biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Quantization][Deprecation] Deprecate HQQ (#32681)

Browse Source 
Signed-off-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: Robert Shaw <robshaw@redhat.com>
This commit is contained in:
Robert Shaw
2026-01-21 09:32:40 -05:00
committed by GitHub
parent cea3c754c4
commit 85f55c943c
8 changed files with 0 additions and 480 deletions
Download Patch File Download Diff File Expand all files Collapse all files

87
tests/kernels/quantization/test_marlin_gemm.py
View File

@@ -26,7 +26,6 @@ from vllm.model_executor.layers.quantization.utils.marlin_utils import (
marlin_make_empty_g_idx,
marlin_make_workspace_new,
marlin_permute_bias,
marlin_permute_scales,
query_marlin_supported_quant_types,
)
from vllm.model_executor.layers.quantization.utils.marlin_utils_fp4 import (
@@ -75,8 +74,6 @@ MARLIN_N_CHUNKS = [64, 256]
MARLIN_24_K_CHUNKS = [128]
MARLIN_24_N_CHUNKS = [512]
HQQ_SUPPORTED_GROUP_SIZES = [64]
MARLIN_REPACK_NK_FACTORS = [
(4, 8),
(7, 5),
@@ -631,90 +628,6 @@ def test_gptq_marlin_24_gemm(k_chunk, n_chunk, quant_type, group_size, mnk_facto
assert max_diff < 0.04
@pytest.mark.skipif(
not is_quant_method_supported("gptq_marlin"),
reason="Marlin is not supported on this GPU type.",
)
@pytest.mark.parametrize("k_chunk", MARLIN_K_CHUNKS)
@pytest.mark.parametrize("n_chunk", MARLIN_N_CHUNKS)
@pytest.mark.parametrize("group_size", HQQ_SUPPORTED_GROUP_SIZES)
@pytest.mark.parametrize("mnk_factors", MNK_FACTORS)
@pytest.mark.parametrize("use_fp32_reduce", USE_FP32_REDUCE_OPTS)
def test_hqq_marlin_gemm(
k_chunk,
n_chunk,
group_size,
mnk_factors,
use_fp32_reduce,
):
m_factor, n_factor, k_factor = mnk_factors
size_m = m_factor
size_k = k_chunk * k_factor
size_n = n_chunk * n_factor
quant_type = scalar_types.uint4
a_input = rand_data((size_m, size_k))
dev = a_input.device
b_weight = torch.randint(0, 10, (size_n, size_k), dtype=torch.uint8, device=dev)
scale = rand_data((size_n, size_k // group_size))
zero = rand_data((size_n, size_k // group_size))
gptq_w_q = gptq_pack(b_weight.transpose(1, 0), 4, size_k, size_n)
sort_indices = torch.empty(0, dtype=torch.int, device=dev)
marlin_w_q = ops.gptq_marlin_repack(gptq_w_q, sort_indices, size_k, size_n, 4).to(
dev
)
marlin_s = marlin_permute_scales(
scale.transpose(1, 0), size_k, size_n, group_size
).to(dev)
marlin_zp = marlin_permute_scales(
zero.transpose(1, 0), size_k, size_n, group_size
).to(dev)
g_idx = marlin_make_empty_g_idx(dev)
g_idx_sort_indices = marlin_make_empty_g_idx(dev)
workspace = marlin_make_workspace_new(b_weight.device)
output = ops.gptq_marlin_gemm(
a_input,
None,
marlin_w_q,
None,
marlin_s,
None,
None,
marlin_zp,
g_idx,
g_idx_sort_indices,
workspace,
quant_type,
a_input.shape[0],
b_weight.shape[0],
a_input.shape[1],
is_k_full=True,
use_fp32_reduce=use_fp32_reduce,
is_zp_float=True,
)
b_flat = b_weight.reshape(-1, group_size)
zp_flat = zero.reshape(-1, 1)
s_flat = scale.reshape(-1, 1)
dequant = (b_flat - zp_flat) * s_flat
output_ref = torch.matmul(a_input, dequant.reshape(b_weight.shape).transpose(1, 0))
torch.cuda.synchronize()
max_diff = compute_max_diff(output, output_ref)
assert max_diff < 0.04
def test_marlin_gemm_subset_input():
quant_type = scalar_types.uint4b8
group_size = 128