fix(docker, predictor): copy libtorch to iota-tools docker image

crates/iota-core/Cargo.toml

@@ -50,16 +50,15 @@ serde.workspace = true
 serde_json.workspace = true
 serde_with.workspace = true
 static_assertions.workspace = true
+sysinfo = "0.33"
 tap.workspace = true
+tch = { version = "0.13", features = ["download-libtorch"], optional = true }
 tempfile.workspace = true
 thiserror.workspace = true
 tokio = { workspace = true, features = ["full", "tracing", "test-util"] }
 tokio-stream.workspace = true
 tracing.workspace = true
 twox-hash = "1.6"
-ordered-float = { version = "5.0", features = ["serde"] }
-tch = { version = "0.13", features = ["download-libtorch"], optional = true }
-sysinfo = "0.33"
 
 # internal dependencies
 consensus-config.workspace = true

crates/iota-core/src/congestion_tracker.rs

@@ -7,8 +7,10 @@ use std::{
     fs::OpenOptions,
     io::Write,
     path::PathBuf,
+    sync::Arc,
 };
 
+use iota_metrics::monitored_scope;
 use iota_types::{
     base_types::ObjectID,
     digests::TransactionDigest,
@@ -18,20 +20,18 @@ use iota_types::{
     transaction::{TransactionData, TransactionDataAPI},
 };
 use moka::{ops::compute::Op, sync::Cache};
+use prometheus::Registry;
 use serde::Deserialize;
+use tracing::info;
+
 #[cfg(feature = "gas-nn")]
 use crate::model_updater::{
-    build_cp_update_batch, build_train_tx_batch, InNodeModelUpdater, ObjectCheckpointStats,
-    ObjectSnapshot, RawTxItem, ModelUpdater, ModelReader,
+    InNodeModelUpdater, ModelReader, ModelUpdater, ObjectCheckpointStats, ObjectSnapshot,
+    RawTxItem, build_cp_update_batch, build_train_tx_batch,
 };
-use tracing::info;
-use iota_metrics::monitored_scope;
-use prometheus::Registry;
-use std::sync::Arc;
-use crate::gas_metrics::{GasMetrics, init_gas_metrics};
-
 use crate::{
-    execution_cache::TransactionCacheRead
+    execution_cache::TransactionCacheRead,
+    gas_metrics::{GasMetrics, init_gas_metrics},
 };
 
 /// Capacity of the congestion tracker's cache.
@@ -162,7 +162,8 @@ pub struct CongestionTracker {
 }
 
 impl CongestionTracker {
-    /// Compose and send model update and training batches for the given checkpoint.
+    /// Compose and send model update and training batches for the given
+    /// checkpoint.
     #[cfg(feature = "gas-nn")]
     fn inform_model(
         &self,
@@ -266,7 +267,8 @@ impl CongestionTracker {
         self.metrics.record_hw_sample("congestion.inform_model");
     }
 
-    /// Fallback when gas-nn is disabled: still record metrics, but skip NN work.
+    /// Fallback when gas-nn is disabled: still record metrics, but skip NN
+    /// work.
     #[cfg(not(feature = "gas-nn"))]
     fn inform_model(
         &self,
@@ -315,7 +317,9 @@ impl CongestionTracker {
         effects: &[TransactionEffects],
     ) {
         let _scope = monitored_scope("CongestionTracker::process_checkpoint_effects");
-        let h = self.metrics.latency_component("congestion.process_checkpoint_effects");
+        let h = self
+            .metrics
+            .latency_component("congestion.process_checkpoint_effects");
         let timer = h.start_timer();
         // Containers for checkpoint's congestion and clearing transactions data.
         let mut congestion_txs_data: Vec<TxData> = Vec::with_capacity(effects.len());
@@ -353,7 +357,7 @@ impl CongestionTracker {
                         .predict_for_tx(tx_data, this.reference_gas_price),
                 )
             }
-                .unwrap_or(self.reference_gas_price);
+            .unwrap_or(self.reference_gas_price);
 
             // Skip system transactions
             if gas_price == 1 {
@@ -413,8 +417,12 @@ impl CongestionTracker {
         );
         // Record touched objects histogram
         let mut touched: std::collections::HashSet<ObjectID> = std::collections::HashSet::new();
-        for tx in &congestion_txs_data { touched.extend(tx.objects.iter().cloned()); }
-        for tx in &clearing_txs_data { touched.extend(tx.objects.iter().cloned()); }
+        for tx in &congestion_txs_data {
+            touched.extend(tx.objects.iter().cloned());
+        }
+        for tx in &clearing_txs_data {
+            touched.extend(tx.objects.iter().cloned());
+        }
         self.metrics.touched_hist().observe(touched.len() as u64);
 
         drop(timer);
@@ -522,7 +530,10 @@ impl CongestionTracker {
     /// Non-blocking: reads from lock-free snapshot of per-object tips.
     #[cfg(feature = "gas-nn")]
     pub fn get_suggested_gas_price_with_nn(&self, transaction: &TransactionData) -> Option<u64> {
-        Some(self.model_reader.predict_for_tx(transaction, self.reference_gas_price))
+        Some(
+            self.model_reader
+                .predict_for_tx(transaction, self.reference_gas_price),
+        )
     }
 
     /// Returns a map of all objects and their hotness values.
@@ -568,7 +579,10 @@ impl CongestionTracker {
 
         // Write header if the file is new
         if !file_exists {
-            writeln!(file, "checkpoint,digest,gasprice,feedback,sui,ogd,nn,cleared")?;
+            writeln!(
+                file,
+                "checkpoint,digest,gasprice,feedback,sui,ogd,nn,cleared"
+            )?;
         }
 
