Rich debug info (#29)

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "openai_trtllm"
-version = "0.1.0"
+version = "0.2.1"
 edition = "2021"
 
 [dependencies]
@@ -23,7 +23,7 @@ tracing-opentelemetry = { version = "0.22" }
 tower-http = { version = "0.5", features = ["trace"] }
 opentelemetry = { version = "0.21.0", features = ["metrics", "logs"] }
 opentelemetry_sdk = { version = "0.21.1", features = ["rt-tokio", "logs"] }
-opentelemetry-otlp = { version = "0.14.0", features = ["tonic", "metrics", "logs"]  }
+opentelemetry-otlp = { version = "0.14.0", features = ["tonic", "metrics", "logs"] }
 opentelemetry-semantic-conventions = { version = "0.13.0" }
 axum-tracing-opentelemetry = "0.16.0"
 liquid = "0.26.4"

src/routes/chat.rs

@@ -97,6 +97,7 @@ async fn chat_completions_stream(
 
             let raw_content = infer_response.raw_output_contents[0].clone();
             let content = deserialize_bytes_tensor(raw_content)?.into_iter().collect::<String>();
+            tracing::debug!("deserialized triton infer response content: {:?}", content);
 
             if !content.is_empty() {
                 let response = ChatCompletionChunk {
@@ -182,6 +183,8 @@ async fn chat_completions(
 
         let raw_content = infer_response.raw_output_contents[0].clone();
         let content = deserialize_bytes_tensor(raw_content)?.into_iter().collect();
+        tracing::debug!("deserialized triton infer response content: {:?}", content);
+
         contents.push(content);
     }
 

src/routes/completions.rs

@@ -91,6 +91,7 @@ async fn completions_stream(
             let content = deserialize_bytes_tensor(raw_content)?
                 .into_iter()
                 .collect::<String>();
+            tracing::debug!("deserialized triton infer response content: {:?}", content);
 
             if !content.is_empty() {
                 let response = Completion {
@@ -166,6 +167,8 @@ async fn completions(
 
         let raw_content = infer_response.raw_output_contents[0].clone();
         let content = deserialize_bytes_tensor(raw_content)?.into_iter().collect();
+        tracing::debug!("deserialized triton infer response content: {:?}", content);
+
         contents.push(content);
     }