Merge b84eff2c1f into e26a6450c8

litellm/llms/deepseek/chat/transformation.py

@@ -51,13 +51,30 @@ class DeepSeekChatConfig(OpenAIGPTConfig):
         thinking_value: Final = optional_params.pop("thinking", None)
         reasoning_effort: Final = optional_params.pop("reasoning_effort", None)
 
+        thinking_enabled = False
         # Handle thinking parameter - accept both enabled and disabled, ignore budget_tokens
         if isinstance(thinking_value, dict) and thinking_value.get("type") in ("enabled", "disabled"):
             optional_params["thinking"] = {"type": thinking_value["type"]}
+            thinking_enabled = thinking_value["type"] == "enabled"
 
         # Otherwise fall back to reasoning_effort: "none" disables, anything else enables
         elif reasoning_effort is not None:
             optional_params["thinking"] = {"type": "disabled" if reasoning_effort == "none" else "enabled"}
+            thinking_enabled = reasoning_effort != "none"
+        else:
+            thinking_enabled = True  # DeepSeek thinking mode is on by default
+
+        # reasoning_effort is a real DeepSeek top-level param (default "high"; valid
+        # values low/medium/high/xhigh/max). The branch above only used it as an on/off
+        # toggle and DROPPED the value, so DeepSeek always ran default HIGH effort — which
+        # can run away, exhausting max_tokens during reasoning and returning empty content
+        # (finish_reason="length", content:null). Preserve the value into the request body
+        # when thinking is on so callers can bound effort (e.g. reasoning_effort="low").
+        # Exclude the "none" sentinel: it is an OpenAI-style thinking-OFF switch, not a real
+        # effort value, so passing it alongside explicit thinking:enabled would be
+        # contradictory. See https://api-docs.deepseek.com/guides/thinking_mode
+        if thinking_enabled and reasoning_effort not in (None, "none"):
+            optional_params["reasoning_effort"] = reasoning_effort
 
         return optional_params
 

tests/llm_translation/test_deepseek_reasoning_effort.py

@@ -0,0 +1,49 @@
+"""Verify DeepSeek `reasoning_effort` passes through the transformation.
+
+Covers the fix that preserves the effort value (low/medium/high/xhigh/max) into
+the request body when thinking is on, instead of mapping it to a binary thinking
+toggle and discarding it. See https://api-docs.deepseek.com/guides/thinking_mode
+"""
+from litellm.llms.deepseek.chat.transformation import DeepSeekChatConfig
+
+
+def _map(non_default_params, drop_params=True):
+    cfg = DeepSeekChatConfig()
+    return cfg.map_openai_params(non_default_params, {}, "deepseek/deepseek-v4-flash", drop_params)
+
+
+def test_reasoning_effort_low_passes_through():
+    out = _map({"reasoning_effort": "low"})
+    assert out["reasoning_effort"] == "low"
+    assert out["thinking"] == {"type": "enabled"}
+
+
+def test_reasoning_effort_max_passes_through():
+    out = _map({"reasoning_effort": "max"})
+    assert out["reasoning_effort"] == "max"
+    assert out["thinking"] == {"type": "enabled"}
+
+
+def test_reasoning_effort_none_disables_thinking():
+    out = _map({"reasoning_effort": "none"})
+    assert out["thinking"] == {"type": "disabled"}
+    assert "reasoning_effort" not in out
+
+
+def test_explicit_thinking_disabled_wins_and_drops_effort():
+    out = _map({"thinking": {"type": "disabled"}, "reasoning_effort": "low"})
+    assert out["thinking"] == {"type": "disabled"}
+    assert "reasoning_effort" not in out
+
+
+def test_no_reasoning_effort_leaves_no_key():
+    out = _map({})
+    assert out.get("reasoning_effort") is None
+
+
+def test_explicit_enabled_thinking_with_none_effort_not_contradictory():
+    # "none" is a thinking-OFF sentinel, not a real effort value: it must not be
+    # passed alongside explicit thinking:enabled (would contradict the provider).
+    out = _map({"thinking": {"type": "enabled"}, "reasoning_effort": "none"})
+    assert out["thinking"] == {"type": "enabled"}
+    assert "reasoning_effort" not in out