chore: Rust to 1.89 and edition 2024 (#2659)

lib/async-openai/src/types/assistant_stream.rs

@@ -13,7 +13,7 @@ use std::pin::Pin;
 use futures::Stream;
 use serde::Deserialize;
 
-use crate::error::{map_deserialization_error, ApiError, OpenAIError};
+use crate::error::{ApiError, OpenAIError, map_deserialization_error};
 
 use super::{
     MessageDeltaObject, MessageObject, RunObject, RunStepDeltaObject, RunStepObject, ThreadObject,

lib/async-openai/src/types/embedding.rs

@@ -8,7 +8,7 @@
 // Modifications Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
 // Licensed under Apache 2.0
 
-use base64::engine::{general_purpose, Engine};
+use base64::engine::{Engine, general_purpose};
 use derive_builder::Builder;
 use serde::{Deserialize, Serialize};
 

lib/async-openai/src/types/impls.rs

@@ -24,7 +24,6 @@ use crate::{
 use bytes::Bytes;
 
 use super::{
-    responses::{CodeInterpreterContainer, Input, InputContent, Role as ResponsesRole},
     AddUploadPartRequest, AudioInput, AudioResponseFormat, ChatCompletionFunctionCall,
     ChatCompletionFunctions, ChatCompletionNamedToolChoice, ChatCompletionRequestAssistantMessage,
     ChatCompletionRequestAssistantMessageContent, ChatCompletionRequestDeveloperMessage,
@@ -40,6 +39,7 @@ use super::{
     EmbeddingInput, FileInput, FilePurpose, FunctionName, Image, ImageInput, ImageModel,
     ImageResponseFormat, ImageSize, ImageUrl, ImagesResponse, ModerationInput, Prompt, Role, Stop,
     TimestampGranularity,
+    responses::{CodeInterpreterContainer, Input, InputContent, Role as ResponsesRole},
 };
 
 /// for `impl_from!(T, Enum)`, implements

lib/async-openai/src/uploads.rs

@@ -9,10 +9,10 @@
 // Licensed under Apache 2.0
 
 use crate::{
+    Client,
     config::Config,
     error::OpenAIError,
     types::{AddUploadPartRequest, CompleteUploadRequest, CreateUploadRequest, Upload, UploadPart},
-    Client,
 };
 
 /// Allows you to upload large files in multiple parts.

lib/async-openai/src/users.rs

@@ -11,10 +11,10 @@
 use serde::Serialize;
 
 use crate::{
+    Client,
     config::Config,
     error::OpenAIError,
     types::{User, UserDeleteResponse, UserListResponse, UserRoleUpdateRequest},
-    Client,
 };
 
 /// Manage users and their role in an organization. Users will be automatically added to the Default project.

lib/async-openai/src/util.rs

@@ -29,7 +29,7 @@ pub(crate) async fn file_stream_body(source: InputSource) -> Result<Body, OpenAI
         _ => {
             return Err(OpenAIError::FileReadError(
                 "Cannot create stream from non-file source".to_string(),
-            ))
+            ));
         }
     };
     Ok(body)

lib/async-openai/src/vector_store_file_batches.rs

@@ -11,12 +11,12 @@
 use serde::Serialize;
 
 use crate::{
+    Client,
     config::Config,
     error::OpenAIError,
     types::{
         CreateVectorStoreFileBatchRequest, ListVectorStoreFilesResponse, VectorStoreFileBatchObject,
     },
-    Client,
 };
 
 /// Vector store file batches represent operations to add multiple files to a vector store.

lib/async-openai/src/vector_store_files.rs

@@ -11,13 +11,13 @@
 use serde::Serialize;
 
 use crate::{
+    Client,
     config::Config,
     error::OpenAIError,
     types::{
         CreateVectorStoreFileRequest, DeleteVectorStoreFileResponse, ListVectorStoreFilesResponse,
         VectorStoreFileContentResponse, VectorStoreFileObject,
     },
-    Client,
 };
 
