| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import torch |
| 5 | from torch import nn |
| 6 | |
| 7 | from vllm.compilation.decorators import support_torch_compile |
| 8 | from vllm.config import ( |
| 9 | CompilationConfig, |
| 10 | CompilationMode, |
| 11 | VllmConfig, |
| 12 | set_current_vllm_config, |
| 13 | ) |
| 14 | from vllm.forward_context import set_forward_context |
| 15 | from vllm.model_executor.models.qwen3_vl import ( |
| 16 | Qwen3VLForConditionalGeneration, |
| 17 | ) |
| 18 | from vllm.sequence import IntermediateTensors |
| 19 | |
| 20 | |
| 21 | @support_torch_compile( |
| 22 | dynamic_arg_dims={"x": 0, "deepstack_input_embeds": 0}, |
| 23 | ) |
| 24 | class _DeepStackDecoder(nn.Module): |
| 25 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 26 | super().__init__() |
| 27 | |
| 28 | def forward( |
| 29 | self, |
| 30 | x: torch.Tensor, |
| 31 | deepstack_input_embeds: IntermediateTensors | None, |
| 32 | ) -> torch.Tensor: |
| 33 | if deepstack_input_embeds is not None: |
| 34 | x = x + deepstack_input_embeds["deepstack_input_embeds_0"] |
| 35 | return x |
| 36 | |
| 37 | |
| 38 | def _run_contract( |
| 39 | deepstack_buffer: Qwen3VLForConditionalGeneration, |
| 40 | ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: |
| 41 | config = VllmConfig( |
| 42 | compilation_config=CompilationConfig(mode=CompilationMode.VLLM_COMPILE) |
| 43 | ) |
| 44 | with set_current_vllm_config(config): |
| 45 | model = _DeepStackDecoder(vllm_config=config).eval().cuda() |
| 46 | |
| 47 | x = torch.arange(8, dtype=torch.float32, device="cuda").reshape(2, 4) |
| 48 | with torch.inference_mode(), set_forward_context(None, vllm_config=config): |
| 49 | empty_payload = deepstack_buffer._get_deepstack_input_embeds(num_tokens=2) |
| 50 | first = model(x, empty_payload) |
| 51 | |
| 52 | payload = torch.full((1, 2, 4), 7.0, device="cuda") |
| 53 | deepstack_buffer._set_deepstack_input_embeds(payload) |
| 54 | active_payload = deepstack_buffer._get_deepstack_input_embeds(num_tokens=2) |
| 55 | assert active_payload is not None |
| 56 | second = model(x, active_payload) |
| 57 | |
| 58 | return x.cpu(), first.cpu(), second.cpu() |
| 59 | |
| 60 | |
| 61 | def test_qwen3_vl_deepstack_payload_after_empty_compile(monkeypatch): |
| 62 | monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", "1") |
| 63 | |
| 64 | deepstack_buffer = object.__new__(Qwen3VLForConditionalGeneration) |
| 65 | deepstack_buffer.deepstack_input_embeds = [torch.zeros(2, 4, device="cuda")] |
| 66 | deepstack_buffer.deepstack_input_embeds_num_tokens = 0 |
| 67 | deepstack_buffer.deepstack_num_level = 1 |
| 68 | |
| 69 | x, first, second = _run_contract(deepstack_buffer) |
| 70 | torch.testing.assert_close(first, x) |
| 71 | torch.testing.assert_close(second, x + 7) |