chore: allow frontend + mockers to run on macos (#6282)

Signed-off-by: PeaBrane <yanrpei@gmail.com>

lib/bindings/python/Cargo.toml

@@ -26,7 +26,6 @@ default = []
 media-ffmpeg = ["dynamo-llm/media-ffmpeg"]
 
 [dependencies]
-dynamo-llm = { path = "../../llm" }
 dynamo-runtime = { path = "../../runtime" }
 dynamo-parsers = { path = "../../parsers" }
 
@@ -74,5 +73,11 @@ cudarc = { version = "0.17.1", features = ["cuda-12020"], optional = true }
 prometheus = "0.14.0"
 
 
+[target.'cfg(target_os = "linux")'.dependencies]
+dynamo-llm = { path = "../../llm" }
+
+[target.'cfg(not(target_os = "linux"))'.dependencies]
+dynamo-llm = { path = "../../llm", default-features = false }
+
 [dev-dependencies]
 rstest = "0.25"

lib/bindings/python/build.rs

+// SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+//! Build script for dynamo-py3.
+//!
+//! On macOS, nixl-sys unconditionally links `-lstdc++` which doesn't exist
+//! (macOS uses libc++). We create an empty static archive to satisfy the
+//! linker since libc++ is already linked.
+
+fn main() {
+    #[cfg(target_os = "macos")]
+    {
+        let out_dir = std::env::var("OUT_DIR").unwrap();
+        let lib_path = format!("{}/libstdc++.a", out_dir);
+
+        // Write a minimal valid static archive (just the magic header).
+        // macOS `ar` refuses to create an empty archive, so write it directly.
+        std::fs::write(&lib_path, b"!<arch>\n").expect("failed to create empty libstdc++.a");
+
+        println!("cargo:rustc-link-search=native={}", out_dir);
+    }
+}

lib/llm/src/local_model.rs

@@ -17,7 +17,7 @@ use crate::entrypoint::RouterConfig;
 use crate::mocker::protocols::{MockEngineArgs, WorkerType};
 use crate::model_card::ModelDeploymentCard;
 use crate::model_type::{ModelInput, ModelType};
-use crate::preprocessor::media::{ImageDecoder, MediaDecoder, MediaFetcher};
+use crate::preprocessor::media::{MediaDecoder, MediaFetcher};
 use crate::request_template::RequestTemplate;
 
 pub mod runtime_config;
@@ -245,12 +245,6 @@ impl LocalModelBuilder {
             self.runtime_config.enable_local_indexer = mocker_engine_args.enable_local_indexer
                 && mocker_engine_args.worker_type != WorkerType::Decode;
             self.runtime_config.data_parallel_size = mocker_engine_args.dp_size;
-            self.media_decoder = Some(MediaDecoder {
-                image: Some(ImageDecoder::default()),
-                #[cfg(feature = "media-ffmpeg")]
-                video: None,
-            });
-            self.media_fetcher = Some(MediaFetcher::default());
 
             // Set bootstrap endpoint for prefill workers with bootstrap_port configured
             if mocker_engine_args.worker_type == WorkerType::Prefill

lib/mocker/src/perf_model.rs

@@ -204,7 +204,7 @@ impl PerfModel {
         };
         // Ensure non-negative timing
         let result = time.max(0.0);
-        tracing::debug!("Prefill time prediction: new_tokens={new_tokens}, time={result:.2}ms");
+        tracing::trace!("Prefill time prediction: new_tokens={new_tokens}, time={result:.2}ms");
         result
     }
 
@@ -229,7 +229,7 @@ impl PerfModel {
         };
         // Ensure non-negative timing
         let result = time.max(0.0);
-        tracing::debug!(
+        tracing::trace!(
             "Decode time prediction: active_kv_tokens={active_kv_tokens}, context_length={context_length}, time={result:.2}ms"
         );
         result