add more models

Signed-off-by: Roger Wang <hey@rogerw.io>

tests/reasoning/test_qwen3_reasoning_parser.py

@@ -11,28 +11,19 @@ parser_name = "qwen3"
 start_token = "<think>"
 end_token = "</think>"
 
-REASONING_MODEL_NAME = "Qwen/Qwen3-0.6B"
+REASONING_MODEL_NAMES = [
+    "Qwen/Qwen3-0.6B",
+    "Qwen/Qwen3.5-397B-A17B",
+    "Qwen/Qwen3-4B-Thinking-2507",
+]
 
 
-@pytest.fixture(scope="module")
-def qwen3_tokenizer():
-    return AutoTokenizer.from_pretrained(REASONING_MODEL_NAME)
+@pytest.fixture(scope="module", params=REASONING_MODEL_NAMES)
+def qwen3_tokenizer(request):
+    return AutoTokenizer.from_pretrained(request.param)
 
 
-# --- Qwen3 style: both <think> and </think> generated by model ---
-
-WITH_THINK = {
-    "output": "<think>This is a reasoning section</think>This is the rest",
-    "reasoning": "This is a reasoning section",
-    "content": "This is the rest",
-}
-WITH_THINK_STREAM = {
-    "output": "<think>This is a reasoning section</think>This is the rest",
-    "reasoning": "This is a reasoning section",
-    "content": "This is the rest",
-}
-
-# --- Qwen3.5 style: <think> in prompt, only </think> in output ---
+# --- <think> in prompt, only </think> in output (typical) ---
 
 WITHOUT_START_TOKEN = {
     "output": "This is a reasoning section</think>This is the rest",
@@ -50,6 +41,19 @@ WITHOUT_START_TOKEN_COMPLETE_REASONING = {
     "content": None,
 }
 
+# --- <think> present in output (old template / edge case) ---
+
+WITH_THINK = {
+    "output": "<think>This is a reasoning section</think>This is the rest",
+    "reasoning": "This is a reasoning section",
+    "content": "This is the rest",
+}
+WITH_THINK_STREAM = {
+    "output": "<think>This is a reasoning section</think>This is the rest",
+    "reasoning": "This is a reasoning section",
+    "content": "This is the rest",
+}
+
 # --- No think tokens at all (thinking disabled) ---
 
 WITHOUT_THINK = {
@@ -58,7 +62,7 @@ WITHOUT_THINK = {
     "content": "This is the rest",
 }
 # In streaming, the parser cannot distinguish "thinking disabled" from
-# "Qwen3.5 reasoning in progress" when no think tokens have appeared yet.
+# "reasoning in progress" when no think tokens have appeared yet.
 # It assumes reasoning. The serving layer handles the "thinking disabled"
 # case by checking prompt_is_reasoning_end_arr before calling the parser.
 WITHOUT_THINK_STREAM = {
@@ -92,16 +96,6 @@ ONLY_OPEN_TAG_STREAM = {
 }
 
 TEST_CASES = [
-    pytest.param(
-        False,
-        WITH_THINK,
-        id="with_think",
-    ),
-    pytest.param(
-        True,
-        WITH_THINK_STREAM,
-        id="with_think_stream",
-    ),
     pytest.param(
         False,
         WITHOUT_START_TOKEN,
@@ -122,6 +116,16 @@ TEST_CASES = [
         WITHOUT_START_TOKEN_COMPLETE_REASONING,
         id="without_start_token_complete_reasoning_stream",
     ),
+    pytest.param(
+        False,
+        WITH_THINK,
+        id="with_think",
+    ),
+    pytest.param(
+        True,
+        WITH_THINK_STREAM,
+        id="with_think_stream",
+    ),
     pytest.param(
         False,
         WITHOUT_THINK,