Skip to content

Commit

Permalink
[Kernel] Re-tune Mixtral MoE configurations for FP8 on H100 (vllm-pro…
Browse files Browse the repository at this point in the history 
…ject#5238)
  • Loading branch information
@pcmoritz
pcmoritz authored and jimpang committed Jul 24, 2024
1 parent 20d60f7 commit 1161cfa
Show file tree
Hide file tree
Showing 3 changed files with 85 additions and 84 deletions.
3 changes: 2 additions & 1 deletion benchmarks/kernels/benchmark_moe.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -255,7 +255,8 @@ def main(args: argparse.Namespace):

if args.batch_size is None:
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:
batch_sizes = [args.batch_size]
Expand Down
104 changes: 52 additions & 52 deletions ...r/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=float8.json
View file
Original file line number Diff line number Diff line change
@@ -1,121 +1,121 @@
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 5
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 2
"num_warps": 4,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 5
"num_warps": 4,
"num_stages": 3
},
"24": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
"num_stages": 3
},
"32": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
"num_stages": 3
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 4
"num_stages": 3
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
"num_stages": 3
},
"96": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
"num_stages": 3
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 5
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 3
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
Expand Down
62 changes: 31 additions & 31 deletions ...r/layers/fused_moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=float8.json
View file
Original file line number Diff line number Diff line change
@@ -1,67 +1,67 @@
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 5
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"4": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
"num_stages": 5
},
"8": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
"num_stages": 5
},
"24": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
"num_stages": 3
},
"32": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
"num_stages": 4
},
"64": {
"BLOCK_SIZE_M": 64,
Expand All @@ -81,19 +81,19 @@
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
"num_stages": 4
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 5
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 3
},
"512": {
"BLOCK_SIZE_M": 128,
Expand All @@ -107,7 +107,7 @@
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
Expand Down

0 comments on commit 1161cfa

Please sign in to comment.