add demo for deepseek (#426)

crates/llm_gateway/src/stream_context.rs

@@ -87,9 +87,11 @@ impl StreamContext {
         ));
 
         debug!(
-            "request received: llm provider hint: {:?}, selected llm: {}",
-            self.get_http_request_header(ARCH_PROVIDER_HINT_HEADER),
-            self.llm_provider.as_ref().unwrap().name
+            "request received: llm provider hint: {}, selected llm: {}, model: {}",
+            self.get_http_request_header(ARCH_PROVIDER_HINT_HEADER)
+                .unwrap_or_default(),
+            self.llm_provider.as_ref().unwrap().name,
+            self.llm_provider.as_ref().unwrap().model
         );
     }
 
@@ -494,7 +496,7 @@ impl HttpContext for StreamContext {
             //HACK: add support for tokenizing mistral and other models
             //filed issue https://github.com/katanemo/arch/issues/222
             if !model.as_ref().unwrap().starts_with("gpt") {
-                warn!(
+                trace!(
                     "tiktoken_rs: unsupported model: {}, using gpt-4 to compute token count",
                     model.as_ref().unwrap()
                 );

crates/llm_gateway/tests/integration.rs

@@ -25,7 +25,7 @@ fn request_headers_expectations(module: &mut Tester, http_context: i32) {
             Some("x-arch-llm-provider-hint"),
         )
         .returning(None)
-        .expect_log(Some(LogLevel::Debug), Some("request received: llm provider hint: Some(\"default\"), selected llm: open-ai-gpt-4"))
+        .expect_log(Some(LogLevel::Debug), Some("request received: llm provider hint: default, selected llm: open-ai-gpt-4, model: gpt-4"))
         .expect_add_header_map_value(
             Some(MapType::HttpRequestHeaders),
             Some("x-arch-llm-provider"),

demos/use_cases/llm_routing/arch_config.yaml

@@ -29,5 +29,12 @@ llm_providers:
     provider_interface: mistral
     model: ministral-3b-latest
 
+  - name: deepseek
+    access_key: $DEEPSEEK_API_KEY
+    provider_interface: openai
+    model: deepseek-reasoner
+    endpoint: api.deepseek.com
+    protocol: https
+
 tracing:
   random_sampling: 100