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, 22:45 UTC
[Model Runner V2][Minor] Remove redundant `do_spec_decode` field (#35039)
Signed-off-by: Nick Hill <nickhill123@gmail.com> Co-authored-by: Woosuk Kwon <woosuk@inferact.ai>
Nick Hill committed
Feb 22, 2026 at 16:18 UTC
944ffb59680c0210ec54ddb43a3c7ef015e1f842
1 file changed
+3
-8
vllm/v1/worker/gpu/model_runner.py
+3
-8
@@ -153,9 +153,9 @@ class GPUModelRunner(LoRAModelRunnerMixin):
153
self.cp_interleave = self.parallel_config.cp_kv_cache_interleave_size
154
155
self.speculator = None
156
+ self.num_speculative_steps = 0
157
self.use_aux_hidden_state_outputs = False
158
if self.speculative_config is not None:
158
- self.do_spec_decode = True
159
self.num_speculative_steps = self.speculative_config.num_speculative_tokens
160
if self.is_last_pp_rank:
161
self.speculator = init_speculator(self.vllm_config, self.device)
@@ -165,9 +165,6 @@ class GPUModelRunner(LoRAModelRunnerMixin):
165
self.use_aux_hidden_state_outputs = True
166
if self.pp_size > 1:
167
raise ValueError("EAGLE3 with pipeline parallel is not supported.")
168
- else:
169
- self.do_spec_decode = False
170
- self.num_speculative_steps = 0
168
169
# Draft tokens propagation - for spec-dec + struct outputs.
170
self.draft_tokens_handler = DraftTokensHandler(self.device)
@@ -251,10 +248,8 @@ class GPUModelRunner(LoRAModelRunnerMixin):
248
)
249
250
prepare_communication_buffer_for_model(self.model)
254
- if self.do_spec_decode:
255
- speculator_model = getattr(self.speculator, "model", None)
256
- if speculator_model is not None:
257
- prepare_communication_buffer_for_model(speculator_model)
251
+ if self.speculator is not None:
252
+ prepare_communication_buffer_for_model(self.speculator)
253
254
def get_model(self) -> nn.Module:
255
return self.model