[BugFix][API] Handle wrapped reasoning parsers in usage accounting (vllm-project#10645)

### What this PR does / why we need it?

Fixes vllm-project#10628.

`OpenAIServingChat` usage accounting is a common API-server path, but
the previous backport lived entirely in
`patch_minimax_usage_accounting.py`. When other parser combinations
failed, the stack trace pointed at the MiniMax patch even though the
failing object was a generic vLLM parser-manager wrapper.

This PR splits that ownership:

- `patch_chat_usage_accounting.py` owns the generic OpenAI chat usage
wrapper and `UsageInfo` extension.
- `patch_minimax_usage_accounting.py` only patches MiniMax-M2 reasoning
parsers to provide MiniMax-specific reasoning-token counts.

The generic usage wrapper now unwraps parser-manager objects through
`reasoning_parser` before calling `count_reasoning_tokens()`. If the
resulting parser still has no token counter, usage accounting skips
reasoning-token details instead of raising a 500. MiniMax concrete and
unified parser paths continue to count reasoning tokens.

- vLLM version: v0.22.1
- vLLM main:
vllm-project/vllm@967c5c3

Signed-off-by: QwertyJack <7554089+QwertyJack@users.noreply.github.com>
Co-authored-by: QwertyJack <7554089+QwertyJack@users.noreply.github.com>

Author: QwertyJack

tests/ut/patch/platform/test_patch_minimax_usage_accounting.py

@@ -4,21 +4,25 @@
 from types import SimpleNamespace
 
 import pytest
+from vllm.parser.parser_manager import ParserManager
 from vllm.reasoning.minimax_m2_reasoning_parser import (
     MiniMaxM2AppendThinkReasoningParser,
     MiniMaxM2ReasoningParser,
 )
 
 from vllm_ascend.patch.platform import (
-    patch_minimax_usage_accounting as minimax_usage_patch,
+    patch_chat_usage_accounting as usage_patch,
 )
+from vllm_ascend.patch.platform import patch_minimax_usage_accounting  # noqa: F401
 
 
 class FakeTokenizer:
     def get_vocab(self):
         return {
             "<think>": 1,
             "</think>": 2,
+            "<minimax:tool_call>": 3,
+            "</minimax:tool_call>": 4,
         }
 
 
@@ -74,7 +78,7 @@ def test_count_reasoning_tokens(
 
 
 def test_update_usage_tracking_state_tracks_prompt_and_completion_tokens():
-    state = minimax_usage_patch._create_usage_tracking_state(
+    state = usage_patch._create_usage_tracking_state(
         num_choices=2,
         reasoning_parser=None,
     )
@@ -89,7 +93,7 @@ def test_update_usage_tracking_state_tracks_prompt_and_completion_tokens():
         ],
     )
 
-    minimax_usage_patch._update_usage_tracking_state(state, res)
+    usage_patch._update_usage_tracking_state(state, res)
 
     assert state.num_prompt_tokens == 3
     assert state.num_cached_tokens == 4
@@ -99,7 +103,7 @@ def test_update_usage_tracking_state_tracks_prompt_and_completion_tokens():
 
 def test_make_usage_info_injects_reasoning_token_details():
     fake_serving = SimpleNamespace(enable_prompt_tokens_details=True)
-    usage = minimax_usage_patch._make_usage_info(
+    usage = usage_patch._make_usage_info(
         fake_serving,
         prompt_tokens=3,
         completion_tokens=4,
@@ -115,7 +119,7 @@ def test_make_usage_info_injects_reasoning_token_details():
 
 def test_make_usage_info_injects_zero_cached_tokens():
     fake_serving = SimpleNamespace(enable_prompt_tokens_details=True)
-    usage = minimax_usage_patch._make_usage_info(
+    usage = usage_patch._make_usage_info(
         fake_serving,
         prompt_tokens=3,
         completion_tokens=4,
@@ -132,13 +136,13 @@ class FakeServing:
         enable_prompt_tokens_details = False
 
         def _make_usage_info(self, **kwargs):
-            return minimax_usage_patch._make_usage_info(self, **kwargs)
+            return usage_patch._make_usage_info(self, **kwargs)
 
     class FakeReasoningParser:
         def count_reasoning_tokens(self, token_ids):
             return 2 if 2 in token_ids else 0
 
