Skip to content
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

[block_wise w8a8]Add block_wise w8w8 tunning scripts #3873

Open
wants to merge 5 commits into
base: main
Choose a base branch
from
+455 −3
Open

[block_wise w8a8]Add block_wise w8w8 tunning scripts #3873

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
ceab14e
[daily]fused moe trition support int8 w8a8
coolhok Feb 25, 2025
bf22891
[fix] name config err
coolhok Feb 26, 2025
adcff7b
[tune]add block wise w8a8 script
coolhok Feb 26, 2025
8def65b
[daily]lint failed
coolhok Feb 26, 2025
fa0f2fe
Merge branch 'main' into worship_int8_w8w8
coolhok Feb 26, 2025
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
44 changes: 41 additions & 3 deletions benchmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,10 +42,11 @@ def benchmark_config(
dtype: torch.dtype,
use_fp8_w8a8: bool,
use_int8_w8a16: bool,
use_int8_w8a8: bool,
block_shape: List[int] = None,
num_iters: int = 100,
) -> float:
init_dtype = torch.float16 if use_fp8_w8a8 else dtype
init_dtype = torch.float16 if (use_fp8_w8a8 or use_int8_w8a8) else dtype
x = torch.randn(num_tokens, hidden_size, dtype=dtype)
if use_int8_w8a16:
w1 = torch.randint(
Expand Down Expand Up @@ -107,6 +108,27 @@ def benchmark_config(

w1 = w1.to(torch.float8_e4m3fnuz if _is_hip_ else torch.float8_e4m3fn)
w2 = w2.to(torch.float8_e4m3fnuz if _is_hip_ else torch.float8_e4m3fn)
if use_int8_w8a8:
if block_shape is None:
w1_scale = torch.randn(num_experts, dtype=torch.float32)
w2_scale = torch.randn(num_experts, dtype=torch.float32)
a1_scale = torch.randn(1, dtype=torch.float32)
a2_scale = torch.randn(1, dtype=torch.float32)
else:
block_n, block_k = block_shape[0], block_shape[1]
n_tiles_w1 = (shard_intermediate_size + block_n - 1) // block_n
n_tiles_w2 = (hidden_size + block_n - 1) // block_n
k_tiles_w1 = (hidden_size + block_k - 1) // block_k
k_tiles_w2 = (shard_intermediate_size // 2 + block_k - 1) // block_k
w1_scale = torch.rand(
(num_experts, n_tiles_w1, k_tiles_w1), dtype=torch.float32
)
w2_scale = torch.rand(
(num_experts, n_tiles_w2, k_tiles_w2), dtype=torch.float32
)

w1 = w1.to(torch.int8)
w2 = w2.to(torch.int8)

input_gating = torch.empty(num_tokens, num_experts, dtype=torch.float32)

Expand All @@ -127,6 +149,7 @@ def run():
inplace=True,
use_fp8_w8a8=use_fp8_w8a8,
use_int8_w8a16=use_int8_w8a16,
use_int8_w8a8=use_int8_w8a8,
w1_scale=w1_scale,
w2_scale=w2_scale,
a1_scale=a1_scale,
Expand Down Expand Up @@ -236,6 +259,7 @@ def benchmark(
dtype: torch.dtype,
use_fp8_w8a8: bool,
use_int8_w8a16: bool,
use_int8_w8a8: bool,
block_shape: List[int],
) -> Tuple[Dict[str, int], float]:
torch.cuda.manual_seed_all(0)
Expand Down Expand Up @@ -271,6 +295,7 @@ def benchmark(
dtype,
use_fp8_w8a8,
use_int8_w8a16,
use_int8_w8a8,
block_shape,
)
return config, kernel_time
Expand All @@ -285,6 +310,7 @@ def tune(
dtype: torch.dtype,
use_fp8_w8a8: bool,
use_int8_w8a16: bool,
use_int8_w8a8: bool,
block_shape: List[int],
search_space: List[Dict[str, int]],
) -> Dict[str, int]:
Expand All @@ -302,6 +328,7 @@ def tune(
dtype,
use_fp8_w8a8,
use_int8_w8a16,
use_int8_w8a8,
block_shape,
num_iters=10,
)
Expand Down Expand Up @@ -341,10 +368,14 @@ def save_configs(
dtype: torch.dtype,
use_fp8_w8a8: bool,
use_int8_w8a16: bool,
use_int8_w8a8: bool,
block_shape: List[int],
) -> None:
dtype_str = get_config_dtype_str(
dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8
dtype,
use_int8_w8a16=use_int8_w8a16,
use_fp8_w8a8=use_fp8_w8a8,
use_int8_w8a8=use_int8_w8a8,
)

# NOTE(woosuk): The current naming convention uses w2.shape[2], which
Expand Down Expand Up @@ -397,6 +428,7 @@ def main(args: argparse.Namespace):
dtype = config.torch_dtype
use_fp8_w8a8 = args.dtype == "fp8_w8a8"
use_int8_w8a16 = args.dtype == "int8_w8a16"
use_int8_w8a8 = args.dtype == "int8_w8a8"
block_shape = None
if (
hasattr(config, "quantization_config")
Expand Down Expand Up @@ -468,6 +500,7 @@ def _distribute(method: str, inputs: List[Any]) -> List[Any]:
dtype,
use_fp8_w8a8,
use_int8_w8a16,
use_int8_w8a8,
block_shape,
search_space,
)
Expand All @@ -486,6 +519,7 @@ def _distribute(method: str, inputs: List[Any]) -> List[Any]:
dtype,
use_fp8_w8a8,
use_int8_w8a16,
use_int8_w8a8,
block_shape,
)
end = time.time()
Expand All @@ -503,6 +537,7 @@ def _distribute(method: str, inputs: List[Any]) -> List[Any]:
dtype,
use_fp8_w8a8,
use_int8_w8a16,
use_int8_w8a8,
block_shape,
)
for batch_size in batch_sizes
Expand All @@ -521,7 +556,10 @@ def _distribute(method: str, inputs: List[Any]) -> List[Any]:
)
parser.add_argument("--tp-size", "-tp", type=int, default=2)
parser.add_argument(
"--dtype", type=str, choices=["auto", "fp8_w8a8", "int8_w8a16"], default="auto"
"--dtype",
type=str,
choices=["auto", "fp8_w8a8", "int8_w8a16", "int8_w8a8"],
default="auto",
)
parser.add_argument("--seed", type=int, default=0)
parser.add_argument("--batch-size", type=int, required=False)
Expand Down
Loading