 /// Vector store files represent files inside a vector store.

lib/async-openai/src/vector_stores.rs

@@ -11,6 +11,7 @@
 use serde::Serialize;
 
 use crate::{
+    Client, VectorStoreFiles,
     config::Config,
     error::OpenAIError,
     types::{
@@ -19,7 +20,6 @@ use crate::{
         VectorStoreSearchResultsPage,
     },
     vector_store_file_batches::VectorStoreFileBatches,
-    Client, VectorStoreFiles,
 };
 
 pub struct VectorStores<'c, C: Config> {

lib/async-openai/tests/bring-your-own-type.rs

@@ -12,9 +12,9 @@
 //! The purpose of this test to make sure that all _byot methods compiles with custom types.
 use std::pin::Pin;
 
-use dynamo_async_openai::{error::OpenAIError, Client};
+use dynamo_async_openai::{Client, error::OpenAIError};
 use futures::Stream;
-use serde_json::{json, Value};
+use serde_json::{Value, json};
 
 impl dynamo_async_openai::traits::AsyncTryFrom<MyJson> for reqwest::multipart::Form {
     type Error = OpenAIError;

lib/async-openai/tests/whisper.rs

@@ -9,7 +9,7 @@
 // Licensed under Apache 2.0
 
 use dynamo_async_openai::types::CreateTranslationRequestArgs;
-use dynamo_async_openai::{types::CreateTranscriptionRequestArgs, Client};
+use dynamo_async_openai::{Client, types::CreateTranscriptionRequestArgs};
 use tokio_test::assert_err;
 
 #[tokio::test]

lib/bindings/c/src/lib.rs

@@ -48,7 +48,7 @@ pub enum DynamoLlmResult {
 
 /// # Safety
 /// the namespace_c_str and component_c_str are passed as pointers to C strings
-#[no_mangle]
+#[unsafe(no_mangle)]
 pub unsafe extern "C" fn dynamo_llm_init(
     namespace_c_str: *const c_char,
     component_c_str: *const c_char,
@@ -108,7 +108,7 @@ pub unsafe extern "C" fn dynamo_llm_init(
     }
 }
 
-#[no_mangle]
+#[unsafe(no_mangle)]
 pub extern "C" fn dynamo_llm_shutdown() -> DynamoLlmResult {
     let wk = match WK.get() {
         Some(wk) => wk,
@@ -123,7 +123,7 @@ pub extern "C" fn dynamo_llm_shutdown() -> DynamoLlmResult {
     DynamoLlmResult::OK
 }
 
-#[no_mangle]
+#[unsafe(no_mangle)]
 pub extern "C" fn dynamo_llm_load_publisher_create() -> DynamoLlmResult {
     DynamoLlmResult::OK
 }
@@ -191,11 +191,7 @@ fn kv_event_create_stored_from_parts(
         if num_toks != (kv_block_size as usize) {
             if WARN_COUNT
                 .fetch_update(Ordering::SeqCst, Ordering::SeqCst, |c| {
-                    if c < 3 {
-                        Some(c + 1)
-                    } else {
-                        None
-                    }
+                    if c < 3 { Some(c + 1) } else { None }
                 })
                 .is_ok()
             {
@@ -256,7 +252,7 @@ pub struct DynamoKvStoredEventParams {
 /// # Safety
 /// parent_hash is passed as pointer to indicate whether the blocks
 /// has a parent hash or not. nullptr is used to represent no parent hash
-#[no_mangle]
+#[unsafe(no_mangle)]
 pub unsafe extern "C" fn dynamo_kv_event_publish_stored(
     event_id: u64,
     token_ids: *const u32,
@@ -293,7 +289,7 @@ pub unsafe extern "C" fn dynamo_kv_event_publish_stored(
     }
 }
 
