some_and(|topology| topology.requires_post_quantization());
let allow_immediate_cli = self.config.imatrix.is_none()
&& self.config.calibration_file.is_none()
&& in_situ_quant.is_some();
let mut immediate_ty = None;
let mut immediate_predicates = Vec::new();
if allow_immediate_cli {
immediate_ty = in_situ_quant;
immediate_predicates =
if matches!(self.config.organization, IsqOrganization::MoeExpertsOnly) {
self.inner.immediate_isq_predicates_moqe(&config)?
} else {
self.inner.immediate_isq_predicates(&config)?
};
info!("Applying ISQ to {in_situ_quant:?}");
if immediate_predicates.is_empty() {
warn!("No predicates for this model and ISQ setting detected. ISQ will not be applied to any weights!");
}
}
let use_immediate = allow_immediate_cli || has_override_isq;
if use_immediate {
let (pool, num_threads) = mistralrs_quant::create_isq_thread_pool(immediate_ty);
info!("Applying immediate ISQ in parallel on {num_threads} threads.");
mistralrs_quant::set_immediate_isq_with_pool(
immediate_ty,
immediate_predicates.clone(),
topology_overrides.clone(),
pool,
);
}
let mut loading_isq = if use_immediate {
false
} else {
in_situ_quant.is_some()
};
if self.config.imatrix.is_some() || self.config.calibration_file.is_some() {
loading_isq = true;
}
loading_isq |= topology_requires_post_quant;
loading_isq |= self.config.from_uqff.is_some();
if self.config.imatrix.is_some() && self.config.calibration_file.is_some() {
anyhow::bail!(
"`imatrix` and `calibration_file` were both specified, this is not allowed."
);
}
let load_device = if !loading_isq || self.config.calibration_file.is_some() {
loading_isq = false;
if use_immediate && !crate::utils::normal::is_integrated_gpu(&device) {
Device::Cpu
} else {
device.clone()
}
} else {
Device::Cpu
};
let is_xlora = self.kind.is_adapted_and(|a| a.is_x_lora());
let attention_mechanism = if paged_attn_config.is_some() {
AttentionImplementation::PagedAttention
} else {
AttentionImplementation::Eager
};
let multi_progress = Arc::new(new_multi_progress());
let matformer_slicing_config = if let Some(matformer_path) =
&self.config.matformer_config_path
{
use crate::matformer::{MatformerConfig, MatformerSliceConfig};
info!("Loading Matformer config from {:?}", matformer_path);
let config = Arc::new(MatformerConfig::from_file(matformer_path)?);
if let Some(slice_name) = &self.config.matformer_slice_name {
info!("Using Matformer slice: {}", slice_name);
Some(MatformerSliceConfig::new(slice_name.clone(), config))
} else {
warn!("Matformer config loaded but no slice name specified. Models will use their default slice.");
None
}
} else {
None
};
let mut model = if use_nccl || cfg!(feature = "ring") {
let (mapper, sharded_vb) = distributed::prepare_distributed_mapper(
dtype,
&device,
&available_devices,
silent,
&config,
loading_isq,
self.config.from_uqff.is_some(),
self.config.organization,
&*self.inner,
paths.as_ref(),
)?;
match self.kind {
ModelKind::Normal => normal_model_loader_sharded!(
sharded_vb,
config,
self.inner,
mapper,
loading_isq,
device.clone(),
attention_mechanism,
multi_progress.clone(),
matformer_slicing_config.clone(),
),
ModelKind::Adapter {
adapter: AdapterKind::XLora,
} => xlora_model_loader!(
paths,
Some(dtype),
&load_device,
layer_devices.clone(),
config,
self.inner,
silent,
mapper,
loading_isq,
device.clone(),
multi_progress.clone(),
matformer_slicing_config.clone(),
),
ModelKind::Adapter {
adapter: AdapterKind::Lora,
} => lora_model_loader!(
paths,
Some(dtype),
&load_device,
layer_devices.clone(),
config,
self.inner,
