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 10, 14:00 UTC
[Bugfix]: prevent leaking tokens in crash log (#30751)
Signed-off-by: Marko Rosenmueller <5467316+dr75@users.noreply.github.com>
Marko Rosenmueller committed
Jan 7, 2026 at 17:15 UTC
1d9e9ae8a4498782de0dd51627ab1fddac4692ef
1 file changed
+25
vllm/v1/core/sched/output.py
+25
@@ -90,10 +90,14 @@ class NewRequestData:
90
prompt_embeds_shape = (
91
self.prompt_embeds.shape if self.prompt_embeds is not None else None
92
)
93
+ prefill_token_ids_len = (
94
+ len(self.prefill_token_ids) if self.prefill_token_ids is not None else None
95
+ )
96
return (
97
f"NewRequestData("
98
f"req_id={self.req_id},"
99
f"prompt_token_ids_len={prompt_token_ids_len},"
100
+ f"prefill_token_ids_len={prefill_token_ids_len},"
101
f"mm_features={self.mm_features},"
102
f"sampling_params={self.sampling_params},"
103
f"block_ids={self.block_ids},"
@@ -122,6 +126,27 @@ class CachedRequestData:
126
num_computed_tokens: list[int]
127
num_output_tokens: list[int]
128
129
+ # Version of dataclass repr with token IDs obfuscated.
130
+ def anon_repr(self) -> str:
131
+ new_token_ids_lens = [len(toks) for toks in self.new_token_ids]
132
+ all_token_ids_lens = {
133
+ req_id: len(toks) for req_id, toks in self.all_token_ids.items()
134
+ }
135
+ return (
136
+ f"CachedRequestData("
137
+ f"req_ids={self.req_ids},"
138
+ f"resumed_req_ids={self.resumed_req_ids},"
139
+ f"new_token_ids_lens={new_token_ids_lens},"
140
+ f"all_token_ids_lens={all_token_ids_lens},"
141
+ f"new_block_ids={self.new_block_ids},"
142
+ f"num_computed_tokens={self.num_computed_tokens},"
143
+ f"num_output_tokens={self.num_output_tokens}"
144
+ f")"
145
+ )
146
+
147
+ def __repr__(self) -> str:
148
+ return self.anon_repr()
149
+
150
@property
151
def num_reqs(self) -> int:
152
return len(self.req_ids)