-#[no_mangle]
+#[unsafe(no_mangle)]
 pub extern "C" fn dynamo_kv_event_publish_removed(
     event_id: u64,
     block_ids: *const u64,

lib/engines/llamacpp/src/lib.rs

@@ -10,17 +10,17 @@ use std::{
 use async_stream::stream;
 use dynamo_runtime::engine::{AsyncEngine, AsyncEngineContextProvider, ResponseStream};
 use dynamo_runtime::pipeline::error as pipeline_error;
-use dynamo_runtime::pipeline::{async_trait, Error, ManyOut, SingleIn};
+use dynamo_runtime::pipeline::{Error, ManyOut, SingleIn, async_trait};
 use dynamo_runtime::protocols::annotated::Annotated;
 use dynamo_runtime::{CancellationToken, ErrorContext, Result};
 use llama_cpp_2::{
-    context::{params::LlamaContextParams, LlamaContext},
+    LogOptions,
+    context::{LlamaContext, params::LlamaContextParams},
     llama_backend::LlamaBackend,
     llama_batch::LlamaBatch,
-    model::{params::LlamaModelParams, LlamaModel},
+    model::{LlamaModel, params::LlamaModelParams},
     sampling::LlamaSampler,
     token::LlamaToken,
-    LogOptions,
 };
 
 use dynamo_llm::protocols::common::llm_backend::LLMEngineOutput;

lib/engines/mistralrs/src/lib.rs

@@ -25,7 +25,7 @@ use dynamo_runtime::protocols::annotated::Annotated;
 
 use dynamo_llm::protocols::openai::{
     chat_completions::{NvCreateChatCompletionRequest, NvCreateChatCompletionStreamResponse},
-    completions::{prompt_to_string, NvCreateCompletionRequest, NvCreateCompletionResponse},
+    completions::{NvCreateCompletionRequest, NvCreateCompletionResponse, prompt_to_string},
     embeddings::{NvCreateEmbeddingRequest, NvCreateEmbeddingResponse},
 };
 
@@ -240,9 +240,10 @@ impl MistralRsEngine {
         }));
 
         // Send warmup request and consume response
-        if let Ok(sender) = engine.mistralrs.get_sender(None) {
-            if let Ok(()) = sender.send(warmup_request).await {
-                if let Some(response) = rx.recv().await {
+        if let Ok(sender) = engine.mistralrs.get_sender(None)
+            && let Ok(()) = sender.send(warmup_request).await
+            && let Some(response) = rx.recv().await
+        {
             match response.as_result() {
                 Ok(r) => {
                     tracing::debug!(request_id, "Warmup response: {r:?}");
@@ -252,8 +253,6 @@ impl MistralRsEngine {
                 }
             }
         }
-            }
-        }
 
         Ok(engine)
     }

lib/llm/benches/tokenizer.rs

@@ -4,13 +4,13 @@
 use std::hint::black_box;
 use std::sync::Arc;
 
-use criterion::{criterion_group, criterion_main, Criterion, Throughput};
+use criterion::{Criterion, Throughput, criterion_group, criterion_main};
 
 use dynamo_llm::backend::Decoder;
 use dynamo_llm::protocols::common::StopConditions;
+use dynamo_llm::tokenizers::DecodeStream;
 use dynamo_llm::tokenizers::hf::HuggingFaceTokenizer;
 use dynamo_llm::tokenizers::traits::{Encoder, Tokenizer};
-use dynamo_llm::tokenizers::DecodeStream;
 use dynamo_llm::types::TokenIdType;
 
 const TEST_TOKENIZER: &str = concat!(

lib/llm/src/backend.rs

@@ -24,22 +24,22 @@ use tracing as log;
 use crate::model_card::{ModelDeploymentCard, TokenizerKind};
 use dynamo_runtime::{
     pipeline::{
-        async_trait, AsyncEngineContextProvider, ManyOut, Operator, ResponseStream,
-        ServerStreamingEngine, SingleIn,
+        AsyncEngineContextProvider, ManyOut, Operator, ResponseStream, ServerStreamingEngine,
+        SingleIn, async_trait,
     },
     protocols::annotated::Annotated,
 };
 
 use crate::protocols::{
+    TokenIdType,
     common::{
+        StopConditions,
         llm_backend::{
             BackendOutput, EmbeddingsEngineOutput, FinishReason, LLMEngineOutput,
             PreprocessedRequest,
         },
         preprocessor::PreprocessedEmbeddingRequest,
-        StopConditions,
     },
-    TokenIdType,
 };
 use crate::tokenizers::{DecodeStream, HuggingFaceTokenizer, Tokenizer};
 use tokenizers::Tokenizer as HfTokenizer;
@@ -149,11 +149,12 @@ impl
                     }
 