-    state = minimax_usage_patch._create_usage_tracking_state(
+    state = usage_patch._create_usage_tracking_state(
         num_choices=2,
         reasoning_parser=FakeReasoningParser(),
     )
@@ -148,7 +152,7 @@ def count_reasoning_tokens(self, token_ids):
     state.completion_tokens = [4, 2]
     state.raw_output_token_ids = [[10, 11, 2, 20], [30, 31]]
 
-    usage = minimax_usage_patch._make_full_response_usage(FakeServing(), state)
+    usage = usage_patch._make_full_response_usage(FakeServing(), state)
 
     assert usage.prompt_tokens == 3
     assert usage.completion_tokens == 6
@@ -157,12 +161,63 @@ def count_reasoning_tokens(self, token_ids):
     assert usage.prompt_tokens_details is None
 
 
+def test_make_full_response_usage_accepts_wrapped_reasoning_parser():
+    class FakeServing:
+        enable_prompt_tokens_details = False
+
+        def _make_usage_info(self, **kwargs):
+            return usage_patch._make_usage_info(self, **kwargs)
+
+    class FakeReasoningParser:
+        def count_reasoning_tokens(self, token_ids):
+            return token_ids.index(2) if 2 in token_ids else len(token_ids)
+
+    state = usage_patch._create_usage_tracking_state(
+        num_choices=1,
+        reasoning_parser=SimpleNamespace(reasoning_parser=FakeReasoningParser()),
+    )
+    state.num_prompt_tokens = 3
+    state.final_res = SimpleNamespace(num_cached_tokens=None)
+    state.completion_tokens = [4]
+    state.raw_output_token_ids = [[10, 11, 2, 20]]
+
+    usage = usage_patch._make_full_response_usage(FakeServing(), state)
+
+    assert usage.completion_tokens_details.reasoning_tokens == 2
+
+
+def test_count_reasoning_tokens_accepts_minimax_unified_parser():
+    parser_cls = ParserManager.get_parser(
+        tool_parser_name="minimax_m2",
+        reasoning_parser_name="minimax_m2",
+        enable_auto_tools=True,
+        model_name="MiniMax-M2",
+    )
+    parser = parser_cls(FakeTokenizer(), tools=[])
+
+    assert not hasattr(parser, "count_reasoning_tokens")
+    assert usage_patch._count_reasoning_tokens_for_usage([10, 11, 2, 20], parser) == 2
+
+
+def test_count_reasoning_tokens_accepts_deepseek_parser_manager_wrapper():
+    parser_cls = ParserManager.get_parser(
+        tool_parser_name="deepseek_v4",
+        reasoning_parser_name="deepseek_v4",
+        enable_auto_tools=True,
+        model_name="DeepSeek-V4",
+    )
+    parser = parser_cls(FakeTokenizer(), tools=[])
+
+    assert not hasattr(parser, "count_reasoning_tokens")
+    assert usage_patch._count_reasoning_tokens_for_usage([10, 11], parser) == 0
+
+
 def test_stream_usage_details_are_injected_without_replacing_source():
     class FakeReasoningParser:
         def count_reasoning_tokens(self, token_ids):
             return token_ids.index(2) if 2 in token_ids else len(token_ids)
 
-    state = minimax_usage_patch._create_usage_tracking_state(
+    state = usage_patch._create_usage_tracking_state(
         num_choices=1,
         reasoning_parser=FakeReasoningParser(),
         enable_prompt_tokens_details=True,
@@ -181,7 +236,7 @@ def count_reasoning_tokens(self, token_ids):
         },
     }
 
-    data = minimax_usage_patch._inject_stream_usage_details(
+    data = usage_patch._inject_stream_usage_details(
         f"data: {json.dumps(chunk)}\n\n",
         state,
     )
@@ -193,12 +248,12 @@ def count_reasoning_tokens(self, token_ids):
     assert payload["usage"]["prompt_tokens_details"] == {
         "cached_tokens": 0,
     }
-    assert not hasattr(minimax_usage_patch, "_extract_class_method_source")
-    assert not hasattr(minimax_usage_patch, "_patch_chat_completion_stream_generator")
+    assert not hasattr(usage_patch, "_extract_class_method_source")
+    assert not hasattr(usage_patch, "_patch_chat_completion_stream_generator")
 
 
 def test_stream_usage_details_inject_prompt_details_without_reasoning():
-    state = minimax_usage_patch._create_usage_tracking_state(
+    state = usage_patch._create_usage_tracking_state(
         num_choices=1,
         reasoning_parser=None,
         enable_prompt_tokens_details=True,
@@ -216,7 +271,7 @@ def test_stream_usage_details_inject_prompt_details_without_reasoning():
         },
     }
 
