A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 2, 03:30 UTC
[ROCm][CI] Skip fp8 reload tests on gfx90a (MI250) (#44369)
Signed-off-by: JartX <sagformas@epdcenter.es>
JartX committed
Jun 3, 2026 at 05:27 UTC
7b476c8f14f1774373ec4e8581b2f3147580f3c2
1 file changed
+19
-3
tests/model_executor/model_loader/test_reload.py
+19
-3
@@ -28,6 +28,22 @@ from vllm.model_executor.model_loader.reload.utils import get_layer_tensors
28
from vllm.platforms import current_platform
29
30
31
+def _fp8_reload_unsupported() -> bool:
32
+ """Whether the FP8 reload/online-quantize tests should be skipped.
33
+
34
+ ``supports_fp8()`` returns True on MI250 (gfx90a) because the general
35
+ quantization paths upcast FP8 weights, but gfx90a has no native FP8 and
36
+ cannot run these reload models, so treat it as unsupported here.
37
+ """
38
+ if not current_platform.supports_fp8():
39
+ return True
40
+ if current_platform.is_rocm():
41
+ from vllm.platforms.rocm import on_gfx90a
42
+
43
+ return on_gfx90a()
44
+ return False
45
+
46
+
47
class _AliasedBufferLayer(torch.nn.Module):
48
def __init__(self):
49
super().__init__()
@@ -284,7 +300,7 @@ def test_reload_weights(base_model, mul_model, add_model, tp_size, vllm_runner):
300
if current_platform.device_count() < tp_size:
301
pytest.skip(reason="Not enough CUDA devices")
302
287
- if "FP8" in base_model and not current_platform.supports_fp8():
303
+ if "FP8" in base_model and _fp8_reload_unsupported():
304
pytest.skip(reason="Requires FP8 support")
305
306
with vllm_runner(
@@ -308,7 +324,7 @@ def test_reload_weights(base_model, mul_model, add_model, tp_size, vllm_runner):
324
325
def test_kv_scale_reload(vllm_runner):
326
"""Test reloading a checkpoint that contains k_scale/v_scale weights."""
311
- if not current_platform.supports_fp8():
327
+ if _fp8_reload_unsupported():
328
pytest.skip(reason="Requires FP8 support")
329
330
model = "nm-testing/Llama-3.2-1B-Instruct-FP8-KV"
@@ -378,7 +394,7 @@ def test_online_quantize_reload(
394
if current_platform.device_count() < tp_size:
395
pytest.skip(reason="Not enough GPU devices")
396
381
- if quantization == "fp8" and not current_platform.supports_fp8():
397
+ if quantization == "fp8" and _fp8_reload_unsupported():
398
pytest.skip(reason="Requires FP8 support")
399
400
with vllm_runner(