[ROCm] Retune MI355 selective_state_update float32 config on the unified effective_batch grid (#48373)

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 14:26 UTC 7ffb98e248c03fc65c38472aac8ed2ee577c09a9
1 file changed +35 -19
vllm/model_executor/layers/mamba/ops/configs/selective_state_update/headdim=64,dstate=128,device_name=AMD_Instinct_MI355_OAM,cache_dtype=float32.json
+35 -19
@@ -1,35 +1,51 @@
1 {
2 "triton_version": "3.6.0",
3 - "10": {
4 - "BLOCK_SIZE_M": 8,
5 - "num_warps": 4
6 - },
7 - "80": {
8 - "BLOCK_SIZE_M": 8,
9 - "num_warps": 4
10 - },
11 - "160": {
3 + "128": {
4 "BLOCK_SIZE_M": 16,
5 "num_warps": 4
6 },
15 - "320": {
7 + "256": {
8 "BLOCK_SIZE_M": 16,
9 "num_warps": 4
10 },
19 - "640": {
20 - "BLOCK_SIZE_M": 16,
21 - "num_warps": 4
11 + "1024": {
12 + "BLOCK_SIZE_M": 8,
13 + "num_warps": 1
14 },
23 - "1280": {
15 + "2048": {
16 + "BLOCK_SIZE_M": 8,
17 + "num_warps": 1
18 + },
19 + "4096": {
20 "BLOCK_SIZE_M": 16,
21 + "num_warps": 8
22 + },
23 + "8192": {
24 + "BLOCK_SIZE_M": 8,
25 "num_warps": 2
26 },
27 - "2560": {
28 - "BLOCK_SIZE_M": 32,
29 - "num_warps": 4
27 + "16384": {
28 + "BLOCK_SIZE_M": 8,
29 + "num_warps": 2
30 },
31 - "5120": {
32 - "BLOCK_SIZE_M": 32,
31 + "32768": {
32 + "BLOCK_SIZE_M": 64,
33 "num_warps": 4
34 + },
35 + "65536": {
36 + "BLOCK_SIZE_M": 64,
37 + "num_warps": 2
38 + },
39 + "131072": {
40 + "BLOCK_SIZE_M": 64,
41 + "num_warps": 2
42 + },
43 + "196608": {
44 + "BLOCK_SIZE_M": 64,
45 + "num_warps": 2
46 + },
47 + "262144": {
48 + "BLOCK_SIZE_M": 64,
49 + "num_warps": 2
50 }
51 }
\ No newline at end of file