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 3, 15:15 UTC
[CI][ROCm] Remove unsupported cases in test_fusion.py (#38680)
Signed-off-by: charlifu <charlifu@amd.com>
Charlie Fu committed
May 14, 2026 at 16:37 UTC
4cfcc0866f233c0f223b78eaa205f936d5522324
3 files changed
+6
-10
tests/compile/passes/test_fuse_act_padding.py
+3
-5
@@ -7,7 +7,7 @@ import torch
7
8
import vllm.config
9
from tests.compile.backend import TestBackend
10
-from vllm._aiter_ops import is_aiter_found_and_supported, rocm_aiter_ops
10
+from vllm._aiter_ops import rocm_aiter_ops
11
from vllm.compilation.passes.utility.noop_elimination import NoOpEliminationPass
12
from vllm.compilation.passes.utility.post_cleanup import PostCleanupPass
13
from vllm.config import (
@@ -72,9 +72,8 @@ class TestModel(torch.nn.Module):
72
@pytest.mark.parametrize("hidden_size", [2880])
73
@pytest.mark.parametrize("num_local_experts", [128])
74
@pytest.mark.parametrize("x_pad_to_multiple", [256])
75
-@pytest.mark.skipif(
76
- not is_aiter_found_and_supported(),
77
- reason="Only test on ROCm with AITER installed and supported",
75
+@pytest.mark.skip(
76
+ reason="Skipping for now because of the accuracy issue. See: https://github.com/ROCm/aiter/issues/2614"
77
)
78
def test_fuse_act_padding(
79
dtype: torch.dtype,
@@ -116,7 +115,6 @@ def test_fuse_act_padding(
115
116
x = torch.rand(1, hidden_size)
117
torch._dynamo.mark_dynamic(x, 0)
119
-
118
outputs_unfused = model(x)
119
120
model_fused = torch.compile(model, backend=backend)
tests/compile/passes/test_fusion.py
-2
@@ -99,8 +99,6 @@ KERNEL_GROUPSHAPE_COMBINATIONS = (
99
100
# For Aiter tests we toggle use_aiter_quant_op
101
AITER_KERNEL_GROUPSHAPE_COMBINATIONS = [
102
- # Per-token with ROCmFP8ScaledMMLinearKernel
103
- (ROCmFP8ScaledMMLinearKernel, GroupShape.PER_TENSOR, False),
102
# Per-token with RowWiseTorchFP8ScaledMMLinearKernel
103
(RowWiseTorchFP8ScaledMMLinearKernel, GroupShape.PER_TOKEN, True),
104
(RowWiseTorchFP8ScaledMMLinearKernel, GroupShape.PER_TOKEN, False),
vllm/model_executor/layers/quantization/utils/fp8_utils.py
+3
-3
@@ -1245,7 +1245,7 @@ def process_fp8_weight_tensor_strategy(
1245
requantize_with_max_scale,
1246
)
1247
1248
- if current_platform.is_fp8_fnuz():
1248
+ if current_platform.is_fp8_fnuz() and weight.dtype == torch.float8_e4m3fn:
1249
weight, weight_scale, input_scale = normalize_e4m3fn_to_e4m3fnuz(
1250
weight=weight, weight_scale=weight_scale, input_scale=input_scale
1251
)
@@ -1271,7 +1271,7 @@ def process_fp8_weight_channel_strategy(
1271
normalize_e4m3fn_to_e4m3fnuz,
1272
)
1273
1274
- if current_platform.is_fp8_fnuz():
1274
+ if current_platform.is_fp8_fnuz() and weight.dtype == torch.float8_e4m3fn:
1275
weight, weight_scale, input_scale = normalize_e4m3fn_to_e4m3fnuz(
1276
weight=weight, weight_scale=weight_scale, input_scale=input_scale
1277
)
@@ -1288,7 +1288,7 @@ def process_fp8_weight_block_strategy(
1288
normalize_e4m3fn_to_e4m3fnuz,
1289
)
1290
1291
- if current_platform.is_fp8_fnuz():
1291
+ if current_platform.is_fp8_fnuz() and weight.dtype == torch.float8_e4m3fn:
1292
weight, weight_scale, _ = normalize_e4m3fn_to_e4m3fnuz(
1293
weight=weight, weight_scale=weight_scale
1294
)