From d251178b57a5d327fb508f96d6909a4a173f8020 Mon Sep 17 00:00:00 2001
From: sarmientoF <fsarmientod@uni.pe>
Date: Fri, 19 Jun 2026 11:30:57 +0900
Subject: [PATCH] fix(utils): preserve stream_options for providers that omit
 it from supported params

Providers whose get_supported_openai_params override hardcodes a list without
stream_options (fireworks_ai, deepinfra, mistral, perplexity, gemini, xai and
~160 other chat configs) had stream_options silently dropped in
map_openai_params. The validation path already force-keeps it
(_check_valid_arg special-cases stream_options/stream/user), but the mapping
path never did, so strict providers that gate the trailing usage chunk on
stream_options.include_usage stopped reporting cached/prompt token usage.

Re-inject stream_options after mapping, mirroring the existing validation
special-case. Explicit drops via additional_drop_params are still honored
because the param is removed from non_default_params upstream.
---
 litellm/utils.py                 |  5 ++++
 tests/test_litellm/test_utils.py | 47 ++++++++++++++++++++++++++++++++
 2 files changed, 52 insertions(+)

diff --git a/litellm/utils.py b/litellm/utils.py
index c9001e7d906..93120dcef41 100644
--- a/litellm/utils.py
+++ b/litellm/utils.py
@@ -4935,6 +4935,11 @@ def get_optional_params(
         openai_params=list(DEFAULT_CHAT_COMPLETION_PARAM_VALUES.keys()),
         additional_drop_params=additional_drop_params,
     )
+    if (
+        "stream_options" not in optional_params
+        and non_default_params.get("stream_options") is not None
+    ):
+        optional_params["stream_options"] = non_default_params["stream_options"]
     print_verbose(f"Final returned optional params: {optional_params}")
     optional_params = _apply_openai_param_overrides(
         optional_params=optional_params,
diff --git a/tests/test_litellm/test_utils.py b/tests/test_litellm/test_utils.py
index 44e0b55ee3b..83b7ee29d6f 100644
--- a/tests/test_litellm/test_utils.py
+++ b/tests/test_litellm/test_utils.py
@@ -3866,6 +3866,53 @@ class TestDropParamsWithPromptCacheKey:
         assert result.get("temperature") == 0.7
 
 
+class TestStreamOptionsPreserved:
+    """
+    stream_options must survive get_optional_params for every provider, even ones
+    whose get_supported_openai_params override omits it (fireworks_ai, deepinfra,
+    mistral, perplexity, gemini, ...). Before this fix map_openai_params silently
+    dropped it, so providers that gate the trailing usage chunk on
+    stream_options.include_usage never reported cached/prompt token usage.
+    """
+
+    @pytest.mark.parametrize(
+        "model, custom_llm_provider",
+        [
+            ("accounts/fireworks/models/llama-v3p1-8b-instruct", "fireworks_ai"),
+            ("deepinfra/meta-llama/Meta-Llama-3.1-8B-Instruct", "deepinfra"),
+            ("mistral/mistral-small-latest", "mistral"),
+            ("perplexity/sonar", "perplexity"),
+            ("gemini/gemini-2.5-flash", "gemini"),
+            ("xai/grok-4", "xai"),
+            ("gpt-4o", "openai"),
+        ],
+    )
+    def test_stream_options_kept(self, model, custom_llm_provider):
+        from litellm.utils import get_optional_params
+
+        result = get_optional_params(
+            model=model,
+            custom_llm_provider=custom_llm_provider,
+            stream=True,
+            stream_options={"include_usage": True},
+        )
+
+        assert result.get("stream_options") == {"include_usage": True}
+
+    def test_explicit_drop_still_respected(self):
+        from litellm.utils import get_optional_params
+
+        result = get_optional_params(
+            model="accounts/fireworks/models/llama-v3p1-8b-instruct",
+            custom_llm_provider="fireworks_ai",
+            stream=True,
+            stream_options={"include_usage": True},
+            additional_drop_params=["stream_options"],
+        )
+
+        assert "stream_options" not in result
+
+
 class TestGetOptionalParamsDeepSeek:
     """Tests that deepseek provider uses DeepSeekChatConfig for parameter mapping."""