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, 20:15 UTC
[Core] Remove unused `num_tokens` parameter from `_init_model_kwargs` (#31517)
Signed-off-by: maang <maang_h@163.com>
maang-h committed
Dec 31, 2025 at 12:47 UTC
b4bb5f312f3479e913c21c93f852077ee391b41f
1 file changed
+6
-6
vllm/v1/worker/gpu_model_runner.py
+6
-6
@@ -684,7 +684,7 @@ class GPUModelRunner(
684
with_numpy=numpy,
685
)
686
687
- def _init_model_kwargs(self, num_tokens: int):
687
+ def _init_model_kwargs(self):
688
model_kwargs = dict[str, Any]()
689
690
if not self.is_pooling_model:
@@ -2579,7 +2579,7 @@ class GPUModelRunner(
2579
2580
input_ids, inputs_embeds = self._prepare_mm_inputs(num_input_tokens)
2581
model_kwargs = {
2582
- **self._init_model_kwargs(num_scheduled_tokens),
2582
+ **self._init_model_kwargs(),
2583
**self._extract_mm_kwargs(scheduler_output),
2584
}
2585
elif self.enable_prompt_embeds and is_first_rank:
@@ -2607,7 +2607,7 @@ class GPUModelRunner(
2607
self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds
2608
2609
inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
2610
- model_kwargs = self._init_model_kwargs(num_input_tokens)
2610
+ model_kwargs = self._init_model_kwargs()
2611
input_ids = None
2612
else:
2613
# For text-only models, we use token ids as input.
@@ -2616,7 +2616,7 @@ class GPUModelRunner(
2616
# then the embedding layer is not included in the CUDA graph.
2617
input_ids = self.input_ids.gpu[:num_input_tokens]
2618
inputs_embeds = None
2619
- model_kwargs = self._init_model_kwargs(num_input_tokens)
2619
+ model_kwargs = self._init_model_kwargs()
2620
2621
if self.uses_mrope:
2622
positions = self.mrope_positions.gpu[:, :num_input_tokens]
@@ -4293,7 +4293,7 @@ class GPUModelRunner(
4293
):
4294
# Make sure padding doesn't exceed max_num_tokens
4295
assert num_tokens_padded <= self.max_num_tokens
4296
- model_kwargs = self._init_model_kwargs(num_tokens_padded)
4296
+ model_kwargs = self._init_model_kwargs()
4297
if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
4298
input_ids, inputs_embeds = self._prepare_mm_inputs(num_tokens_padded)
4299
@@ -4304,7 +4304,7 @@ class GPUModelRunner(
4304
elif self.enable_prompt_embeds:
4305
input_ids = None
4306
inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
4307
- model_kwargs = self._init_model_kwargs(num_tokens_padded)
4307
+ model_kwargs = self._init_model_kwargs()
4308
else:
4309
input_ids = self.input_ids.gpu[:num_tokens_padded]
4310
inputs_embeds = None