-    data = minimax_usage_patch._inject_stream_usage_details(
+    data = usage_patch._inject_stream_usage_details(
         f"data: {json.dumps(chunk)}\n\n",
         state,
     )

vllm_ascend/patch/__init__.py

@@ -144,25 +144,39 @@
 #       Drop the alias once upstream registry includes it or the checkpoint
 #       standardizes architecture strings.
 #
-# ** 7. File: platform/patch_minimax_usage_accounting.py**
+# ** 7. File: platform/patch_chat_usage_accounting.py**
 # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 #   1. `vllm.entrypoints.openai.chat_completion.serving.OpenAIServingChat`
-#      `vllm.reasoning.minimax_m2_reasoning_parser`
 #    Why:
-#       MiniMax-M2 reasoning usage accounting needs to report
+#       Chat usage accounting needs to report
 #       `completion_tokens_details.reasoning_tokens` for both streaming and
 #       non-streaming chat completions.
 #    How：
-#       Monkey-patch MiniMax reasoning token counters, extend `UsageInfo`, and
-#       update chat usage construction to count reasoning tokens from raw output
-#       token ids.
+#       Extend `UsageInfo` and update chat usage construction to count reasoning
+#       tokens from raw output token ids when the active reasoning parser
+#       supports token counting.
 #    Related PR (if no, explain why):
 #       https://github.com/vllm-project/vllm/pull/37955
 #    Future Plan:
 #       Remove this patch once the runtime vLLM version contains the upstream
-#       MiniMax usage-accounting fix.
+#       reasoning usage-accounting support.
 #
-# ** 7a. File: platform/patch_glm_tool_call_streaming.py**
+# ** 7a. File: platform/patch_minimax_usage_accounting.py**
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+#   1. `vllm.reasoning.minimax_m2_reasoning_parser`
+#    Why:
+#       MiniMax-M2 reasoning parsers need token-count support for the generic
+#       chat usage-accounting wrapper.
+#    How：
+#       Monkey-patch MiniMax reasoning token counters so usage accounting can
+#       derive reasoning-token counts from raw output token ids.
+#    Related PR (if no, explain why):
+#       https://github.com/vllm-project/vllm/pull/37955
+#    Future Plan:
+#       Remove this patch once the runtime vLLM version contains the upstream
+#       MiniMax reasoning-token counting fix.
+#
+# ** 7b. File: platform/patch_glm_tool_call_streaming.py**
 # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 #   1. `vllm.entrypoints.openai.chat_completion.serving.OpenAIServingChat`
 #    Why:
@@ -182,7 +196,7 @@
 #       Remove this patch once the supported vLLM version contains the upstream
 #       GLM tool-call final chunk fixes.
 #
-# ** 7b. File: platform/patch_glm47_tool_call_parser.py**
+# ** 7c. File: platform/patch_glm47_tool_call_parser.py**
 # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 #   1. `vllm.tool_parsers.glm47_moe_tool_parser.Glm47MoeModelToolParser`
 #    Why:
@@ -200,7 +214,7 @@
 #       Remove this patch once the supported vLLM version contains the upstream
 #       GLM47 inline zero-argument streaming parser fix.
 #
-# ** 7c. File: platform/patch_anthropic_system_message.py**
+# ** 7d. File: platform/patch_anthropic_system_message.py**
 # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 #   1. `vllm.entrypoints.anthropic.protocol.AnthropicMessage`
 #      `vllm.entrypoints.anthropic.serving.AnthropicServingMessages`

vllm_ascend/patch/platform/__init__.py

@@ -29,6 +29,7 @@
 else:
     import vllm_ascend.patch.platform.patch_mamba_config_310  # noqa
 import vllm_ascend.patch.platform.patch_minimax_m2_config  # noqa
+import vllm_ascend.patch.platform.patch_chat_usage_accounting  # noqa
 import vllm_ascend.patch.platform.patch_minimax_usage_accounting  # noqa
 import vllm_ascend.patch.platform.patch_glm_tool_call_streaming  # noqa
 import vllm_ascend.patch.platform.patch_glm47_tool_call_parser  # noqa