From 234df712ed6da81d8c8213a16cd913a747380214 Mon Sep 17 00:00:00 2001 From: Nick Hill Date: Fri, 25 Sep 2026 07:59:27 -0700 Subject: [PATCH] [MRV2] Minor model_runner.py code cleanup (#58610) Signed-off-by: Nick Hill --- tests/v1/worker/test_gpu_model_runner_v2.py | 14 -------------- vllm/v1/worker/gpu/model_runner.py | 16 +++------------- 2 files changed, 3 insertions(+), 27 deletions(-) diff --git a/tests/v1/worker/test_gpu_model_runner_v2.py b/tests/v1/worker/test_gpu_model_runner_v2.py index 83cd1e9d07..308c9fe9ea 100644 --- a/tests/v1/worker/test_gpu_model_runner_v2.py +++ b/tests/v1/worker/test_gpu_model_runner_v2.py @@ -21,20 +21,6 @@ from vllm.v1.worker.gpu.block_table import BlockTables from vllm.v1.worker.gpu.model_runner import GPUModelRunner -def test_prepare_padding_mask_marks_sequence_parallel_padding(): - runner = GPUModelRunner.__new__(GPUModelRunner) - runner.input_buffers = SimpleNamespace(is_padding=torch.empty(8, dtype=torch.bool)) - - mask = runner._prepare_padding_mask(1, 8) - - assert mask.tolist() == [False, True, True, True, True, True, True, True] - assert mask.data_ptr() == runner.input_buffers.is_padding.data_ptr() - - mask = runner._prepare_padding_mask(0, 8) - - assert mask.all() - - def test_qsa_circular_group_uses_custom_slot_mapping(monkeypatch): runner = GPUModelRunner.__new__(GPUModelRunner) runner.max_model_len = 262144 diff --git a/vllm/v1/worker/gpu/model_runner.py b/vllm/v1/worker/gpu/model_runner.py index 5c617533f1..c4dbe25843 100644 --- a/vllm/v1/worker/gpu/model_runner.py +++ b/vllm/v1/worker/gpu/model_runner.py @@ -797,8 +797,7 @@ class GPUModelRunner(LoRAModelRunnerMixin): assert num_tokens % self.decode_query_len == 0 elif self.speculator is not None: num_reqs = min( - num_reqs, - self.max_num_tokens // self.speculator.num_query_per_req, + num_reqs, self.max_num_tokens // self.speculator.num_query_per_req ) # Distribute the remainder evenly so no dummy request exceeds # ceil(num_tokens / num_reqs) <= max_model_len tokens. @@ -1257,14 +1256,6 @@ class GPUModelRunner(LoRAModelRunnerMixin): num_reqs, num_toks, max_query_len, batch_state.has_prefill ) - def _prepare_padding_mask( - self, num_tokens: int, num_tokens_after_padding: int - ) -> torch.Tensor: - is_padding = self.input_buffers.is_padding[:num_tokens_after_padding] - is_padding[:num_tokens].fill_(False) - is_padding[num_tokens:].fill_(True) - return is_padding - def prepare_inputs( self, scheduler_output: SchedulerOutput, @@ -1277,9 +1268,8 @@ class GPUModelRunner(LoRAModelRunnerMixin): assert num_tokens > 0 is_padding = self.input_buffers.is_padding[:num_tokens_after_padding] if envs.VLLM_MOE_SKIP_PADDING: - is_padding = self._prepare_padding_mask( - num_tokens, num_tokens_after_padding - ) + is_padding[:num_tokens].fill_(False) + is_padding[num_tokens:].fill_(True) req_ids = batch_req_state.req_ids num_scheduled_tokens_np = batch_req_state.num_scheduled_tokens