                     // if we have a data field without an event, then we might need to update the data
-                    if let Some(data) = &output.data {
-                        if data.text.is_some() && !state.validate_engine_decode {
+                    if let Some(data) = &output.data
+                        && data.text.is_some()
+                        && !state.validate_engine_decode
+                    {
                         return Some((output, state));
                     }
-                    }
 
                     let data = output.data.as_ref().unwrap();
 
@@ -425,8 +426,9 @@ impl Decoder {
 
         // check stop sequences - the jail will always hold at least the largest stop sequence
         // if jail_max_bytes is 0, then there are no stop sequences
-        if self.jail_max_bytes > 0 {
-            if let Some(token) = &token {
+        if self.jail_max_bytes > 0
+            && let Some(token) = &token
+        {
             let pre_append = self.jail.len();
             log::debug!("pre_append: {}", pre_append);
             log::debug!("jail: {}", self.jail);
@@ -436,8 +438,7 @@ impl Decoder {
 
             for seq in &self.hidden_stop_sequences {
                 log::debug!("stop seq: {}", seq);
-                    if let Some(offset) =
-                        galil_seiferas::gs_find(self.jail.as_bytes(), seq.as_bytes())
+                if let Some(offset) = galil_seiferas::gs_find(self.jail.as_bytes(), seq.as_bytes())
                 {
                     log::debug!("offset: {}", offset);
                     // return only new bytes after pre_append .. offset+seq.len()
@@ -465,7 +466,6 @@ impl Decoder {
                 self.jail.drain(0..drain_len);
             }
         }
-        }
 
         Ok(StepResult::ok(token))
     }
@@ -485,12 +485,10 @@ impl Decoder {
                 .map(|x| x.should_hide_text())
                 .unwrap_or(false);
 
-            if !hide_text {
-                if let Some(token) = &token {
+            if !hide_text && let Some(token) = &token {
                 text.get_or_insert_with(|| String::with_capacity(token_ids.len()))
                     .push_str(token);
             }
-            }
             tokens.push(token);
 
             if let Some(stop_trigger) = stop_trigger {

lib/llm/src/block_manager.rs

@@ -38,18 +38,18 @@ pub mod controller;
 
 pub use crate::common::dtype::DType;
 pub use block::{
+    BasicMetadata, BlockMetadata, Blocks, ImmutableBlock, MutableBlock,
     locality::{self, LocalityProvider, LogicalResources},
     nixl::{BlockDescriptorList, IsImmutable, IsMutable, MutabilityKind, RemoteBlock},
-    BasicMetadata, BlockMetadata, Blocks, ImmutableBlock, MutableBlock,
 };
 pub use config::*;
 
-pub use layout::{nixl::NixlLayout, LayoutConfig, LayoutConfigBuilder, LayoutError, LayoutType};
+pub use layout::{LayoutConfig, LayoutConfigBuilder, LayoutError, LayoutType, nixl::NixlLayout};
 pub use offload::request::BlockResult;
 pub use pool::{BlockPool, ManagedBlockPool};
 pub use storage::{
-    nixl::NixlRegisterableStorage, DeviceStorage, DiskStorage, PinnedStorage, Storage,
-    StorageAllocator,
+    DeviceStorage, DiskStorage, PinnedStorage, Storage, StorageAllocator,
+    nixl::NixlRegisterableStorage,
 };
 pub use tokio_util::sync::CancellationToken;
 

lib/llm/src/block_manager/block.rs

@@ -21,7 +21,7 @@ pub mod registry;
 pub mod state;
 pub mod transfer;
 