silent,
mapper,
loading_isq,
self.config.from_uqff.is_some(),
device.clone(),
attention_mechanism,
matches!(self.config.organization, IsqOrganization::MoeExpertsOnly),
multi_progress.clone(),
matformer_slicing_config.clone(),
),
_ => unreachable!(),
}
} else {
match self.kind {
ModelKind::Normal => normal_model_loader!(
paths,
Some(dtype),
&load_device,
layer_devices.clone(),
config,
self.inner,
silent,
mapper,
loading_isq,
self.config.from_uqff.is_some(),
device.clone(),
attention_mechanism,
matches!(self.config.organization, IsqOrganization::MoeExpertsOnly),
multi_progress.clone(),
matformer_slicing_config.clone(),
),
ModelKind::Adapter {
adapter: AdapterKind::XLora,
} => xlora_model_loader!(
paths,
Some(dtype),
&load_device,
layer_devices.clone(),
config,
self.inner,
silent,
mapper,
loading_isq,
device.clone(),
multi_progress.clone(),
matformer_slicing_config.clone(),
),
ModelKind::Adapter {
adapter: AdapterKind::Lora,
} => lora_model_loader!(
paths,
Some(dtype),
&load_device,
layer_devices.clone(),
config,
self.inner,
silent,
mapper,
loading_isq,
self.config.from_uqff.is_some(),
device.clone(),
attention_mechanism,
matches!(self.config.organization, IsqOrganization::MoeExpertsOnly),
multi_progress.clone(),
matformer_slicing_config.clone(),
),
_ => unreachable!(),
}
};
let tokenizer = get_tokenizer(paths.get_tokenizer_filename(), None)?;
let gen_conf: Option<GenerationConfig> = paths.get_gen_conf_filename().and_then(|f| {
match serde_json::from_str::<GenerationConfig>(&fs::read_to_string(f).unwrap()) {
Ok(conf) => Some(conf),
Err(e) => {
warn!("Failed to parse generation_config.json: {}", e);
None
}
}
});
let chat_template_explicit = paths
.get_chat_template_explicit()
.as_ref()
.map(|x| x.to_string_lossy().to_string());
let chat_template = get_chat_template(
paths,
self.jinja_explicit.as_ref(),
chat_template_explicit.as_ref(),
self.chat_template.as_ref(),
None,
);
if let Some(calibration_file) = &self.config.calibration_file {
let calibration_data = std::fs::read_to_string(calibration_file)?;
let tokens = tokenizer
.encode_fast(calibration_data, false)
.map_err(anyhow::Error::msg)?
.get_ids()
.to_vec();
info!(
"Collecting imatrix from calibration file `{}` of {} tokens.",
calibration_file.display(),
tokens.len()
);
let bos_tok_id = chat_template
.bos_tok()
.as_deref()
.and_then(|tok| tokenizer.token_to_id(tok));
match self.config.organization {
IsqOrganization::Default => model.begin_track_stats()?,
IsqOrganization::MoeExpertsOnly => model.begin_track_stats_moe_experts_only()?,
}
const CHUNK_SIZE: usize = 1024;
let n_chunks = tokens.len().div_ceil(CHUNK_SIZE);
let start = Instant::now();
for (i, chunk) in tokens.chunks(CHUNK_SIZE).enumerate() {
let mut chunk = chunk.to_vec();
if let Some(bos_tok_id) = bos_tok_id {
chunk.insert(0, bos_tok_id);
}
let chunk_len = chunk.len();
let start = Instant::now();
let inputs = make_prompt_chunk(
0,
vec![&chunk],
&[0],
&load_device,
None,
false,
None,
Some(pipeline_mapper.as_ref()),
None,
model.config().sliding_window,
)?;
model.forward(
&inputs.input.to_device(model.device())?,
&inputs.positions,
inputs.context_lens.clone(),
inputs.position_ids.clone(),
None,
&inputs.flash_meta.clone(),
)?;
match model.cache_mut() {
EitherCache::Full(full) => {
for layer in &mut *full.lock() {
*layer = None
}
}
EitherCache::Normal(normal) => {
for layer in &mut *normal.lock().unwrap().0 {
layer.reset();
}
}
EitherCache::Hybrid(hybrid) => {
hybrid.lock().unwrap().reset();
}
}
let end = Instant::now();
info!(