         // Build row

crates/iota-core/src/gas_metrics.rs

@@ -1,8 +1,14 @@
+// Copyright (c) 2025 IOTA Stiftung
+// SPDX-License-Identifier: Apache-2.0
+
 use std::sync::{Arc, Mutex};
 
 use iota_metrics::histogram::{Histogram, HistogramVec};
 use once_cell::sync::OnceCell;
-use prometheus::{IntCounterVec, IntGauge, IntGaugeVec, Registry, register_int_counter_vec_with_registry, register_int_gauge_vec_with_registry, register_int_gauge_with_registry};
+use prometheus::{
+    IntCounterVec, IntGauge, IntGaugeVec, Registry, register_int_counter_vec_with_registry,
+    register_int_gauge_vec_with_registry, register_int_gauge_with_registry,
+};
 use sysinfo::{ProcessRefreshKind, ProcessesToUpdate, System, UpdateKind, get_current_pid};
 
 /// Percentiles to export for latency histograms (in 1/1000ths).
@@ -21,12 +27,14 @@ pub struct GasMetrics {
     /// Pending training items gauge.
     pub pending_train_items: IntGauge,
 
-    /// Batches accounting. Labels: type ("update"|"train"), status ("received"|"dropped").
+    /// Batches accounting. Labels: type ("update"|"train"), status
+    /// ("received"|"dropped").
     pub batches_total: IntCounterVec,
     /// Inference accounting. Label: status ("success"|"fallback"|"none").
     pub infer_total: IntCounterVec,
 
-    /// Per-operation process CPU usage (percent) as histogram; label: component.
+    /// Per-operation process CPU usage (percent) as histogram; label:
+    /// component.
     pub proc_cpu_pct: HistogramVec,
     /// Per-operation process RSS memory (bytes) as histogram; label: component.
     pub proc_mem_bytes: HistogramVec,
@@ -63,7 +71,8 @@ impl GasMetrics {
             "gas_predict_pending_train_items",
             "Pending training items count",
             registry,
-        ).unwrap();
+        )
+        .unwrap();
 
         let batches_total = register_int_counter_vec_with_registry!(
             "gas_predict_batches_total",
@@ -111,8 +120,9 @@ impl GasMetrics {
         }
     }
 
-    /// Pre-create time series for common components so that dashboards show lines even
-    /// when NN is disabled or idle. This sends a single zero sample per series.
+    /// Pre-create time series for common components so that dashboards show
+    /// lines even when NN is disabled or idle. This sends a single zero
+    /// sample per series.
     pub fn warmup_series(&self) {
         let components = [
             "model_updater.predict",
@@ -139,26 +149,30 @@ impl GasMetrics {
     }
 
     pub fn touched_hist(&self) -> Histogram {
-        self.touched_per_cp.with_label_values(&["congestion.touched"]) 
+        self.touched_per_cp
+            .with_label_values(&["congestion.touched"])
     }
 
-    /// Sample process CPU% and RSS memory and record to histograms for the given component.
+    /// Sample process CPU% and RSS memory and record to histograms for the
+    /// given component.
     pub fn record_hw_sample(&self, component: &str) {
         if let Some((cpu_pct, rss_bytes)) = self.sampler.sample_proc() {
             self.proc_cpu_pct
                 .with_label_values(&[component])
                 .observe(cpu_pct as u64);
             self.proc_mem_bytes
                 .with_label_values(&[component])
-                .observe(rss_bytes as u64);
+                .observe(rss_bytes);
         }
     }
 }
 
 static GLOBAL: OnceCell<Arc<GasMetrics>> = OnceCell::new();
 
 pub fn init_gas_metrics(registry: &Registry) -> Arc<GasMetrics> {
-    let m = GLOBAL.get_or_init(|| Arc::new(GasMetrics::new(registry))).clone();
+    let m = GLOBAL
+        .get_or_init(|| Arc::new(GasMetrics::new(registry)))
+        .clone();
     // warmup once per process
     m.warmup_series();
     m
@@ -193,7 +207,11 @@ impl GasHwSampler {
 
     fn sample_proc(&self) -> Option<(f32, u64)> {
         let mut sys = self.sys.lock().ok()?;
-        let _ = sys.refresh_processes_specifics(ProcessesToUpdate::Some(&[self.pid]), false, self.refresh_kind);
+        let _ = sys.refresh_processes_specifics(
+            ProcessesToUpdate::Some(&[self.pid]),
+            false,
+            self.refresh_kind,
+        );
         let p = sys.process(self.pid)?;
         let cpu = p.cpu_usage();
         let rss = p.memory();

crates/iota-core/src/lib.rs

@@ -22,13 +22,13 @@ pub mod epoch;
 pub mod execution_cache;
 mod execution_driver;
 mod fallback_fetch;
+pub mod gas_metrics;
 pub mod jsonrpc_index;
 pub mod metrics;
 pub mod mock_consensus;
 pub mod model;
 #[cfg(feature = "gas-nn")]
 pub mod model_updater;
-pub mod gas_metrics;
 pub mod module_cache_metrics;
 pub mod mysticeti_adapter;
 pub mod overload_monitor;

crates/iota-core/src/model.rs

@@ -1,3 +1,6 @@
+// Copyright (c) 2025 IOTA Stiftung
+// SPDX-License-Identifier: Apache-2.0
+
 //! Libtorch (tch) port of the hotness-aware gas predictor.
 //!
 //! Parity with Python model:
@@ -241,7 +244,7 @@ impl GasLearner {
     /// One-time warm-up to avoid first-call stalls (parity with Python
     /// startup).
     pub fn warmup(&self) {
-        let _ = tch::no_grad(|| {
+        tch::no_grad(|| {
             let dummy = Tensor::zeros([1, T, F], (Kind::Float, self.vs.device()));
             let (_pred, _attn) = self
                 .model