-pub use data::{view, BlockData, BlockDataExt, BlockDataProvider, BlockDataProviderMut};
+pub use data::{BlockData, BlockDataExt, BlockDataProvider, BlockDataProviderMut, view};
 pub use locality::LocalityProvider;
 
 pub use crate::tokens::TokenBlockError;
@@ -37,10 +37,10 @@ use crate::block_manager::{
 use crate::tokens::{SaltHash, SequenceHash, Token, TokenBlock, Tokens};
 
 use super::{
+    WorkerID,
     events::PublishHandle,
     layout::{BlockLayout, LayoutError, LayoutType},
     storage::StorageType,
-    WorkerID,
 };
 
 use derive_getters::Getters;
@@ -657,12 +657,12 @@ impl<S: Storage, L: LocalityProvider, M: BlockMetadata> std::fmt::Debug for Muta
 impl<S: Storage, L: LocalityProvider, M: BlockMetadata> Drop for MutableBlock<S, L, M> {
     fn drop(&mut self) {
         tracing::debug!("drop: {:?}", self);
-        if let Some(block) = self.block.take() {
-            if self.return_tx.send(block).is_err() {
+        if let Some(block) = self.block.take()
+            && self.return_tx.send(block).is_err()
+        {
             tracing::warn!("block pool shutdown before block was returned");
         }
     }
-    }
 }
 
 impl<S: Storage, L: LocalityProvider, M: BlockMetadata> Deref for MutableBlock<S, L, M> {
@@ -957,9 +957,9 @@ pub mod nixl {
     use super::view::{BlockKind, Kind, LayerKind};
 
     use super::super::{
+        WorkerID,
         layout::nixl::{NixlLayout, SerializedNixlBlockLayout},
         storage::nixl::{MemType, NixlRegisterableStorage, NixlStorage},
-        WorkerID,
     };
 
     use derive_getters::{Dissolve, Getters};
@@ -1360,9 +1360,7 @@ pub mod nixl {
         #[error("Input block list cannot be empty")]
         EmptyInput,
 
-        #[error(
-            "Blocks in the input list are not homogeneous (worker_id, block_set_idx mismatch)"
-        )]
+        #[error("Blocks in the input list are not homogeneous (worker_id, block_set_idx mismatch)")]
         NotHomogeneous,
 
         #[error("Serialization failed: {0}")]

lib/llm/src/block_manager/block/data.rs

@@ -46,7 +46,11 @@ pub trait BlockDataExt<S: Storage>: Send + Sync + 'static + std::fmt::Debug {
     fn is_local_mut(&mut self) -> Option<&mut dyn BlockDataViews<S>>;
 
     /// Get a read-only view of this block's storage for a layer
-    fn layer_view(&self, layer_idx: usize, outer_idx: usize) -> BlockResult<view::LayerView<S>> {
+    fn layer_view(
+        &self,
+        layer_idx: usize,
+        outer_idx: usize,
+    ) -> BlockResult<view::LayerView<'_, S>> {
         match self.is_local() {
             Some(views) => views.local_layer_view(layer_idx, outer_idx),
             None => Err(BlockError::ViewsNotAvailableOnLogicalBlocks),
@@ -58,7 +62,7 @@ pub trait BlockDataExt<S: Storage>: Send + Sync + 'static + std::fmt::Debug {
         &mut self,
         layer_idx: usize,
         outer_idx: usize,
-    ) -> BlockResult<view::LayerViewMut<S>> {
+    ) -> BlockResult<view::LayerViewMut<'_, S>> {
         match self.is_local_mut() {
             Some(views) => views.local_layer_view_mut(layer_idx, outer_idx),
             None => Err(BlockError::ViewsNotAvailableOnLogicalBlocks),
@@ -66,7 +70,7 @@ pub trait BlockDataExt<S: Storage>: Send + Sync + 'static + std::fmt::Debug {
     }
 
