[Hardware][AMD][CI] Patch Whisper multi LoRA test to use TRITON_ATTN for now (#46928)
Signed-off-by: Matthew Wong <Matthew.Wong2@amd.com>
Matt committed
Jun 27, 2026 at 17:30 UTC
9036c89ee410b30913ca8b7d362a7d0805583b51
1 file changed
+10
-2
tests/lora/test_whisper.py
+10
-2
@@ -12,6 +12,7 @@ import pytest
12
import vllm
13
from vllm.assets.audio import AudioAsset
14
from vllm.lora.request import LoRARequest
15
+from vllm.platforms import current_platform
16
17
from ..utils import create_new_process_for_each_test
18
@@ -30,7 +31,9 @@ def use_spawn_for_whisper(monkeypatch):
31
monkeypatch.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
32
33
33
-def create_whisper_llm(enable_lora: bool = True, max_loras: int = 2):
34
+def create_whisper_llm(
35
+ enable_lora: bool = True, max_loras: int = 2, attn_backend: str | None = None
36
+):
37
"""Create a Whisper LLM instance with optional LoRA support."""
38
return vllm.LLM(
39
model=WHISPER_MODEL,
@@ -40,6 +43,7 @@ def create_whisper_llm(enable_lora: bool = True, max_loras: int = 2):
43
max_model_len=448,
44
dtype="half",
45
enforce_eager=True, # For stability in tests
46
+ attention_config={"backend": attn_backend},
47
)
48
49
@@ -109,7 +113,11 @@ def test_whisper_multi_lora(whisper_lora_files):
113
This test verifies that the same LoRA adapter can be loaded with
114
different IDs and produce consistent results.
115
"""
112
- llm = create_whisper_llm(enable_lora=True, max_loras=4)
116
+ llm = create_whisper_llm(
117
+ enable_lora=True,
118
+ max_loras=4,
119
+ attn_backend="TRITON_ATTN" if current_platform.is_rocm() else None,
120
+ )
121
122
# Test with different LoRA IDs using the same adapter
123
outputs_lora1 = run_whisper_inference(llm, lora_path=whisper_lora_files, lora_id=1)