"Processed chunk {}/{n_chunks} ({chunk_len} tokens), {:.2}s",
i + 1,
end.duration_since(start).as_secs_f32()
);
}
load_device.synchronize()?;
let end = Instant::now();
info!(
"Finished collecting imatrix in {:.2}s",
end.duration_since(start).as_secs_f32()
);
}
let should_serialize = self.config.write_uqff.is_some();
let should_quantize_pass = loading_isq;
if (should_quantize_pass || should_serialize) && self.config.from_uqff.is_none() {
let imatrix_source = if should_quantize_pass {
match (
self.config.imatrix.as_ref(),
self.config.calibration_file.is_some(),
) {
(None, false) => None,
(Some(file), false) => Some(ImatrixDataSource::File(file)),
(None, true) => Some(ImatrixDataSource::Collected),
(Some(_), true) => unreachable!(),
}
} else {
None
};
if should_quantize_pass {
info!("Applying ISQ to all ranks.");
} else {
info!("Serializing existing ISQ tensors without additional quantization.");
}
let multi_progress = Arc::new(new_multi_progress());
model.quantize(
in_situ_quant,
model.device().clone(),
self.config.topology.as_ref(),
silent,
imatrix_source,
self.config.organization,
should_quantize_pass,
self.config.write_uqff.as_ref(),
UqffFullSer {
tokenizer: &tokenizer,
template_filename: paths.get_template_filename(),
generation_config: paths.get_gen_conf_filename(),
config: config.clone(),
processor_filename: &None,
preprocessor_filename: &None,
modules: None,
module_paths: None,
},
multi_progress.clone(),
)?;
} else if let Some(from_uqff) = &*self.from_uqff.read().unwrap() {
model.load_from_artifacts(
device.clone(),
self.config.topology.as_ref(),
silent,
from_uqff,
)?;
}
let paged_attn_config = if matches!(
self.kind,
ModelKind::Adapter {
adapter: AdapterKind::XLora
}
) {
warn!(
"Adapter parallel_models do not currently support PagedAttention, running without"
);
None
} else {
paged_attn_config
};
let model_metadata = model.model_config();
let (cache_config, cache_engine) = if let Some(paged_attn_config) = paged_attn_config {
let cache_config = calculate_cache_config(
paged_attn_config.mem_gpu,
paged_attn_config.block_size,
dtype,
paged_attn_config.cache_type,
model_metadata.as_ref(),
&device,
&pipeline_mapper
.get_unique_devices()
.into_iter()
.map(Some)
.collect::<Vec<_>>(),
silent,
None,
max_kv_tokens,
)?;
let mut layer_devices = Vec::new();
for layer in 0..self.inner.num_layers(&config)? {
let device = model.get_layers().1.device_for(layer, false).cloned();
layer_devices.push(device);
}
let cache_engine = CacheEngine::new(
model_metadata.as_ref(),
&cache_config,
dtype,
model.device(),
layer_devices.clone(),
)?;
(Some(cache_config), Some(cache_engine))
} else {
(None, None)
};
let max_seq_len = model.max_seq_len();
let llg_factory = build_llg_factory(tokenizer.clone())?;
let num_hidden_layers = match model.cache() {
EitherCache::Full(full) => full.lock().len(),
EitherCache::Normal(normal) => normal.lock().unwrap().0.len(),
EitherCache::Hybrid(hybrid) => hybrid.lock().unwrap().num_layers(),
};
let generation_defaults = gen_conf
.as_ref()
.and_then(GenerationConfig::generation_defaults);
let eos = calculate_eos_tokens(&chat_template, gen_conf.as_ref(), &tokenizer);
let sliding_window = model.config().sliding_window;
Ok(Arc::new(Mutex::new(NormalPipeline {
model,
tokenizer: tokenizer.into(),
no_kv_cache: self.no_kv_cache,
chat_template: Arc::new(chat_template),
non_granular_state: self.tgt_non_granular_index.map(|tgt_non_granular_index| {
NonGranularState {
non_granular_index: Arc::new(Mutex::new(0)),
tgt_non_granular_index,
}
}),
model_id: self.model_id.clone(),