     /// Get a read-only view of this block's storage
-    fn block_view(&self) -> BlockResult<view::BlockView<S>> {
+    fn block_view(&self) -> BlockResult<view::BlockView<'_, S>> {
         match self.is_local() {
             Some(views) => views.local_block_view(),
             None => Err(BlockError::ViewsNotAvailableOnLogicalBlocks),
@@ -74,7 +78,7 @@ pub trait BlockDataExt<S: Storage>: Send + Sync + 'static + std::fmt::Debug {
     }
 
     /// Get a mutable view of this block's storage
-    fn block_view_mut(&mut self) -> BlockResult<view::BlockViewMut<S>> {
+    fn block_view_mut(&mut self) -> BlockResult<view::BlockViewMut<'_, S>> {
         match self.is_local_mut() {
             Some(views) => views.local_block_view_mut(),
             None => Err(BlockError::ViewsNotAvailableOnLogicalBlocks),
@@ -88,20 +92,20 @@ pub trait BlockDataViews<S: Storage> {
         &self,
         layer_idx: usize,
         outer_idx: usize,
-    ) -> BlockResult<view::LayerView<S>>;
+    ) -> BlockResult<view::LayerView<'_, S>>;
 
     /// Get a mutable view of this block's storage for a layer
     fn local_layer_view_mut(
         &mut self,
         layer_idx: usize,
         outer_idx: usize,
-    ) -> BlockResult<view::LayerViewMut<S>>;
+    ) -> BlockResult<view::LayerViewMut<'_, S>>;
 
     /// Get a read-only view of this block's storage
-    fn local_block_view(&self) -> BlockResult<view::BlockView<S>>;
+    fn local_block_view(&self) -> BlockResult<view::BlockView<'_, S>>;
 
     /// Get a mutable view of this block's storage
-    fn local_block_view_mut(&mut self) -> BlockResult<view::BlockViewMut<S>>;
+    fn local_block_view_mut(&mut self) -> BlockResult<view::BlockViewMut<'_, S>>;
 }
 
 pub trait BlockDataProvider: StorageTypeProvider {

lib/llm/src/block_manager/block/data/local.rs

@@ -101,7 +101,7 @@ impl<S: Storage> BlockDataViews<S> for LocalBlockData<S> {
         &self,
         layer_idx: usize,
         outer_idx: usize,
-    ) -> BlockResult<view::LayerView<S>> {
+    ) -> BlockResult<view::LayerView<'_, S>> {
         let mr = self
             .layout
             .memory_region(self.block_idx, layer_idx, outer_idx)?;
@@ -113,14 +113,14 @@ impl<S: Storage> BlockDataViews<S> for LocalBlockData<S> {
         &mut self,
         layer_idx: usize,
         outer_idx: usize,
-    ) -> BlockResult<view::LayerViewMut<S>> {
+    ) -> BlockResult<view::LayerViewMut<'_, S>> {
         let mr = self
             .layout
             .memory_region(self.block_idx, layer_idx, outer_idx)?;
         unsafe { view::LayerViewMut::new(self, mr.addr(), mr.size(), mr.storage_type()) }
     }
 
-    fn local_block_view(&self) -> BlockResult<view::BlockView<S>> {
+    fn local_block_view(&self) -> BlockResult<view::BlockView<'_, S>> {
         if self.is_fully_contiguous() {
             let mr = self.layout.memory_region(self.block_idx, 0, 0)?;
             let offset = mr.addr();
@@ -134,7 +134,7 @@ impl<S: Storage> BlockDataViews<S> for LocalBlockData<S> {
         }
     }
 
-    fn local_block_view_mut(&mut self) -> BlockResult<view::BlockViewMut<S>> {
+    fn local_block_view_mut(&mut self) -> BlockResult<view::BlockViewMut<'_, S>> {
         if self.is_fully_contiguous() {
             let mr = self.layout.memory_region(self.block_idx, 0, 0)?;
             let offset = mr.addr();