From 470229c37efaf69c86e8bc97482b0b1ff7551c65 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Juan=20P=C3=A9rez=20de=20Algaba?= <124347725+jperezdealgaba@users.noreply.github.com> Date: Sat, 13 Jun 2026 12:17:38 +0200 Subject: [PATCH] [Security] Fix DoS via prompt_embeds on M-RoPE models (#45252) Signed-off-by: jperezde --- tests/v1/worker/test_mrope_prompt_embeds.py | 79 +++++++++++++++++++++ vllm/v1/worker/gpu_model_runner.py | 19 +++-- 2 files changed, 94 insertions(+), 4 deletions(-) create mode 100644 tests/v1/worker/test_mrope_prompt_embeds.py diff --git a/tests/v1/worker/test_mrope_prompt_embeds.py b/tests/v1/worker/test_mrope_prompt_embeds.py new file mode 100644 index 0000000000..209b88f522 --- /dev/null +++ b/tests/v1/worker/test_mrope_prompt_embeds.py @@ -0,0 +1,79 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +"""Test that M-RoPE position initialization handles prompt_embeds-only inputs. + +Regression test for GHSA-33cg-gxv8-3p8g: sending /v1/completions with +prompt_embeds and no prompt_token_ids on M-RoPE models crashed the +EngineCore via an assertion failure. +""" + +from unittest.mock import Mock + +import pytest +import torch + +from vllm.model_executor.models.interfaces import SupportsMRoPE +from vllm.v1.worker.gpu_input_batch import CachedRequestState +from vllm.v1.worker.gpu_model_runner import GPUModelRunner + + +class FakeMRoPEModel(SupportsMRoPE): + """Minimal model that passes supports_mrope() check.""" + + def get_mrope_input_positions(self, input_tokens, mm_features): + seq_len = len(input_tokens) + positions = torch.arange(seq_len).unsqueeze(0).expand(3, -1) + return positions.clone(), 0 + + +def _make_runner_and_req(prompt_token_ids, prompt_embeds): + """Create a minimal GPUModelRunner instance and request state.""" + model = FakeMRoPEModel() + instance = object.__new__(GPUModelRunner) + instance.get_model = lambda: model + + req_state = Mock(spec=CachedRequestState) + req_state.prompt_token_ids = prompt_token_ids + req_state.prompt_embeds = prompt_embeds + req_state.mm_features = [] + req_state.mrope_positions = None + req_state.mrope_position_delta = None + return instance, req_state + + +class TestMRopePromptEmbeds: + """Verify _init_mrope_positions handles prompt_embeds-only inputs.""" + + def test_prompt_embeds_only_does_not_crash(self): + """Prompt-embeds-only request must not raise AssertionError.""" + instance, req_state = _make_runner_and_req( + prompt_token_ids=None, + prompt_embeds=torch.randn(15, 896), + ) + + instance._init_mrope_positions(req_state) + + assert req_state.mrope_positions is not None + assert req_state.mrope_positions.shape == (3, 15) + + def test_prompt_token_ids_still_works(self): + """Normal path with prompt_token_ids continues working.""" + instance, req_state = _make_runner_and_req( + prompt_token_ids=[1, 2, 3, 4, 5], + prompt_embeds=None, + ) + + instance._init_mrope_positions(req_state) + + assert req_state.mrope_positions is not None + assert req_state.mrope_positions.shape == (3, 5) + + def test_neither_token_ids_nor_embeds_raises(self): + """When both are None, a ValueError should be raised.""" + instance, req_state = _make_runner_and_req( + prompt_token_ids=None, + prompt_embeds=None, + ) + + with pytest.raises(ValueError, match="prompt_token_ids or prompt_embeds"): + instance._init_mrope_positions(req_state) diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py index cb607c0b7b..afda4ec0bb 100644 --- a/vllm/v1/worker/gpu_model_runner.py +++ b/vllm/v1/worker/gpu_model_runner.py @@ -1588,9 +1588,6 @@ class GPUModelRunner( def _init_mrope_positions(self, req_state: CachedRequestState): model = self.get_model() assert supports_mrope(model), "M-RoPE support is not implemented." - assert req_state.prompt_token_ids is not None, ( - "M-RoPE requires prompt_token_ids to be available." - ) mrope_model = cast(SupportsMRoPE, model) # `prompt_embeds` is a passthrough modality (no grid_thw), models' @@ -1599,9 +1596,23 @@ class GPUModelRunner( mrope_features = [ f for f in req_state.mm_features if f.modality != "prompt_embeds" ] + + if req_state.prompt_token_ids is not None: + input_tokens = req_state.prompt_token_ids + elif req_state.prompt_embeds is not None: + # For embeddings-only inputs, get_mrope_input_positions only + # needs the sequence length when mm_features is empty (which is + # the case here since prompt_embeds are filtered out above). + seq_len = req_state.prompt_embeds.shape[0] + input_tokens = list(range(seq_len)) + else: + raise ValueError( + "M-RoPE requires either prompt_token_ids or prompt_embeds." + ) + req_state.mrope_positions, req_state.mrope_position_delta = ( mrope_model.get_mrope_input_positions( - req_state.prompt_token_ids, + input_tokens, mrope_features, ) )