metadata: Arc::new(GeneralMetadata {
max_seq_len,
llg_factory: Some(llg_factory),
no_kv_cache: self.no_kv_cache,
no_prefix_cache: is_xlora,
num_hidden_layers,
eos_tok: eos,
kind: self.kind.clone(),
is_xlora,
activation_dtype: dtype,
sliding_window,
cache_config,
cache_engine,
model_metadata: Some(model_metadata),
modalities: Modalities {
input: vec![SupportedModality::Text],
output: vec![SupportedModality::Text],
},
}),
topology: self.config.topology.clone(),
silent,
organization: self.config.organization,
template_filename: paths.get_template_filename().clone(),
generation_config: paths.get_gen_conf_filename().cloned(),
generation_defaults,
config,
imatrix: self.config.imatrix.clone(),
mapper: pipeline_mapper,
})))
}
fn get_id(&self) -> String {
self.model_id.clone()
}
fn get_kind(&self) -> ModelKind {
self.kind.clone()
}
}
impl PreProcessingMixin for NormalPipeline {
fn get_chat_template(&self) -> Option<Arc<ChatTemplate>> {
Some(self.chat_template.clone())
}
fn get_input_processor_config(&self) -> Option<Arc<dyn Any>> {
None
}
}
impl IsqPipelineMixin for NormalPipeline {
fn re_isq_model(&mut self, dtype: IsqType) -> Result<()> {
let device = self.device().clone();
let multi_progress = Arc::new(new_multi_progress());
self.model.quantize(
Some(dtype),
device.clone(),
self.topology.as_ref(),
self.silent,
self.imatrix.as_ref().map(ImatrixDataSource::File),
self.organization,
true,
None,
UqffFullSer {
tokenizer: &self.tokenizer,
template_filename: &self.template_filename,
generation_config: self.generation_config.as_ref(),
config: self.config.clone(),
processor_filename: &None,
preprocessor_filename: &None,
modules: None,
module_paths: None,
},
multi_progress.clone(),
)?;
Ok(())
}
}
impl CacheManagerMixin for NormalPipeline {
fn clone_in_cache(&self, seqs: &mut [&mut Sequence]) {
match self.model.cache() {
EitherCache::Full(_) => FullCacheManager.clone_in_cache(self, seqs, false),
EitherCache::Normal(_) => NormalCacheManager.clone_in_cache(self, seqs, false),
EitherCache::Hybrid(_) => HybridCacheManager.clone_in_cache(self, seqs, false),
}
}
fn clone_out_cache(&self, seqs: &mut [&mut Sequence]) {
match self.model.cache() {
EitherCache::Full(_) => FullCacheManager.clone_out_cache(self, seqs, false),
EitherCache::Normal(_) => NormalCacheManager.clone_out_cache(self, seqs, false),
EitherCache::Hybrid(_) => HybridCacheManager.clone_out_cache(self, seqs, false),
}
}
fn set_none_cache(
&self,
seqs: &mut [&mut Sequence],
reset_non_granular: bool,
modify_draft_cache: bool,
load_preallocated_cache: bool,
) {
match self.model.cache() {
EitherCache::Full(_) => {
FullCacheManager.set_none_cache(self, seqs, modify_draft_cache, false)
}
EitherCache::Normal(_) => NormalCacheManager.set_none_cache(
self,
seqs,
modify_draft_cache,
load_preallocated_cache,
),
EitherCache::Hybrid(_) => HybridCacheManager.set_none_cache(
self,
seqs,
modify_draft_cache,
load_preallocated_cache,
),
}
if reset_non_granular {
self.reset_non_granular_state()
}
}
fn cache(&self) -> &EitherCache {
self.model.cache()
}
}
impl MetadataMixin for NormalPipeline {
fn device(&self) -> Device {
self.model.device().clone()
}
fn tokenizer(&self) -> Option<Arc<Tokenizer>> {
Some(self.tokenizer.clone())
}
fn name(&self) -> String {
self.model_id.clone()
}
fn reset_non_granular_state(&self) {
if let Some(s) = self.non_granular_state.as_ref() {
*self.cache().full().get_scalings_cache() = None;
*get_mut_arcmutex!(s.non_granular_index) = 0;
}
}
fn get_metadata(&self) -> Arc<GeneralMetadata> {
