Correcting the mock classes for MM GC tests (#43321)

Signed-off-by: Weida Hong <wdhongtw@google.com>
This commit is contained in:
Weida Hong
2026-05-22 15:21:35 +08:00
committed by GitHub
parent 025d4f5cd2
commit 6bb8753db1
+3 -4
View File
@@ -17,6 +17,7 @@ from typing import Any
import pytest
import torch
from vllm.model_executor.models.interfaces import SupportsEncoderCudaGraph
from vllm.platforms import current_platform
from vllm.v1.worker.encoder_cudagraph import (
EncoderCudaGraphManager,
@@ -75,7 +76,7 @@ class _MockVllmConfig:
self.parallel_config = _MockParallelConfig()
class _MockModel:
class _MockModel(SupportsEncoderCudaGraph):
"""Minimal mock implementing SupportsEncoderCudaGraph for __init__."""
def __init__(self, min_budget: int = 4, max_budget: int = 128):
@@ -250,7 +251,7 @@ def _count_output_tokens(
return sum(t * (h // m) * (w // m) for t, h, w in grid_thw_list)
class SimpleMockViTModel(torch.nn.Module):
class SimpleMockViTModel(torch.nn.Module, SupportsEncoderCudaGraph):
"""Minimal ViT model for CUDA graph tests.
Implements the SupportsEncoderCudaGraph protocol by providing
@@ -258,8 +259,6 @@ class SimpleMockViTModel(torch.nn.Module):
simulates spatial merge by averaging groups of m^2 patches.
"""
supports_encoder_cudagraph = True
def __init__(self):
super().__init__()
self.proj = torch.nn.Linear(_FLAT, _HIDDEN)