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

[Kernel/Quant] Remove the original marlin format and qqq (#23204)

Browse Source 
Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2025-08-20 15:13:36 -04:00
committed by GitHub
parent ebe56a0064
commit 0cdbf5e61c
26 changed files with 92 additions and 3698 deletions
Download Patch File Download Diff File Expand all files Collapse all files

139
csrc/quantization/machete/generate.py
View File

@@ -571,78 +571,79 @@ def generate():
itertools.repeat(default_heuristic))
]
# Stored as "condition": ((tile_shape_mn), (cluster_shape_mnk))
# TODO (LucasWilkinson): Further tuning required
qqq_tile_heuristic_config = {
#### M = 257+
# ((128, 256), (2, 1, 1)) Broken for QQQ types
# TODO (LucasWilkinson): Investigate further
# "M > 256 && K <= 16384 && N <= 4096": ((128, 128), (2, 1, 1)),
# "M > 256": ((128, 256), (2, 1, 1)),
"M > 256": ((128, 128), (2, 1, 1)),
#### M = 129-256
"M > 128 && K <= 4096 && N <= 4096": ((128, 64), (2, 1, 1)),
"M > 128 && K <= 8192 && N <= 8192": ((128, 128), (2, 1, 1)),
# ((128, 256), (2, 1, 1)) Broken for QQQ types
# TODO (LucasWilkinson): Investigate further
# "M > 128": ((128, 256), (2, 1, 1)),
"M > 128": ((128, 128), (2, 1, 1)),
#### M = 65-128
"M > 64 && K <= 4069 && N <= 4069": ((128, 32), (2, 1, 1)),
"M > 64 && K <= 4069 && N <= 8192": ((128, 64), (2, 1, 1)),
"M > 64 && K >= 8192 && N >= 12288": ((256, 128), (2, 1, 1)),
"M > 64": ((128, 128), (2, 1, 1)),
#### M = 33-64
"M > 32 && K <= 6144 && N <= 6144": ((128, 16), (1, 1, 1)),
# Broken for QQQ types
# TODO (LucasWilkinson): Investigate further
#"M > 32 && K >= 16384 && N >= 12288": ((256, 64), (2, 1, 1)),
"M > 32": ((128, 64), (2, 1, 1)),
#### M = 17-32
"M > 16 && K <= 12288 && N <= 8192": ((128, 32), (2, 1, 1)),
"M > 16": ((256, 32), (2, 1, 1)),
#### M = 1-16
"N >= 26624": ((256, 16), (1, 1, 1)),
None: ((128, 16), (1, 1, 1)),
}
# TODO: Support W4A8 when ready
# # Stored as "condition": ((tile_shape_mn), (cluster_shape_mnk))
# # TODO (LucasWilkinson): Further tuning required
# qqq_tile_heuristic_config = {
# #### M = 257+
# # ((128, 256), (2, 1, 1)) Broken for QQQ types
# # TODO (LucasWilkinson): Investigate further
# # "M > 256 && K <= 16384 && N <= 4096": ((128, 128), (2, 1, 1)),
# # "M > 256": ((128, 256), (2, 1, 1)),
# "M > 256": ((128, 128), (2, 1, 1)),
# #### M = 129-256
# "M > 128 && K <= 4096 && N <= 4096": ((128, 64), (2, 1, 1)),
# "M > 128 && K <= 8192 && N <= 8192": ((128, 128), (2, 1, 1)),
# # ((128, 256), (2, 1, 1)) Broken for QQQ types
# # TODO (LucasWilkinson): Investigate further
# # "M > 128": ((128, 256), (2, 1, 1)),
# "M > 128": ((128, 128), (2, 1, 1)),
# #### M = 65-128
# "M > 64 && K <= 4069 && N <= 4069": ((128, 32), (2, 1, 1)),
# "M > 64 && K <= 4069 && N <= 8192": ((128, 64), (2, 1, 1)),
# "M > 64 && K >= 8192 && N >= 12288": ((256, 128), (2, 1, 1)),
# "M > 64": ((128, 128), (2, 1, 1)),
# #### M = 33-64
# "M > 32 && K <= 6144 && N <= 6144": ((128, 16), (1, 1, 1)),
# # Broken for QQQ types
# # TODO (LucasWilkinson): Investigate further
# #"M > 32 && K >= 16384 && N >= 12288": ((256, 64), (2, 1, 1)),
# "M > 32": ((128, 64), (2, 1, 1)),
# #### M = 17-32
# "M > 16 && K <= 12288 && N <= 8192": ((128, 32), (2, 1, 1)),
# "M > 16": ((256, 32), (2, 1, 1)),
# #### M = 1-16
# "N >= 26624": ((256, 16), (1, 1, 1)),
# None: ((128, 16), (1, 1, 1)),
# }
# For now we use the same heuristic for all types
# Heuristic is currently tuned for H100s
qqq_heuristic = [
(cond, ScheduleConfig(*tile_config,
**sch_common_params)) # type: ignore
for cond, tile_config in qqq_tile_heuristic_config.items()
]
# # For now we use the same heuristic for all types
# # Heuristic is currently tuned for H100s
# qqq_heuristic = [
# (cond, ScheduleConfig(*tile_config,
# **sch_common_params)) # type: ignore
# for cond, tile_config in qqq_tile_heuristic_config.items()
# ]
QQQ_kernel_types = [
*(TypeConfig(
a=DataType.s8,
b=VLLMDataType.u4b8,
b_group_scale=b_group_scale,
b_group_zeropoint=DataType.void,
b_channel_scale=DataType.f32,
a_token_scale=DataType.f32,
out=DataType.f16,
accumulator=DataType.s32,
) for b_group_scale in (DataType.f16, DataType.void)),
*(TypeConfig(
a=DataType.e4m3,
b=VLLMDataType.u4b8,
b_group_scale=b_group_scale,
b_group_zeropoint=DataType.void,
b_channel_scale=DataType.f32,
a_token_scale=DataType.f32,
out=DataType.f16,
accumulator=DataType.f32,
) for b_group_scale in (DataType.f16, DataType.void)),
]
# QQQ_kernel_types = [
# *(TypeConfig(
# a=DataType.s8,
# b=VLLMDataType.u4b8,
# b_group_scale=b_group_scale,
# b_group_zeropoint=DataType.void,
# b_channel_scale=DataType.f32,
# a_token_scale=DataType.f32,
# out=DataType.f16,
# accumulator=DataType.s32,
# ) for b_group_scale in (DataType.f16, DataType.void)),
# *(TypeConfig(
# a=DataType.e4m3,
# b=VLLMDataType.u4b8,
# b_group_scale=b_group_scale,
# b_group_zeropoint=DataType.void,
# b_channel_scale=DataType.f32,
# a_token_scale=DataType.f32,
# out=DataType.f16,
# accumulator=DataType.f32,
# ) for b_group_scale in (DataType.f16, DataType.void)),
# ]
impl_configs += [
ImplConfig(x[0], x[1], x[2])
for x in zip(QQQ_kernel_types,
itertools.repeat(get_unique_schedules(qqq_heuristic)),
itertools.repeat(qqq_heuristic))
]
# impl_configs += [
# ImplConfig(x[0], x[1], x[2])
# for x in zip(QQQ_kernel_types,
# itertools.repeat(get_unique_schedules(qqq_heuristic)),
# itertools.repeat(qqq_heuristic))
# ]
output_dir = os.path.join(SCRIPT_DIR, "generated")