Skip to content
GitLab
Menu
Projects
Groups
Snippets
Loading...
Help
Help
Support
Community forum
Keyboard shortcuts
?
Submit feedback
Contribute to GitLab
Sign in
Toggle navigation
Menu
Open sidebar
OpenDAS
vllm_cscc
Commits
51a08e7d
"...git@developer.sourcefind.cn:2222/OpenDAS/vllm_cscc.git" did not exist on "fd46df331de3dd1832e601cf589204f424312241"
Unverified
Commit
51a08e7d
authored
Jun 05, 2024
by
Philipp Moritz
Committed by
GitHub
Jun 05, 2024
Browse files
[Kernel] Re-tune Mixtral MoE configurations for FP8 on H100 (#5238)
parent
eb8fcd26
Changes
3
Hide whitespace changes
Inline
Side-by-side
Showing
3 changed files
with
85 additions
and
84 deletions
+85
-84
benchmarks/kernels/benchmark_moe.py
benchmarks/kernels/benchmark_moe.py
+2
-1
vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=float8.json
...=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=float8.json
+52
-52
vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=float8.json
...=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=float8.json
+31
-31
No files found.
benchmarks/kernels/benchmark_moe.py
View file @
51a08e7d
...
@@ -255,7 +255,8 @@ def main(args: argparse.Namespace):
...
@@ -255,7 +255,8 @@ def main(args: argparse.Namespace):
if
args
.
batch_size
is
None
:
if
args
.
batch_size
is
None
:
batch_sizes
=
[
batch_sizes
=
[
1
,
2
,
4
,
8
,
16
,
32
,
64
,
128
,
256
,
512
,
1024
,
1536
,
2048
,
3072
,
4096
1
,
2
,
4
,
8
,
16
,
24
,
32
,
48
,
64
,
96
,
128
,
256
,
512
,
1024
,
1536
,
2048
,
3072
,
4096
]
]
else
:
else
:
batch_sizes
=
[
args
.
batch_size
]
batch_sizes
=
[
args
.
batch_size
]
...
...
vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=float8.json
View file @
51a08e7d
{
{
"1"
:
{
"1"
:
{
"BLOCK_SIZE_M"
:
1
6
,
"BLOCK_SIZE_M"
:
6
4
,
"BLOCK_SIZE_N"
:
32
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
64
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
32
,
"num_warps"
:
8
,
"num_warps"
:
4
,
"num_stages"
:
4
"num_stages"
:
5
},
},
"2"
:
{
"2"
:
{
"BLOCK_SIZE_M"
:
1
6
,
"BLOCK_SIZE_M"
:
6
4
,
"BLOCK_SIZE_N"
:
256
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
64
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
1
,
"num_warps"
:
8
,
"num_warps"
:
4
,
"num_stages"
:
3
"num_stages"
:
4
},
},
"4"
:
{
"4"
:
{
"BLOCK_SIZE_M"
:
25
6
,
"BLOCK_SIZE_M"
:
6
4
,
"BLOCK_SIZE_N"
:
64
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
256
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
1
6
,
"num_warps"
:
4
,
"num_warps"
:
4
,
"num_stages"
:
2
"num_stages"
:
4
},
},
"8"
:
{
"8"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
256
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
64
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
32
,
"GROUP_SIZE_M"
:
32
,
"num_warps"
:
8
,
"num_warps"
:
4
,
"num_stages"
:
2
"num_stages"
:
4
},
},
"16"
:
{
"16"
:
{
"BLOCK_SIZE_M"
:
25
6
,
"BLOCK_SIZE_M"
:
6
4
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
64
,
"BLOCK_SIZE_K"
:
64
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
1
,
"num_warps"
:
8
,
"num_warps"
:
4
,
"num_stages"
:
5
"num_stages"
:
3
},
},
"24"
:
{
"24"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
25
6
,
"BLOCK_SIZE_N"
:
6
4
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
64
,
"GROUP_SIZE_M"
:
64
,
"num_warps"
:
4
,
"num_warps"
:
4
,
"num_stages"
:
4
"num_stages"
:
3
},
},
"32"
:
{
"32"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
64
,
"BLOCK_SIZE_K"
:
256
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
64
,
"num_warps"
:
4
,
"num_warps"
:
4
,
"num_stages"
:
4
"num_stages"
:
3
},
},
"48"
:
{
"48"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
25
6
,
"BLOCK_SIZE_N"
:
6
4
,
"BLOCK_SIZE_K"
:
64
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
6
4
,
"GROUP_SIZE_M"
:
1
6
,
"num_warps"
:
4
,
"num_warps"
:
4
,
"num_stages"
:
4
"num_stages"
:
3
},
},
"64"
:
{
"64"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
64
,
"BLOCK_SIZE_K"
:
256
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
1
,
"num_warps"
:
4
,
"num_warps"
:
4
,
"num_stages"
:
2
"num_stages"
:
3
},
},
"96"
:
{
"96"
:
{
"BLOCK_SIZE_M"
:
128
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
32
,
"BLOCK_SIZE_N"
:
64
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
1
,
"num_warps"
:
4
,
"num_warps"
:
4
,
"num_stages"
:
5
"num_stages"
:
3
},
},
"128"
:
{
"128"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
64
,
"BLOCK_SIZE_N"
:
64
,
"BLOCK_SIZE_K"
:
256
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
32
,
"GROUP_SIZE_M"
:
1
,
"num_warps"
:
8
,
"num_warps"
:
4
,
"num_stages"
:
5
"num_stages"
:
3
},
},
"256"
:
{
"256"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
256
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
1
6
,
"num_warps"
:
8
,
"num_warps"
:
4
,
"num_stages"
:
3
"num_stages"
:
3
},
},
"512"
:
{
"512"
:
{
"BLOCK_SIZE_M"
:
128
,
"BLOCK_SIZE_M"
:
128
,
"BLOCK_SIZE_N"
:
6
4
,
"BLOCK_SIZE_N"
:
25
6
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
16
,
"GROUP_SIZE_M"
:
16
,
"num_warps"
:
8
,
"num_warps"
:
8
,
"num_stages"
:
3
"num_stages"
:
4
},
},
"1024"
:
{
"1024"
:
{
"BLOCK_SIZE_M"
:
128
,
"BLOCK_SIZE_M"
:
128
,
"BLOCK_SIZE_N"
:
256
,
"BLOCK_SIZE_N"
:
256
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
64
,
"GROUP_SIZE_M"
:
32
,
"num_warps"
:
8
,
"num_warps"
:
8
,
"num_stages"
:
4
"num_stages"
:
4
},
},
...