self.metadata.clone()
}
fn generation_defaults(&self) -> Option<crate::ModelGenerationDefaults> {
self.generation_defaults.clone()
}
fn device_mapper(&self) -> Option<&dyn DeviceMapper> {
Some(&*self.mapper)
}
}
#[async_trait::async_trait]
impl Pipeline for NormalPipeline {
fn forward_inputs(
&mut self,
inputs: Box<dyn Any>,
return_raw_logits: bool,
) -> Result<ForwardInputsResult, candle_core::Error> {
let ModelInputs {
input_ids,
input_ids_full,
seqlen_offsets,
seqlen_offsets_full,
context_lens,
position_ids,
paged_attn_meta,
flash_meta,
flash_meta_full,
} = *inputs.downcast().expect("Downcast failed.");
let metadata = self.get_metadata();
let paged_attn_meta = match (&metadata.cache_engine, &paged_attn_meta) {
(Some(cache_engine), Some(meta)) => Some((cache_engine, meta)),
(Some(_), None) => {
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.")
}
(None, Some(_)) => {
candle_core::bail!("Forward step got a PagedAttention input metadata but there is no cache engine. Please raise an issue.")
}
(None, None) => None,
};
let logits = match self.model.is_xlora() {
false => {
let paged_attn_meta = paged_attn_meta
.as_ref()
.map(|meta| (meta.0.get_kv_cache().clone(), meta.1.clone()));
self.model.forward(
&input_ids,
&seqlen_offsets,
context_lens,
position_ids,
paged_attn_meta.as_ref().map(|(a, b)| (a.clone(), b)),
&flash_meta,
)?
}
true => self.model.xlora_forward(
&input_ids,
input_ids_full.as_ref().unwrap_or(&input_ids),
&seqlen_offsets,
seqlen_offsets_full.as_ref().unwrap_or(&seqlen_offsets),
self.no_kv_cache,
&self.non_granular_state,
context_lens,
position_ids,
&flash_meta,
flash_meta_full.as_ref().unwrap_or(&flash_meta),
)?,
};
if return_raw_logits {
Ok(ForwardInputsResult::RawLogits { logits })
} else {
Ok(ForwardInputsResult::CausalGeneration { logits })
}
}
async fn sample_causal_gen(
&self,
seqs: &mut [&mut Sequence],
logits: Vec<Tensor>,
prefix_cacher: &mut PrefixCacheManagerV2,
disable_eos_stop: bool,
rng: Arc<std::sync::Mutex<Isaac64Rng>>,
) -> Result<(), candle_core::Error> {
sample_and_add_toks(self, seqs, logits, prefix_cacher, disable_eos_stop, rng).await
}
fn category(&self) -> ModelCategory {
ModelCategory::Text
}
}
impl AnyMoePipelineMixin for NormalPipeline {
fn amoe_finish_training(&mut self, gate_model_id: Option<String>) -> candle_core::Result<()> {
self.model.finish_training(gate_model_id)
}
fn amoe_layer_vars(&self) -> Vec<Vec<Var>> {
self.model.get_vars()
}
fn amoe_base_model_trainable_params(&self) -> usize {
self.model.trainable_params()
}
fn amoe_take_cached_gating_outputs(&mut self) -> Vec<Tensor> {
self.model.take_cached_gating_outputs()
}
fn amoe_create_layers(
&mut self,
model_ids: Vec<String>,
token: &TokenSource,
revision: Option<String>,
match_regex: &str,
config: crate::amoe::AnyMoeConfig,
dtype: candle_core::DType,
dev: &Device,
(prefix, mlp): (String, String),
layers: Vec<usize>,
expert_type: AnyMoeExpertType,
silent: bool,
gate_model_id: Option<String>,
) -> candle_core::Result<()> {
let mut vbs = Vec::new();
let regex = Regex::new(match_regex).map_err(candle_core::Error::msg)?;
for model_id in model_ids {
let model_id_str = &model_id;
let model_id = Path::new(&model_id);
let api = {
let cache = GLOBAL_HF_CACHE.get().cloned().unwrap_or_default();
let mut api = ApiBuilder::from_cache(cache)
.with_progress(!silent)
.with_token(get_token(token).map_err(candle_core::Error::msg)?);
if let Some(cache_dir) = crate::hf_hub_cache_dir() {
api = api.with_cache_dir(cache_dir);
}
api.build().map_err(candle_core::Error::msg)?
