Skip to content

FP8 MoE implementation, take 2 #3954

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
pcmoritz wants to merge 12 commits into from
Closed

FP8 MoE implementation, take 2 #3954

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
55675b4
simplify code
pcmoritz Apr 8, 2024
d3dcb8d
update
pcmoritz Apr 8, 2024
1b1aec1
Add split k
pcmoritz Apr 8, 2024
5631741
update
pcmoritz Apr 8, 2024
bdb8644
update
pcmoritz Apr 9, 2024
3d791f3
fp8
pcmoritz Apr 9, 2024
7270baf
update
pcmoritz Apr 9, 2024
efa49e0
add back scales
pcmoritz Apr 9, 2024
2b745df
update
pcmoritz Apr 9, 2024
58d4091
h100 tp4
pcmoritz Apr 10, 2024
1e692f7
fix
pcmoritz Apr 10, 2024
fc135d4
fix the fix
pcmoritz Apr 10, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
21 changes: 15 additions & 6 deletions benchmarks/kernels/benchmark_mixtral_moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,7 @@ def run_grid(bs, method):
d_model = 4096
num_total_experts = 8
top_k = 2
tp_size = 2
tp_size = 4
model_intermediate_size = 14336
num_layers = 32
num_calls = 100
Expand All @@ -48,13 +48,13 @@ def run_grid(bs, method):
for block_size_n in [32, 64, 128, 256]:
for block_size_m in BLOCK_SIZES_M:
for block_size_k in [64, 128, 256]:
for group_size_m in [1, 16, 32, 64]:
for split_k_size in [1, 2, 4]:
for num_warps in [4, 8]:
configs.append({
"BLOCK_SIZE_M": block_size_m,
"BLOCK_SIZE_N": block_size_n,
"BLOCK_SIZE_K": block_size_k,
"GROUP_SIZE_M": group_size_m,
"SPLIT_K_SIZE": split_k_size,
"num_warps": num_warps,
"num_stages": 4,
})
Expand Down Expand Up @@ -134,20 +134,27 @@ def run_timing(num_calls: int, bs: int, d_model: int, num_total_experts: int,
hidden_states = torch.rand(
(bs, d_model),
device="cuda:0",
dtype=torch.bfloat16,
dtype=torch.float16,
)

ws = torch.rand(
(num_total_experts, 2 * shard_intermediate_size, d_model),
device=hidden_states.device,
dtype=hidden_states.dtype,
)
).to(torch.float8_e4m3fn)

w2s = torch.rand(
(num_total_experts, d_model, shard_intermediate_size),
device=hidden_states.device,
dtype=hidden_states.dtype,
)
).to(torch.float8_e4m3fn)

s = torch.rand(d_model,
device=hidden_states.device,
dtype=torch.float16)
s2 = torch.rand(shard_intermediate_size,
device=hidden_states.device,
dtype=torch.float16)

gating_output = F.softmax(torch.rand(
(num_calls, bs, num_total_experts),
Expand All @@ -165,6 +172,8 @@ def run_timing(num_calls: int, bs: int, d_model: int, num_total_experts: int,
hidden_states=hidden_states,
w1=ws,
w2=w2s,
s=s,
s2=s2,
gating_output=gating_output[i],
topk=2,
renormalize=True,
Expand Down
114 changes: 57 additions & 57 deletions ...model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,145 +1,145 @@
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"SPLIT_K_SIZE": 2,
"num_warps": 8,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"SPLIT_K_SIZE": 2,
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"SPLIT_K_SIZE": 4,
"num_warps": 8,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K_SIZE": 2,
"num_warps": 4,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"SPLIT_K_SIZE": 2,
"num_warps": 4,
"num_stages": 4
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K_SIZE": 4,
"num_warps": 4,
"num_stages": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"SPLIT_K_SIZE": 1,
"num_warps": 4,
"num_stages": 4
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"SPLIT_K_SIZE": 2,
"num_warps": 4,
"num_stages": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"SPLIT_K_SIZE": 1,
"num_warps": 8,
"num_stages": 4
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"SPLIT_K_SIZE": 2,
"num_warps": 4,
"num_stages": 4
},
"256": {
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"SPLIT_K_SIZE": 1,
"num_warps": 8,
"num_stages": 4
},
"512": {
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"SPLIT_K_SIZE": 1,
"num_warps": 4,
"num_stages": 4
},
"1024": {
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"BLOCK_SIZE_K": 128,
"SPLIT_K_SIZE": 1,
"num_warps": 4,
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"SPLIT_K_SIZE": 1,
"num_warps": 8,
"num_stages": 4
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"BLOCK_SIZE_K": 128,
"SPLIT_K_SIZE": 1,
"num_warps": 8,
"num_stages": 4
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"BLOCK_SIZE_K": 128,
"SPLIT_K_SIZE": 1,
"num_warps": 8,
"num_stages": 4
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"BLOCK_SIZE_K": 128,
"SPLIT_K_SIZE": 1,
"num_warps": 8,
"num_stages": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"BLOCK_SIZE_K": 128,
"SPLIT_K_SIZE": 1,
"num_warps": 8,
"num_stages": 4
}
Expand Down
Loading