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 7, 15:15 UTC
[CI] Actually run tests/kernels/quantization/test_block_fp8.py in CI (#34274)
Michael Goin committed
Feb 26, 2026 at 19:58 UTC
4fec53cfcb3135f223ce52d2c6b7bbf9ddf2be63
3 files changed
+7
-9
.buildkite/test_areas/kernels.yaml
+1
-1
@@ -70,7 +70,7 @@ steps:
70
- tests/kernels/moe/test_batched_deepgemm.py
71
- tests/kernels/attention/test_deepgemm_attention.py
72
commands:
73
- - pytest -v -s kernels/quantization/test_block_fp8.py -k deep_gemm
73
+ - pytest -v -s kernels/quantization/test_block_fp8.py
74
- pytest -v -s kernels/moe/test_deepgemm.py
75
- pytest -v -s kernels/moe/test_batched_deepgemm.py
76
- pytest -v -s kernels/attention/test_deepgemm_attention.py
tests/kernels/quantization/test_block_fp8.py
+5
-7
@@ -37,13 +37,15 @@ vllm_config = VllmConfig()
37
38
# Test configurations
39
DTYPES = [torch.bfloat16] # [torch.half, torch.bfloat16, torch.float32]
40
+# Quantization test configs
41
NUM_TOKENS = [7, 2050]
42
D = [512, 4096, 5120, 13824]
43
GROUP_SIZE = [64, 128, 512]
44
COLUMN_MAJOR_SCALES = [True, False]
45
TMA_ALIGNED_SCALES = [True, False]
45
-M = [1, 7, 8, 83, 84, 4096]
46
-N = [128, 512, 7168, 7748, 13824]
46
+# Matmul test configs
47
+M = [1, 7, 8, 83, 4096]
48
+N = [128, 512, 576, 7168, 13824]
49
K = [256, 3884, 4096, 13824, 16384]
50
# Deepseek-V3's intermediate size 18432, so N is 18432*2/8=4608 at TP8
51
# and its hidden size is 7168.
@@ -162,8 +164,6 @@ def test_w8a8_block_fp8_cutlass_matmul():
164
k_tiles = (K + block_k - 1) // block_k
165
166
Bs = torch.rand(n_tiles, k_tiles, dtype=torch.float32) * factor_for_scale
165
- # Hopper requires row-major format for scales
166
- Bs_cutlass = Bs.T.contiguous() if current_platform.is_device_capability(90) else Bs
167
168
A_fp8, As = per_token_group_quant_fp8(
169
A_fp32, block_size[1], column_major_scales=False
@@ -174,9 +174,7 @@ def test_w8a8_block_fp8_cutlass_matmul():
174
)
175
176
ref_out = native_w8a8_block_matmul(A_fp8, B_fp8, As, Bs, block_size, out_dtype)
177
- out = cutlass_scaled_mm(
178
- A_fp8_cutlass, B_fp8, As_cutlass, Bs_cutlass, block_size, out_dtype
179
- )
177
+ out = cutlass_scaled_mm(A_fp8_cutlass, B_fp8, As_cutlass, Bs, block_size, out_dtype)
178
179
rel_diff = torch.mean(
180
torch.abs(out.to(torch.float32) - ref_out.to(torch.float32))
vllm/utils/flashinfer.py
+1
-1
@@ -734,7 +734,7 @@ def should_use_flashinfer_for_blockscale_fp8_gemm(
734
735
# Verify DeepGEMM N/K dims requirements
736
# NOTE: Also synchronized with test_w8a8_block_fp8_deep_gemm_matmul
737
- # test inside kernels/quatization/test_block_fp8.py
737
+ # test inside kernels/quantization/test_block_fp8.py
738
N_MULTIPLE = 64
739
K_MULTIPLE = 128
740