A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 8, 10:00 UTC
Add config file for fused MoE for Nemotron (TP4, B200) (#34411)
Signed-off-by: Daniel Serebrenik <daserebrenik@nvidia.com>
danisereb committed
Feb 12, 2026 at 16:09 UTC
dea63512bb9bdf7521d591546c52138d9d79e8ce
1 file changed
+59
vllm/model_executor/layers/fused_moe/configs/E=512,N=672,device_name=NVIDIA_B200.json
new
+59
@@ -0,0 +1,59 @@
1
+{
2
+ "triton_version": "3.6.0",
3
+ "8": {
4
+ "BLOCK_SIZE_M": 16,
5
+ "BLOCK_SIZE_N": 128,
6
+ "BLOCK_SIZE_K": 64,
7
+ "GROUP_SIZE_M": 32,
8
+ "num_warps": 4,
9
+ "num_stages": 4
10
+ },
11
+ "32": {
12
+ "BLOCK_SIZE_M": 16,
13
+ "BLOCK_SIZE_N": 128,
14
+ "BLOCK_SIZE_K": 64,
15
+ "GROUP_SIZE_M": 64,
16
+ "num_warps": 4,
17
+ "num_stages": 3
18
+ },
19
+ "64": {
20
+ "BLOCK_SIZE_M": 16,
21
+ "BLOCK_SIZE_N": 128,
22
+ "BLOCK_SIZE_K": 64,
23
+ "GROUP_SIZE_M": 32,
24
+ "num_warps": 4,
25
+ "num_stages": 4
26
+ },
27
+ "128": {
28
+ "BLOCK_SIZE_M": 16,
29
+ "BLOCK_SIZE_N": 128,
30
+ "BLOCK_SIZE_K": 64,
31
+ "GROUP_SIZE_M": 16,
32
+ "num_warps": 4,
33
+ "num_stages": 5
34
+ },
35
+ "256": {
36
+ "BLOCK_SIZE_M": 16,
37
+ "BLOCK_SIZE_N": 128,
38
+ "BLOCK_SIZE_K": 64,
39
+ "GROUP_SIZE_M": 64,
40
+ "num_warps": 4,
41
+ "num_stages": 5
42
+ },
43
+ "512": {
44
+ "BLOCK_SIZE_M": 16,
45
+ "BLOCK_SIZE_N": 128,
46
+ "BLOCK_SIZE_K": 64,
47
+ "GROUP_SIZE_M": 16,
48
+ "num_warps": 4,
49
+ "num_stages": 5
50
+ },
51
+ "1024": {
52
+ "BLOCK_SIZE_M": 16,
53
+ "BLOCK_SIZE_N": 128,
54
+ "BLOCK_SIZE_K": 64,
55
+ "GROUP_SIZE_M": 16,
56
+ "num_warps": 4,
57
+ "num_stages": 2
58
+ }
59
+}