fix(sglang): use incremental streaming output for completions

components/src/dynamo/sglang/_compat.py

@@ -18,6 +18,7 @@
 import ipaddress
 import logging
 import socket
+from typing import Any
 
 logger = logging.getLogger(__name__)
 
@@ -98,8 +99,42 @@ def to_tcp(self) -> str:
             return f"tcp://{self.host}:{self.port}"
 
 
+def enable_disjoint_streaming_output(server_args: Any) -> None:
+    """
+    Enable SGLang's disjoint streaming output across ServerArgs field renames.
+
+    Covers sglang <= 0.5.x (`stream_output`) and newer releases
+    (`incremental_streaming_output`).
+    """
+    fields = getattr(type(server_args), "__dataclass_fields__", None)
+    if isinstance(fields, dict):
+        if "incremental_streaming_output" in fields:
+            server_args.incremental_streaming_output = True
+            return
+        if "stream_output" in fields:
+            server_args.stream_output = True
+            return
+        raise AttributeError(
+            "SGLang ServerArgs has neither 'incremental_streaming_output' nor "
+            "'stream_output'"
+        )
+
+    if hasattr(server_args, "incremental_streaming_output"):
+        server_args.incremental_streaming_output = True
+        return
+    if hasattr(server_args, "stream_output"):
+        server_args.stream_output = True
+        return
+
+    logger.debug(
+        "Skipping streaming output compatibility for non-ServerArgs object: %s",
+        type(server_args).__name__,
+    )
+
+
 __all__ = [
     "NetworkAddress",
+    "enable_disjoint_streaming_output",
     "get_local_ip_auto",
     "get_zmq_socket",
     "_SGLANG_HAS_NETWORK_MODULE",

components/src/dynamo/sglang/args.py

@@ -24,6 +24,7 @@
 from dynamo.common.utils.runtime import parse_endpoint
 from dynamo.llm import fetch_model
 from dynamo.runtime.logging import configure_dynamo_logging
+from dynamo.sglang._compat import enable_disjoint_streaming_output
 from dynamo.sglang.backend_args import DynamoSGLangArgGroup, DynamoSGLangConfig
 
 configure_dynamo_logging()
@@ -374,12 +375,10 @@ async def parse_args(args: list[str]) -> Config:
         )
 
     # Dynamo's streaming handlers expect disjoint output_ids from SGLang (only new
-    # tokens since last output), not cumulative tokens.
-    # sglang renamed stream_output -> incremental_streaming_output in PR #20614.
-    if hasattr(ServerArgs, "incremental_streaming_output"):
-        server_args.incremental_streaming_output = True
-    else:
-        server_args.stream_output = True
+    # tokens since last output), not cumulative tokens. Modern SGLang gates this
+    # behavior behind incremental_streaming_output, while older releases used
+    # stream_output.
+    enable_disjoint_streaming_output(server_args)
 
     if dynamo_config.use_sglang_tokenizer:
         warnings.warn(

lib/llm/src/protocols/openai/completions/delta.rs

@@ -284,6 +284,11 @@ impl crate::protocols::openai::DeltaGeneratorExt<NvCreateCompletionResponse> for
             // Update prompt_tokens from worker if provided (e.g., for embeddings)
             self.usage.prompt_tokens = completion_usage.prompt_tokens;
 
+            // Propagate completion token details if provided
+            if let Some(completion_details) = completion_usage.completion_tokens_details.as_ref() {
+                self.usage.completion_tokens_details = Some(completion_details.clone());
+            }
+
             // Propagate prompt token details if provided
             if let Some(prompt_details) = completion_usage.prompt_tokens_details.as_ref() {
                 self.usage.prompt_tokens_details = Some(prompt_details.clone());