[ROCm] Retune MI355 selective_state_update float16 config on the unified effective_batch grid (#48372)
Signed-off-by: vanshbhatia-amd <210711135+vanshbhatia-amd@users.noreply.github.com> Co-authored-by: vanshbhatia-amd <210711135+vanshbhatia-amd@users.noreply.github.com>
vanshbhatia-amd committed
Jul 14, 2026 at 01:16 UTC
31be872f5596fa9b68218edcf7fa22e1a3b7ca5c
1 file changed
+32
-16
vllm/model_executor/layers/mamba/ops/configs/selective_state_update/headdim=64,dstate=128,device_name=AMD_Instinct_MI355_OAM,cache_dtype=float16.json
+32
-16
@@ -1,35 +1,51 @@
1
{
2
"triton_version": "3.6.0",
3
- "10": {
3
+ "128": {
4
"BLOCK_SIZE_M": 16,
5
"num_warps": 4
6
},
7
- "80": {
8
- "BLOCK_SIZE_M": 16,
7
+ "256": {
8
+ "BLOCK_SIZE_M": 32,
9
"num_warps": 4
10
},
11
- "160": {
12
- "BLOCK_SIZE_M": 16,
13
- "num_warps": 4
11
+ "1024": {
12
+ "BLOCK_SIZE_M": 8,
13
+ "num_warps": 1
14
},
15
- "320": {
15
+ "2048": {
16
"BLOCK_SIZE_M": 16,
17
- "num_warps": 4
17
+ "num_warps": 2
18
},
19
- "640": {
20
- "BLOCK_SIZE_M": 32,
21
- "num_warps": 4
22
- },
23
- "1280": {
19
+ "4096": {
20
"BLOCK_SIZE_M": 16,
21
"num_warps": 1
22
},
27
- "2560": {
23
+ "8192": {
24
"BLOCK_SIZE_M": 16,
25
"num_warps": 1
26
},
31
- "5120": {
27
+ "16384": {
28
"BLOCK_SIZE_M": 16,
33
- "num_warps": 1
29
+ "num_warps": 2
30
+ },
31
+ "32768": {
32
+ "BLOCK_SIZE_M": 16,
33
+ "num_warps": 2
34
+ },
35
+ "65536": {
36
+ "BLOCK_SIZE_M": 16,
37
+ "num_warps": 2
38
+ },
39
+ "131072": {
40
+ "BLOCK_SIZE_M": 64,
41
+ "num_warps": 2
42
+ },
43
+ "196608": {
44
+ "BLOCK_SIZE_M": 16,
45
+ "num_warps": 2
46
+ },
47
+ "262144": {
48
+ "BLOCK_SIZE_M": 16,
49
+ "num_warps": 2
50
}
51
}
\ No newline at end of file