misc(backend): allow rebinding numa core affinity

backends/llamacpp/csrc/backend.cpp

@@ -3,7 +3,6 @@
 //
 
 #include <filesystem>
-#include <span>
 
 #include <ggml.h>
 #include <llama.h>

backends/llamacpp/csrc/ffi.hpp

@@ -111,7 +111,7 @@ namespace huggingface::tgi::backends::llamacpp {
     struct numa_cpumask_deleter { void operator()(struct bitmask* cpumask){ numa_free_cpumask(cpumask); }};
     typedef std::unique_ptr<struct bitmask, numa_cpumask_deleter> unique_cpumask_ptr;
 
-    void set_numactl_core_affinity(rust::Slice<const size_t> affinity) {
+    void set_numa_core_affinity(rust::Slice<const size_t> affinity) {
 //    void set_numactl_core_affinity(std::vector<size_t> affinity) {
 #ifdef NUMA_AVAILABLE
         if(numa_available()) {
@@ -173,6 +173,14 @@ namespace huggingface::tgi::backends::llamacpp {
         SPDLOG_WARN("TGI's llama.cpp backend was compiled without NUMA support");
 #endif
     }
+
+    /**
+     * 
+     */
+    void update_numa_affinity() {
+        SPDLOG_INFO("Rebinding NUMA affinity for current worker on thread: {}", std::this_thread::get_id());
+        llama_numa_init(ggml_numa_strategy::GGML_NUMA_STRATEGY_NUMACTL);
+    }
 }
 
 

backends/llamacpp/src/backend.rs

@@ -1,6 +1,6 @@
 use crate::ffi::{
-    create_worker_frontend, set_numactl_core_affinity, GenerationParams, LlamaCppWorkerFrontend,
+    create_worker_frontend, set_numa_core_affinity, update_numa_affinity, GenerationParams,
-    SamplingParams,
+    LlamaCppWorkerFrontend, SamplingParams,
 };
 use async_channel::{unbounded as mpmc_unbounded, Receiver as MpmcReceiver, Sender as MpmcSender};
 use async_trait::async_trait;
@@ -8,7 +8,6 @@ use cxx::UniquePtr;
 use log::warn;
 use std::ops::Range;
 use std::path::{Path, PathBuf};
-use std::sync::mpsc::{channel, Receiver, Sender};
 use std::sync::Arc;
 use std::thread::spawn;
 use text_generation_router::infer::InferError::GenerationError;
@@ -25,17 +24,6 @@ use tokio::time::Instant;
 use tokio_stream::wrappers::UnboundedReceiverStream;
 use tracing::{debug, error, info};
 
-macro_rules! send_or_warn {
-    ($send: expr, $err: expr) => {
-        if let Err(se) = $send.send(err) {
-            warn!(
-                "Failed to send message back to the user: {}. Originating error: {}",
-                se, e
-            );
-        }
-    };
-}
-
 fn get_num_cores() -> usize {
     match option_env!("TGI_USE_PHYSICAL_CORES")
         .unwrap_or("OFF")
@@ -272,8 +260,9 @@ fn worker_loop(
     // This loop will mostly decode single token at every step, so no need to rely on parallelism
     tokenizers::utils::parallelism::set_parallelism(false);
 
-    // Bind cores for the current thread
+    // Bind cores for the current thread and make sure it's taken into account
-    set_numactl_core_affinity(&affinity);
+    set_numa_core_affinity(&affinity);
+    update_numa_affinity();
 
     loop {
         if let Ok((generation, stream)) = backlog.recv_blocking() {

backends/llamacpp/src/lib.rs

@@ -54,7 +54,8 @@ mod ffi {
             num_threads: u32,
         ) -> Result<UniquePtr<LlamaCppWorkerFrontend>>;
 
-        fn set_numactl_core_affinity(affinity: &[usize]);
+        fn set_numa_core_affinity(affinity: &[usize]);
+        fn update_numa_affinity();
 
         unsafe fn stream(
             self: Pin<&mut LlamaCppWorkerFrontend>,