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

Convert formatting to use ruff instead of yapf + isort (#26247)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-10-05 15:06:22 +01:00
committed by GitHub
parent 17edd8a807
commit d6953beb91
1508 changed files with 115244 additions and 94146 deletions
Download Patch File Download Diff File Expand all files Collapse all files

48
csrc/quantization/gptq_marlin/generate_kernels.py
View File

@@ -17,28 +17,32 @@ FILE_HEAD = """
namespace MARLIN_NAMESPACE_NAME {
""".strip()
TEMPLATE = ("template __global__ void Marlin<"
"{{scalar_t}}, "
"{{w_type_id}}, "
"{{s_type_id}}, "
"{{threads}}, "
"{{thread_m_blocks}}, "
"{{thread_n_blocks}}, "
"{{thread_k_blocks}}, "
"{{'true' if m_block_size_8 else 'false'}}, "
"{{stages}}, "
"{{group_blocks}}, "
"{{'true' if is_zp_float else 'false'}}>"
"( MARLIN_KERNEL_PARAMS );")
TEMPLATE = (
"template __global__ void Marlin<"
"{{scalar_t}}, "
"{{w_type_id}}, "
"{{s_type_id}}, "
"{{threads}}, "
"{{thread_m_blocks}}, "
"{{thread_n_blocks}}, "
"{{thread_k_blocks}}, "
"{{'true' if m_block_size_8 else 'false'}}, "
"{{stages}}, "
"{{group_blocks}}, "
"{{'true' if is_zp_float else 'false'}}>"
"( MARLIN_KERNEL_PARAMS );"
)
# int8 with zero point case (vllm::kU8) is also supported,
# we don't add it to reduce wheel size.
SCALAR_TYPES = [
"vllm::kU4", "vllm::kU4B8", "vllm::kU8B128", "vllm::kFE4M3fn",
"vllm::kFE2M1f"
"vllm::kU4",
"vllm::kU4B8",
"vllm::kU8B128",
"vllm::kFE4M3fn",
"vllm::kFE2M1f",
]
THREAD_CONFIGS = [(128, 128, 256), (64, 256, 256), (64, 128, 128),
(128, 64, 128)]
THREAD_CONFIGS = [(128, 128, 256), (64, 256, 256), (64, 128, 128), (128, 64, 128)]
THREAD_M_BLOCKS = [0.5, 1, 2, 3, 4]
# group_blocks:
@@ -59,11 +63,12 @@ def generate_new_kernels():
all_template_str_list = []
for group_blocks, m_blocks, thread_configs in itertools.product(
GROUP_BLOCKS, THREAD_M_BLOCKS, THREAD_CONFIGS):
GROUP_BLOCKS, THREAD_M_BLOCKS, THREAD_CONFIGS
):
# act order case only support gptq-int4 and gptq-int8
if group_blocks == 0 and scalar_type not in [
"vllm::kU4B8", "vllm::kU8B128"
"vllm::kU4B8",
"vllm::kU8B128",
]:
continue
if thread_configs[2] == 256:
@@ -93,8 +98,7 @@ def generate_new_kernels():
c_dtype = "half" if dtype == "fp16" else "nv_bfloat16"
is_zp_float_list = [False]
if dtype == "fp16" and scalar_type == "vllm::kU4" and \
group_blocks == 4:
if dtype == "fp16" and scalar_type == "vllm::kU4" and group_blocks == 4:
# HQQ (is_zp_float = true) only supports
# 4bit quantization and fp16
is_zp_float_list.append(True)