Skip to content

[Kernel][Triton][AMD] Use block size heuristic for avg 2.8x speedup for int8 models #11698

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

Merged
merged 3 commits into from
Jan 8, 2025
Merged

[Kernel][Triton][AMD] Use block size heuristic for avg 2.8x speedup for int8 models #11698

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
5675c6b
Change defeault block size for triton_scaled_mm to 128 for 4-5x speedup
rasmith Jan 2, 2025
a45f569
Use heuristic based on cutlass_gemm_sm90_int8_dispatch
rasmith Jan 3, 2025
eb8126e
Use heuristic to pick block size for better performance across input/…
rasmith Jan 7, 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
17 changes: 16 additions & 1 deletion vllm/model_executor/layers/quantization/compressed_tensors/triton_scaled_mm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -128,7 +128,8 @@ def triton_scaled_mm(input: torch.Tensor,
bias: Optional[torch.Tensor] = None,
block_size_m: int = 32,
block_size_n: int = 32,
block_size_k: int = 32) -> torch.Tensor:
block_size_k: int = 32,
use_heuristic=True) -> torch.Tensor:
M, K = input.shape
N = weight.shape[1]

Expand All @@ -152,6 +153,20 @@ def triton_scaled_mm(input: torch.Tensor,

has_scalar = lambda x: x.shape[0] == 1 and x.shape[1] == 1

if use_heuristic:
is_small_N = N < 8192
next_power_of_2_M = max(32, triton.next_power_of_2(M))
if next_power_of_2_M <= 32:
tile_shape = (64, 64, 256) if is_small_N else (64, 128, 256)
elif next_power_of_2_M <= 64:
tile_shape = (64, 64, 256)
elif next_power_of_2_M <= 128:
tile_shape = (64, 128, 128)
else:
tile_shape = (128, 128, 128)

block_size_m, block_size_n, block_size_k = tile_shape

block_size_sa = 1 if has_scalar(scale_a) else block_size_m
block_size_sb = 1 if has_scalar(scale_b) else block_size_n

Expand Down
Loading