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 7, 05:15 UTC
[Misc] Remove deprecated items that are due for removal (#36006)
Signed-off-by: Martin Hickey <martin.hickey@ie.ibm.com>
Martin Hickey committed
Mar 5, 2026 at 06:14 UTC
c3598d02fa638119ae4ac933850dbcd3d629fa1c
3 files changed
-51
vllm/config/cache.py
-18
@@ -92,24 +92,6 @@ class CacheConfig:
92
benefits before turning this on.\n
93
- "xxhash_cbor" combines canonical CBOR serialization with xxHash for
94
reproducible hashing. Requires the optional ``xxhash`` package."""
95
- cpu_offload_gb: float = Field(default=0, ge=0)
96
- """The space in GiB to offload to CPU, per GPU. Default is 0, which means
97
- no offloading. Intuitively, this argument can be seen as a virtual way to
98
- increase the GPU memory size. For example, if you have one 24 GB GPU and
99
- set this to 10, virtually you can think of it as a 34 GB GPU. Then you can
100
- load a 13B model with BF16 weight, which requires at least 26GB GPU memory.
101
- Note that this requires fast CPU-GPU interconnect, as part of the model is
102
- loaded from CPU memory to GPU memory on the fly in each model forward pass.
103
-
104
- DEPRECATED: This field is deprecated and will be removed in v0.16.
105
- Please use OffloadConfig.uva.cpu_offload_gb instead.
106
- """
107
- cpu_offload_params: set[str] = Field(default_factory=set)
108
- """The set of parameter name segments to target for CPU offloading.
109
-
110
- DEPRECATED: This field is deprecated and will be removed in v0.16.
111
- Please use OffloadConfig.uva.cpu_offload_params instead.
112
- """
95
calculate_kv_scales: bool = False
96
"""This enables dynamic calculation of `k_scale` and `v_scale` when
97
kv_cache_dtype is fp8. If `False`, the scales will be loaded from the model
vllm/config/compilation.py
-18
@@ -381,13 +381,6 @@ class CompilationConfig:
381
certain small batchsizes, where inductor is good at optimizing.
382
"""
383
384
- # Top-level Compilation control
385
- level: int = Field(default=None)
386
- """
387
- Level is deprecated and will be removed in the next release,
388
- either 0.12.0 or 0.11.2 whichever is soonest.
389
- Please use mode. Currently all levels are mapped to mode.
390
- """
384
# Top-level Compilation control
385
mode: CompilationMode = Field(default=None)
386
"""The compilation approach used for torch.compile-based compilation of the
@@ -801,17 +794,6 @@ class CompilationConfig:
794
return handler(value)
795
796
def __post_init__(self) -> None:
804
- if self.level is not None:
805
- logger.warning(
806
- "Level is deprecated and will be removed in the next release,"
807
- "either 0.12.0 or 0.11.2 whichever is soonest."
808
- "Use mode instead."
809
- "If both level and mode are given,"
810
- "only mode will be used."
811
- )
812
- if self.mode is None:
813
- self.mode = self.level
814
-
797
count_none = self.custom_ops.count("none")
798
count_all = self.custom_ops.count("all")
799
assert count_none + count_all <= 1, "Can only specify 'none' or 'all'"
vllm/multimodal/processing/processor.py
-15
@@ -1074,21 +1074,6 @@ class BaseMultiModalProcessor(ABC, Generic[_I]):
1074
mm_items.get_all_counts(),
1075
)
1076
1077
- for modality, prompt_updates in mm_prompt_updates.items():
1078
- for item_idx, item_prompt_updates in enumerate(prompt_updates):
1079
- if len(item_prompt_updates) > 1:
1080
- logger.warning_once(
1081
- "Detected %d prompt updates for `mm_items[%r][%s]`. "
1082
- "Multiple prompt updates per item is now "
1083
- "deprecated and may be removed in v0.13. "
1084
- "Instead, please specify dynamic update targets "
1085
- "in the same prompt update definition by passing "
1086
- "a function to `PromptUpdate.target`.",
1087
- len(prompt_updates),
1088
- modality,
1089
- item_idx,
1090
- )
1091
-
1077
return mm_prompt_updates
1078
1079
def _find_mm_placeholders(