};
let revision = revision.clone().unwrap_or("main".to_string());
let api = api.repo(Repo::with_revision(
model_id_str.clone(),
RepoType::Model,
revision.clone(),
));
let mut filenames = vec![];
for rfilename in
api_dir_list!(api, model_id, true).filter(|x| x.ends_with(".safetensors"))
{
filenames.push(api_get_file!(api, &rfilename, model_id));
}
let regex = regex.clone();
let match_regex_clone = match_regex.to_string();
let layers_clone = layers.clone();
let vb = from_mmaped_safetensors(
filenames,
vec![],
Some(dtype),
dev,
vec![None],
silent,
None,
move |key| {
if regex.is_match(&key) {
let last_layer_idx = key.find(&match_regex_clone).unwrap() - 1;
let first_layer_idx = key[..last_layer_idx].rfind('.').unwrap();
let layer_n = key[first_layer_idx + 1..last_layer_idx]
.parse::<usize>()
.unwrap();
layers_clone.contains(&layer_n) || layers_clone.is_empty()
} else {
false
}
},
Arc::new(|_| DeviceForLoadTensor::Base),
)?;
vbs.push(vb);
}
let gate_vb = if let Some(gate_model_id) = gate_model_id {
let model_id_str = &gate_model_id;
let model_id = Path::new(&gate_model_id);
let api = {
let cache = GLOBAL_HF_CACHE.get().cloned().unwrap_or_default();
let mut api = ApiBuilder::from_cache(cache)
.with_progress(!silent)
.with_token(get_token(token).map_err(candle_core::Error::msg)?);
if let Some(cache_dir) = crate::hf_hub_cache_dir() {
api = api.with_cache_dir(cache_dir);
}
api.build().map_err(candle_core::Error::msg)?
};
let revision = revision.clone().unwrap_or("main".to_string());
let api = api.repo(Repo::with_revision(
model_id_str.clone(),
RepoType::Model,
revision.clone(),
));
let mut gate_filenames = vec![];
for rfilename in
api_dir_list!(api, model_id, true).filter(|x| x.ends_with(".safetensors"))
{
gate_filenames.push(api_get_file!(api, &rfilename, model_id));
}
assert_eq!(
gate_filenames.len(),
1,
"Gate model ID must contain only one .safetensors file"
);
let vb = from_mmaped_safetensors(
gate_filenames.clone(),
vec![],
Some(dtype),
dev,
vec![None],
silent,
None,
|_| true,
Arc::new(|_| DeviceForLoadTensor::Base),
)?;
info!(
"Loaded gating layers from `{}`",
gate_filenames[0].display()
);
Some(vb)
} else {
None
};
self.model.create_anymoe_layers(
vbs.clone(),
config.clone(),
(prefix.clone(), mlp.clone()),
layers.clone(),
expert_type.clone(),
gate_vb.clone(),
)?;
Ok(())
}
fn amoe_supported(&self) -> bool {
self.model.amoe_supported()
}
}