A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 9, 18:15 UTC
[BugFix] deepseek_v32_encoding: Replace asserts with proper exceptions (#32884)
Signed-off-by: RishabhSaini <rishabhsaini01@gmail.com>
Rishabh Saini committed
Jan 22, 2026 at 22:44 UTC
f61c9da711d846220cb59738d10c0a937ffb80e2
1 file changed
+39
-28
vllm/tokenizers/deepseek_v32_encoding.py
+39
-28
@@ -154,10 +154,12 @@ def find_last_user_index(messages: list[dict[str, Any]]) -> int:
154
def render_message(
155
index: int, messages: list[dict[str, Any]], thinking_mode: str
156
) -> str:
157
- assert 0 <= index < len(messages)
158
- assert thinking_mode in ["chat", "thinking"], (
159
- f"Invalid thinking_mode `{thinking_mode}`"
160
- )
157
+ if not (0 <= index < len(messages)):
158
+ raise ValueError(
159
+ f"Index {index} out of range for messages list of length {len(messages)}"
160
+ )
161
+ if thinking_mode not in ["chat", "thinking"]:
162
+ raise ValueError(f"Invalid thinking_mode `{thinking_mode}`")
163
164
prompt = ""
165
msg = messages[index]
@@ -187,7 +189,8 @@ def render_message(
189
)
190
191
elif role == "developer":
190
- assert content, f"Invalid message for role `{role}`: {msg}"
192
+ if not content:
193
+ raise ValueError(f"Invalid message for role `{role}`: {msg}")
194
content_developer = ""
195
if tools:
196
content_developer += "\n\n" + render_tools(tools)
@@ -220,17 +223,17 @@ def render_message(
223
prev_assistant_idx -= 1
224
assistant_msg = messages[prev_assistant_idx]
225
223
- assert (
226
+ if not (
227
index == 0
228
or prev_assistant_idx >= 0
229
and assistant_msg.get("role") == "assistant"
227
- ), f"Invalid messages at {index}:\n{assistant_msg}"
230
+ ):
231
+ raise ValueError(f"Invalid messages at {index}:\n{assistant_msg}")
232
233
tool_call_order = index - prev_assistant_idx
234
assistant_tool_calls = assistant_msg.get("tool_calls")
231
- assert assistant_tool_calls and len(assistant_tool_calls) >= tool_call_order, (
232
- "No tool calls but found tool output"
233
- )
235
+ if not (assistant_tool_calls and len(assistant_tool_calls) >= tool_call_order):
236
+ raise ValueError("No tool calls but found tool output")
237
238
if tool_call_order == 1:
239
prompt += "\n\n<function_results>"
@@ -266,9 +269,10 @@ def render_message(
269
summary_content = content or ""
270
271
if thinking_mode == "thinking" and index > last_user_idx:
269
- assert reasoning_content or tool_calls, (
270
- f"ThinkingMode: {thinking_mode}, invalid message without reasoning_content/tool_calls `{msg}` after last user message"
271
- )
272
+ if not (reasoning_content or tool_calls):
273
+ raise ValueError(
274
+ f"ThinkingMode: {thinking_mode}, invalid message without reasoning_content/tool_calls `{msg}` after last user message"
275
+ )
276
thinking_part = (
277
thinking_template.format(reasoning_content=reasoning_content or "")
278
+ thinking_end_token
@@ -362,12 +366,14 @@ def parse_tool_calls(index: int, text: str):
366
index, _, stop_token = _read_until_stop(
367
index, text, [f"<{dsml_token}invoke", tool_calls_end_token]
368
)
365
- assert _ == ">\n", "Tool call format error"
369
+ if _ != ">\n":
370
+ raise RuntimeError("Tool call format error")
371
372
if stop_token == tool_calls_end_token:
373
break
374
370
- assert stop_token is not None, "Missing special token"
375
+ if stop_token is None:
376
+ raise RuntimeError("Missing special token")
377
378
index, tool_name_content, stop_token = _read_until_stop(
379
index, text, [f"<{dsml_token}parameter", f"</{dsml_token}invoke"]
@@ -376,7 +382,8 @@ def parse_tool_calls(index: int, text: str):
382
p_tool_name = re.findall(
383
r'^\s*name="(.*?)">\n$', tool_name_content, flags=re.DOTALL
384
)
379
- assert len(p_tool_name) == 1, "Tool name format error"
385
+ if len(p_tool_name) != 1:
386
+ raise RuntimeError("Tool name format error")
387
tool_name = p_tool_name[0]
388
389
tool_args: dict[str, tuple[str, str]] = {}
@@ -390,16 +397,19 @@ def parse_tool_calls(index: int, text: str):
397
param_content,
398
flags=re.DOTALL,
399
)
393
- assert len(param_kv) == 1, "Parameter format error"
400
+ if len(param_kv) != 1:
401
+ raise RuntimeError("Parameter format error")
402
param_name, string, param_value = param_kv[0]
403
396
- assert param_name not in tool_args, "Duplicate parameter name"
404
+ if param_name in tool_args:
405
+ raise RuntimeError("Duplicate parameter name")
406
tool_args[param_name] = (param_value, string)
407
408
index, content, stop_token = _read_until_stop(
409
index, text, [f"<{dsml_token}parameter", f"</{dsml_token}invoke"]
410
)
402
- assert content == ">\n", "Parameter format error"
411
+ if content != ">\n":
412
+ raise RuntimeError("Parameter format error")
413
414
tool_call = decode_dsml_to_arguments(tool_name=tool_name, tool_args=tool_args)
415
tool_calls.append(tool_call)
@@ -422,7 +432,8 @@ def parse_message_from_completion_text(text: str, thinking_mode: str):
432
index, text, [thinking_end_token, tool_calls_start_token]
433
)
434
reasoning_content = content_delta
425
- assert stop_token == thinking_end_token, "Invalid thinking format"
435
+ if stop_token != thinking_end_token:
436
+ raise RuntimeError("Invalid thinking format")
437
438
index, content_delta, stop_token = _read_until_stop(
439
index, text, [eos_token, tool_calls_start_token]
@@ -431,17 +442,18 @@ def parse_message_from_completion_text(text: str, thinking_mode: str):
442
if stop_token == tool_calls_start_token:
443
is_tool_calling = True
444
else:
434
- assert stop_token == eos_token, "Invalid summary format"
445
+ if stop_token != eos_token:
446
+ raise RuntimeError("Invalid summary format")
447
448
if is_tool_calling:
449
index, stop_token, tool_calls = parse_tool_calls(index, text)
450
451
index, tool_ends_text, stop_token = _read_until_stop(index, text, [eos_token])
440
- assert not tool_ends_text, "Unexpected content after tool calls"
452
+ if tool_ends_text:
453
+ raise RuntimeError("Unexpected content after tool calls")
454
442
- assert len(text) == index and stop_token in [eos_token, None], (
443
- "Unexpected content at end"
444
- )
455
+ if not (len(text) == index and stop_token in [eos_token, None]):
456
+ raise RuntimeError("Unexpected content at end")
457
458
for sp_token in [
459
bos_token,
@@ -450,9 +462,8 @@ def parse_message_from_completion_text(text: str, thinking_mode: str):
462
thinking_end_token,
463
dsml_token,
464
]:
453
- assert sp_token not in summary_content and sp_token not in reasoning_content, (
454
- "Unexpected special token in content"
455
- )
465
+ if sp_token in summary_content or sp_token in reasoning_content:
466
+ raise RuntimeError("Unexpected special token in content")
467
468
return {
469
"role": "assistant",