Skip to content

Enable tf32 precision acceleration for h100 #4767

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

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Enable tf32 precision acceleration for h100 #4767

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c0731b4
Fine tune CMF 500x large&medium K shapes to boost flops (#4766)
Aug 25, 2025
68134fa
Enable tf32 precision acceleration for CMF 500x on h100 (#4767)
Aug 25, 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
20 changes: 20 additions & 0 deletions fbgemm_gpu/experimental/gemm/triton_gemm/fp8_gemm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -149,6 +149,16 @@ def get_configs_io_bound() -> List[Config]:
return configs


def dummy_prune_configs(configs, named_args, **kwargs):

M = named_args["M"]
N = named_args["N"]
K = named_args["K"]

logger.info(f"{len(configs)=} {len(configs)=} for {M=} {N=} {K=}")
return configs


MATMUL_CONFIGS: List[Config] = [
# basic configs for compute-bound matmuls
Config(
Expand Down Expand Up @@ -176,6 +186,11 @@ def get_configs_io_bound() -> List[Config]:
num_stages=4,
num_warps=4,
),
Config(
{"BLOCK_M": 64, "BLOCK_N": 128, "BLOCK_K": 256, "SPLIT_K": 1},
num_stages=4,
num_warps=4,
),
Config(
{"BLOCK_M": 128, "BLOCK_N": 128, "BLOCK_K": 32, "SPLIT_K": 1},
num_stages=4,
Expand Down Expand Up @@ -252,6 +267,9 @@ def get_configs_io_bound() -> List[Config]:

@triton.autotune(
configs=MATMUL_CONFIGS,
prune_configs_by={
"early_config_prune": dummy_prune_configs,
},
key=[
"m_key",
"n_key",
Expand Down Expand Up @@ -1227,6 +1245,8 @@ def matmul_fp8_row(
a = a.view(-1, a.size(-1))
# View inputs into proper torch fp8 dtype.
if torch.version.cuda:
# Enable tf32 tensor core to accelerate CUDA
torch.set_float32_matmul_precision("high")
assert a.dtype in (torch.float8_e4m3fn, torch.float8_e5m2)
else:
assert a.dtype in (torch.float8_e4m3fnuz, torch.float8_e5m2fnuz)
Expand Down
Loading