@@ -115,7 +115,7 @@
...
@@ -115,7 +115,7 @@
"BLOCK_SIZE_M"
:
128
,
"BLOCK_SIZE_M"
:
128
,
"BLOCK_SIZE_N"
:
256
,
"BLOCK_SIZE_N"
:
256
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
6
4
,
"GROUP_SIZE_M"
:
1
6
,
"num_warps"
:
8
,
"num_warps"
:
8
,
"num_stages"
:
4
"num_stages"
:
4
},
},
...
...
vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=float8.json
View file @
51a08e7d
{
{
"1"
:
{
"1"
:
{
"BLOCK_SIZE_M"
:
1
6
,
"BLOCK_SIZE_M"
:
6
4
,
"BLOCK_SIZE_N"
:
32
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
6
4
,
"BLOCK_SIZE_K"
:
25
6
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
64
,
"num_warps"
:
4
,
"num_warps"
:
4
,
"num_stages"
:
5
"num_stages"
:
4
},
},
"2"
:
{
"2"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
64
,
"BLOCK_SIZE_N"
:
64
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
256
,
"GROUP_SIZE_M"
:
32
,
"GROUP_SIZE_M"
:
1
,
"num_warps"
:
8
,
"num_warps"
:
4
,
"num_stages"
:
4
"num_stages"
:
5
},
},
"4"
:
{
"4"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
64
,
"BLOCK_SIZE_K"
:
256
,
"BLOCK_SIZE_K"
:
256
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
1
,
"num_warps"
:
4
,
"num_warps"
:
4
,
"num_stages"
:
2
"num_stages"
:
5
},
},
"8"
:
{
"8"
:
{
"BLOCK_SIZE_M"
:
128
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
256
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
32
,
"num_warps"
:
4
,
"num_warps"
:
4
,
"num_stages"
:
3
"num_stages"
:
4
},
},
"16"
:
{
"16"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
256
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
32
,
"num_warps"
:
4
,
"num_warps"
:
4
,
"num_stages"
:
4
"num_stages"
:
5
},
},
"24"
:
{
"24"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
64
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
256
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
1
,
"num_warps"
:
4
,
"num_warps"
:
4
,
"num_stages"
:
5
"num_stages"
:
3
},
},
"32"
:
{
"32"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
256
,
"BLOCK_SIZE_K"
:
256
,
"GROUP_SIZE_M"
:
64
,
"GROUP_SIZE_M"
:
1
,
"num_warps"
:
4
,
"num_warps"
:
4
,
"num_stages"
:
4
"num_stages"
:
4
},
},
"48"
:
{
"48"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
64
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
256
,
"BLOCK_SIZE_K"
:
256
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
1
,
"num_warps"
:
4
,
"num_warps"
:
4
,
"num_stages"
:
3
"num_stages"
:
4
},
},
"64"
:
{
"64"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
64
,
...
@@ -81,19 +81,19 @@
...
@@ -81,19 +81,19 @@
},
},
"128"
:
{
"128"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
64
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
256
,
"BLOCK_SIZE_K"
:
256
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
1
,
"num_warps"
:
4
,
"num_warps"
:
4
,
"num_stages"
:
3
"num_stages"
:
4
},
},
"256"
:
{
"256"
:
{
"BLOCK_SIZE_M"
:
128
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
64
,
"num_warps"
:
8
,
"num_warps"
:
4
,
"num_stages"
:
5
"num_stages"
:
3
},
},
"512"
:
{
"512"
:
{
"BLOCK_SIZE_M"
:
128
,
"BLOCK_SIZE_M"
:
128
,
...
@@ -107,7 +107,7 @@
...
@@ -107,7 +107,7 @@
"BLOCK_SIZE_M"
:
128
,
"BLOCK_SIZE_M"
:
128
,
"BLOCK_SIZE_N"
:
256
,
"BLOCK_SIZE_N"
:
256
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
64
,
"GROUP_SIZE_M"
:
32
,
"num_warps"
:
8
,
"num_warps"
:
8
,
"num_stages"
:
4
"num_stages"
:
4
},
},
...
...
Write
Preview
Markdown
is supported
0%
Try again
or
attach a new file
.
Attach a file
Cancel
You are about to add
0
people
to the discussion. Proceed with caution.
Finish editing this message first!
Cancel
Please
register
or
sign in
to comment