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 13, 23:15 UTC
[Cleanup] Remove no-longer-used `SpeculativeConfig.enable_chunked_prefill` (#27826)
Signed-off-by: Nick Hill <nhill@redhat.com>
Nick Hill committed
Oct 31, 2025 at 10:57 UTC
9e5bd3076e0b2dc9336ac230428424351426e2ef
3 files changed
+1
-16
vllm/config/speculative.py
-10
@@ -78,10 +78,6 @@ class SpeculativeConfig:
78
draft_tensor_parallel_size: int | None = Field(default=None, ge=1)
79
"""The degree of the tensor parallelism for the draft model. Can only be 1
80
or the same as the target model's tensor parallel size."""
81
- disable_logprobs: bool = True
82
- """If set to True, token log probabilities are not returned during
83
- speculative decoding. If set to False, token log probabilities are returned
84
- according to the log probability settings in SamplingParams."""
81
82
# Draft model configuration
83
quantization: me_quant.QuantizationMethods | None = None
@@ -126,12 +122,6 @@ class SpeculativeConfig:
122
"""The configuration of the target model."""
123
target_parallel_config: SkipValidation[ParallelConfig] = None # type: ignore
124
"""The parallel configuration for the target model."""
129
- enable_chunked_prefill: SkipValidation[bool] = None # type: ignore
130
- """Whether vLLM is configured to use chunked prefill or not. Used for
131
- raising an error since it's not yet compatible with speculative decode."""
132
- disable_log_stats: SkipValidation[bool] = None # type: ignore
133
- """Whether to disable the periodic printing of stage times in speculative
134
- decoding."""
125
126
# params generated in the post-init stage
127
draft_model_config: SkipValidation[ModelConfig] = None # type: ignore
vllm/engine/arg_utils.py
-6
@@ -1246,8 +1246,6 @@ class EngineArgs:
1246
self,
1247
target_model_config: ModelConfig,
1248
target_parallel_config: ParallelConfig,
1249
- enable_chunked_prefill: bool,
1250
- disable_log_stats: bool,
1249
) -> SpeculativeConfig | None:
1250
"""Initializes and returns a SpeculativeConfig object based on
1251
`speculative_config`.
@@ -1267,8 +1265,6 @@ class EngineArgs:
1265
{
1266
"target_model_config": target_model_config,
1267
"target_parallel_config": target_parallel_config,
1270
- "enable_chunked_prefill": enable_chunked_prefill,
1271
- "disable_log_stats": disable_log_stats,
1268
}
1269
)
1270
return SpeculativeConfig(**self.speculative_config)
@@ -1561,8 +1557,6 @@ class EngineArgs:
1557
speculative_config = self.create_speculative_config(
1558
target_model_config=model_config,
1559
target_parallel_config=parallel_config,
1564
- enable_chunked_prefill=self.enable_chunked_prefill,
1565
- disable_log_stats=self.disable_log_stats,
1560
)
1561
1562
# make sure num_lookahead_slots is set appropriately depending on
vllm/entrypoints/openai/api_server.py
+1
@@ -241,6 +241,7 @@ async def build_async_engine_client_from_engine_args(
241
)
242
243
# Don't keep the dummy data in memory
244
+ assert async_llm is not None
245
await async_llm.reset_mm_cache()
246
247
yield async_llm