Skip to content
GitLab
Menu
Projects
Groups
Snippets
Loading...
Help
Help
Support
Community forum
Keyboard shortcuts
?
Submit feedback
Contribute to GitLab
Sign in / Register
Toggle navigation
Menu
Open sidebar
change
sglang
Commits
ddf39d3f
Unverified
Commit
ddf39d3f
authored
Feb 18, 2025
by
yigex
Committed by
GitHub
Feb 17, 2025
Browse files
[ROCm] Optimal MOE Tuning for AMD Radeon Graphics (#3567)
parent
2eab1132
Changes
2
Show whitespace changes
Inline
Side-by-side
Showing
2 changed files
with
51 additions
and
51 deletions
+51
-51
benchmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py
...hmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py
+1
-1
python/sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=256,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json
...adeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json
+50
-50
No files found.
benchmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py
View file @
ddf39d3f
...
@@ -175,7 +175,7 @@ def get_rocm_configs_compute_bound() -> List[Dict[str, int]]:
...
@@ -175,7 +175,7 @@ def get_rocm_configs_compute_bound() -> List[Dict[str, int]]:
for
block_m
in
[
32
,
64
,
128
,
256
]:
for
block_m
in
[
32
,
64
,
128
,
256
]:
for
block_k
in
[
32
,
64
,
128
,
256
]:
for
block_k
in
[
32
,
64
,
128
,
256
]:
for
block_n
in
[
16
,
32
,
64
,
128
,
256
]:
for
block_n
in
[
16
,
32
,
64
,
128
,
256
]:
for
num_warps
in
[
4
,
8
]:
for
num_warps
in
[
1
,
2
,
4
,
8
]:
for
group_size
in
[
1
,
4
,
8
,
16
,
32
]:
for
group_size
in
[
1
,
4
,
8
,
16
,
32
]:
configs
.
append
(
configs
.
append
(
{
{
...
...
python/sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=256,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json
View file @
ddf39d3f
{
{
"1"
:
{
"1"
:
{
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_N"
:
32
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
1
6
,
"GROUP_SIZE_M"
:
1
,
"num_warps"
:
4
,
"num_warps"
:
2
,
"num_stages"
:
2
,
"num_stages"
:
2
,
"waves_per_eu"
:
0
"waves_per_eu"
:
0
},
},
"2"
:
{
"2"
:
{
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_N"
:
64
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
8
,
"num_warps"
:
4
,
"num_warps"
:
2
,
"num_stages"
:
2
,
"num_stages"
:
2
,
"waves_per_eu"
:
0
"waves_per_eu"
:
0
},
},
"4"
:
{
"4"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_N"
:
64
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
1
6
,
"GROUP_SIZE_M"
:
1
,
"num_warps"
:
4
,
"num_warps"
:
2
,
"num_stages"
:
2
,
"num_stages"
:
2
,
"waves_per_eu"
:
0
"waves_per_eu"
:
0
},
},
"8"
:
{
"8"
:
{
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
64
,
"GROUP_SIZE_M"
:
32
,
"GROUP_SIZE_M"
:
1
,
"num_warps"
:
4
,
"num_warps"
:
2
,
"num_stages"
:
2
,
"num_stages"
:
2
,
"waves_per_eu"
:
0
"waves_per_eu"
:
0
},
},
"16"
:
{
"16"
:
{
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
64
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
4
,
"num_warps"
:
4
,
"num_warps"
:
2
,
"num_stages"
:
2
,
"num_stages"
:
2
,
"waves_per_eu"
:
0
"waves_per_eu"
:
0
},
},
...
@@ -48,17 +48,17 @@
...
@@ -48,17 +48,17 @@
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
4
,
"GROUP_SIZE_M"
:
8
,
"num_warps"
:
4
,
"num_warps"
:
2
,
"num_stages"
:
2
,
"num_stages"
:
2
,
"waves_per_eu"
:
0
"waves_per_eu"
:
0
},
},
"32"
:
{
"32"
:
{
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
64
,
"GROUP_SIZE_M"
:
8
,
"GROUP_SIZE_M"
:
4
,
"num_warps"
:
4
,
"num_warps"
:
2
,
"num_stages"
:
2
,
"num_stages"
:
2
,
"waves_per_eu"
:
0
"waves_per_eu"
:
0
},
},
...
