xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 02:25:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

Speed up the kernels/quantization/ tests (#18669)

Browse Source 
Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin 2025-05-25 01:02:59 -04:00 committed by GitHub
parent 75f81750f3
commit 63934543a0
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 17 additions and 25 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
tests/kernels/quantization/test_block_fp8.py
View File

@ -36,16 +36,16 @@ vllm_config.scheduler_config.max_model_len = 8192
# Test configurations
DTYPES = [torch.bfloat16] # [torch.half, torch.bfloat16, torch.float32]
NUM_TOKENS = [7, 83, 2048]
NUM_TOKENS = [7, 2050]
D = [512, 4096, 5120, 13824]
GROUP_SIZE = [64, 128, 256, 512]
M = [1, 7, 8, 83, 84, 512, 2048, 4096]
N = [128, 512, 1024, 4096, 7168, 7748, 13824]
K = [256, 4096, 5120, 3884, 13824, 16384]
GROUP_SIZE = [64, 128, 512]
M = [1, 7, 8, 83, 84, 4096]
N = [128, 512, 7168, 7748, 13824]
K = [256, 3884, 4096, 13824, 16384]
# Deepseek-V3's intermediate size 18432, so N is 18432*2/8=4608 at TP8
# and its hidden size is 7168.
M_moe = [1, 2, 7, 83, 128, 512, 2048]
M_moe_dg = [128, 192, 512, 1335, 2048]
M_moe = [1, 2, 7, 83, 128, 2048]
M_moe_dg = [128, 192, 1335, 2048]
N_moe = [128, 256, 1024, 4608] # [13824]
K_moe = [256, 512, 7168] # [13824]
BLOCK_SIZE = [[128, 128]]

4
tests/kernels/quantization/test_gguf.py
View File

@ -35,11 +35,11 @@ def get_gguf_MoE_tensors(
return GGUFReader(sample_file).tensors
DTYPES = [torch.half, torch.bfloat16, torch.float32]
DTYPES = [torch.bfloat16] # [torch.half, torch.bfloat16, torch.float32]
# Hidden_size for testing, must match the sample file in HF repo,
# we have `hidden_size = 256, 1024` for test in HF repo currently.
HIDDEN_SIZES = [256, 1024]
NUM_TOKENS = [7, 83, 128, 2048] # Arbitrary values for testing
NUM_TOKENS = [7, 2050] # Arbitrary values for testing
SEEDS = [0]
QUANT_TYPES = [
# i-matrix

24
tests/kernels/quantization/test_triton_scaled_mm.py
View File

@ -13,8 +13,13 @@ from vllm.platforms import current_platform
device = "cuda"
triton_scaled_mm_module = importlib.import_module(
"vllm.model_executor.layers.quantization.compressed_tensors."
"triton_scaled_mm")
triton_scaled_mm = triton_scaled_mm_module.triton_scaled_mm
def scaled_mm_torch(a: torch.Tensor,
def torch_scaled_mm(a: torch.Tensor,
b: torch.Tensor,
scale_a: torch.Tensor,
scale_b: torch.Tensor,
@ -101,21 +106,8 @@ def test_scaled_mm(M, N, K, in_dtype, out_dtype, use_scalar_scale_a,
if use_bias:
bias = torch.rand((N, ), device=device, dtype=out_dtype)
triton_scaled_mm_module = importlib.import_module(
"vllm.model_executor.layers.quantization.compressed_tensors."
"triton_scaled_mm")
triton_scaled_mm = triton_scaled_mm_module.triton_scaled_mm
c_check = triton_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)
a_cpu = a.cpu()
b_cpu = b.cpu()
scale_a_cpu = scale_a.cpu()
scale_b_cpu = scale_b.cpu()
bias_cpu = None if bias is None else bias.cpu()
c_actual = torch_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)
c_actual = scaled_mm_torch(a_cpu, b_cpu, scale_a_cpu, scale_b_cpu,
out_dtype, bias_cpu)
c_check_cpu = c_check.cpu()
torch.testing.assert_close(c_check_cpu, c_actual, rtol=1e-1, atol=1e-1)
torch.testing.assert_close(c_check, c_actual, rtol=1e-1, atol=1e-1)