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 14, 05:30 UTC
[Misc] Remove unused encoder-decoder error strings (#25374)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Sep 22, 2025 at 19:04 UTC
417a164af66a11978f11a9ba64d4d76b4cf3e72a
2 files changed
-63
vllm/attention/backends/utils.py
-5
@@ -18,11 +18,6 @@ from vllm.utils import async_tensor_h2d, make_tensor_with_pad
18
19
logger = init_logger(__name__)
20
21
-# Error string(s) for encoder/decoder
22
-# unsupported attention scenarios
23
-STR_NOT_IMPL_ENC_DEC_ROCM_HIP = ("ROCm/HIP is not currently supported "
24
- "with encoder/decoder models.")
25
-
21
PAD_SLOT_ID = -1
22
23
# Switch to numpy implementation of compute_slot_mapping
vllm/utils/__init__.py
-58
@@ -88,64 +88,6 @@ DEFAULT_MAX_NUM_BATCHED_TOKENS = 2048
88
POOLING_MODEL_MAX_NUM_BATCHED_TOKENS = 32768
89
MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS = 5120
90
91
-# Exception strings for non-implemented encoder/decoder scenarios
92
-
93
-# Reminder: Please update docs/features/compatibility_matrix.md
94
-# If the feature combo become valid
95
-
96
-STR_NOT_IMPL_ENC_DEC_SWA = \
97
- "Sliding window attention for encoder/decoder models " + \
98
- "is not currently supported."
99
-
100
-STR_NOT_IMPL_ENC_DEC_PREFIX_CACHE = \
101
- "Prefix caching for encoder/decoder models " + \
102
- "is not currently supported."
103
-
104
-STR_NOT_IMPL_ENC_DEC_CHUNKED_PREFILL = \
105
- "Chunked prefill for encoder/decoder models " + \
106
- "is not currently supported."
107
-
108
-STR_NOT_IMPL_ENC_DEC_LOGIT_SOFTCAP = (
109
- "Models with logits_soft_cap "
110
- "require FlashInfer backend, which is "
111
- "currently not supported for encoder/decoder "
112
- "models.")
113
-
114
-STR_NOT_IMPL_ENC_DEC_LORA = ("LoRA is not currently "
115
- "supported with encoder/decoder "
116
- "models.")
117
-
118
-STR_NOT_IMPL_ENC_DEC_PP = ("Pipeline parallelism is not "
119
- "currently supported with "
120
- "encoder/decoder models.")
121
-
122
-STR_NOT_IMPL_ENC_DEC_MM = ("Multimodal is not currently "
123
- "supported with encoder/decoder "
124
- "models.")
125
-
126
-STR_NOT_IMPL_ENC_DEC_SPEC_DEC = ("Speculative decoding is not "
127
- "currently supported with encoder/"
128
- "decoder models.")
129
-
130
-STR_NOT_IMPL_ENC_DEC_BACKEND = ("XFormers and Flash-Attention are the only "
131
- "backends currently supported with encoder/"
132
- "decoder models.")
133
-
134
-# Efficiently import all enc/dec error strings
135
-# rather than having to import all of the above
136
-STR_NOT_IMPL_ENC_DEC_ERR_STRS = {
137
- "STR_NOT_IMPL_ENC_DEC_SWA": STR_NOT_IMPL_ENC_DEC_SWA,
138
- "STR_NOT_IMPL_ENC_DEC_PREFIX_CACHE": STR_NOT_IMPL_ENC_DEC_PREFIX_CACHE,
139
- "STR_NOT_IMPL_ENC_DEC_CHUNKED_PREFILL":
140
- STR_NOT_IMPL_ENC_DEC_CHUNKED_PREFILL,
141
- "STR_NOT_IMPL_ENC_DEC_LOGIT_SOFTCAP": STR_NOT_IMPL_ENC_DEC_LOGIT_SOFTCAP,
142
- "STR_NOT_IMPL_ENC_DEC_LORA": STR_NOT_IMPL_ENC_DEC_LORA,
143
- "STR_NOT_IMPL_ENC_DEC_PP": STR_NOT_IMPL_ENC_DEC_PP,
144
- "STR_NOT_IMPL_ENC_DEC_MM": STR_NOT_IMPL_ENC_DEC_MM,
145
- "STR_NOT_IMPL_ENC_DEC_SPEC_DEC": STR_NOT_IMPL_ENC_DEC_SPEC_DEC,
146
- "STR_NOT_IMPL_ENC_DEC_BACKEND": STR_NOT_IMPL_ENC_DEC_BACKEND,
147
-}
148
-
91
# Constants related to forcing the attention backend selection
92
93
# String name of register which may be set in order to