@@ -66,8 +66,8 @@
...
@@ -66,8 +66,8 @@
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
4
,
"GROUP_SIZE_M"
:
1
,
"num_warps"
:
4
,
"num_warps"
:
2
,
"num_stages"
:
2
,
"num_stages"
:
2
,
"waves_per_eu"
:
0
"waves_per_eu"
:
0
},
},
...
@@ -75,8 +75,8 @@
...
@@ -75,8 +75,8 @@
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
4
,
"GROUP_SIZE_M"
:
1
,
"num_warps"
:
4
,
"num_warps"
:
2
,
"num_stages"
:
2
,
"num_stages"
:
2
,
"waves_per_eu"
:
0
"waves_per_eu"
:
0
},
},
...
@@ -84,77 +84,77 @@
...
@@ -84,77 +84,77 @@
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
8
,
"GROUP_SIZE_M"
:
4
,
"num_warps"
:
4
,
"num_warps"
:
2
,
"num_stages"
:
2
,
"num_stages"
:
2
,
"waves_per_eu"
:
0
"waves_per_eu"
:
0
},
},
"128"
:
{
"128"
:
{
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_N"
:
1
6
,
"BLOCK_SIZE_N"
:
1
28
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
4
,
"GROUP_SIZE_M"
:
4
,
"num_warps"
:
4
,
"num_warps"
:
2
,
"num_stages"
:
2
,
"num_stages"
:
2
,
"waves_per_eu"
:
0
"waves_per_eu"
:
0
},
},
"256"
:
{
"256"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_N"
:
1
6
,
"BLOCK_SIZE_N"
:
1
28
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
4
,
"num_warps"
:
4
,
"num_warps"
:
2
,
"num_stages"
:
2
,
"num_stages"
:
2
,
"waves_per_eu"
:
0
"waves_per_eu"
:
0
},
},
"512"
:
{
"512"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_N"
:
64
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
32
,
"GROUP_SIZE_M"
:
4
,
"num_warps"
:
4
,
"num_warps"
:
2
,
"num_stages"
:
2
,
"num_stages"
:
2
,
"waves_per_eu"
:
0
"waves_per_eu"
:
0
},
},
"1024"
:
{
"1024"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_N"
:
64
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
4
,
"GROUP_SIZE_M"
:
1
,
"num_warps"
:
8
,
"num_warps"
:
2
,
"num_stages"
:
2
,
"num_stages"
:
2
,
"waves_per_eu"
:
0
"waves_per_eu"
:
0
},
},
"1536"
:
{
"1536"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
6
4
,
"BLOCK_SIZE_N"
:
25
6
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
64
,
"GROUP_SIZE_M"
:
8
,
"GROUP_SIZE_M"
:
4
,
"num_warps"
:
4
,
"num_warps"
:
4
,
"num_stages"
:
2
,
"num_stages"
:
2
,
"waves_per_eu"
:
0
"waves_per_eu"
:
0
},
},
"2048"
:
{
"2048"
:
{
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_N"
:
64
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
1
,
"num_warps"
:
4
,
"num_warps"
:
2
,
"num_stages"
:
2
,
"num_stages"
:
2
,
"waves_per_eu"
:
0
"waves_per_eu"
:
0
},
},
"3072"
:
{
"3072"
:
{
"BLOCK_SIZE_M"
:
32
,
"BLOCK_SIZE_M"
:
128
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
256
,
"BLOCK_SIZE_K"
:
128
,
"BLOCK_SIZE_K"
:
128
,
"GROUP_SIZE_M"
:
1
,
"GROUP_SIZE_M"
:
4
,
"num_warps"
:
4
,
"num_warps"
:
4
,
"num_stages"
:
2
,
"num_stages"
:
2
,
"waves_per_eu"
:
0
"waves_per_eu"
:
0
},
},
"4096"
:
{
"4096"
:
{
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_M"
:
64
,
"BLOCK_SIZE_N"
:
128
,
"BLOCK_SIZE_N"
:
256
,
"BLOCK_SIZE_K"
:
64
,
"BLOCK_SIZE_K"
:
64
,
"GROUP_SIZE_M"
:
4
,
"GROUP_SIZE_M"
:
4
,
"num_warps"
:
4
,
"num_warps"
:
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