Skip to content

[OP] support deepgeem for sm103 #7073

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
BingooYang wants to merge 3 commits into
base: develop
Choose a base branch
from
+6 −3
Open

[OP] support deepgeem for sm103 #7073

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
d8fdf9d
support deepgeem for sm103
BingooYang Mar 30, 2026
703c828
add assert
BingooYang Mar 30, 2026
5a2b811
modify code style
BingooYang Mar 30, 2026
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
5 changes: 4 additions & 1 deletion fastdeploy/model_executor/layers/quantization/block_wise_fp8.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -125,7 +125,10 @@ def deep_gemm_fp8_gemm_nt(
layer_output_size: int,
bias: paddle.Tensor = None,
):
if get_sm_version() == 100 and current_platform.is_cuda():
sm_version = get_sm_version()
if sm_version >= 100:
assert x_scale_tensor.dtype == paddle.uint8, "For sm100+, x_scale_tensor must be uint8 dtype."
if sm_version == 100 and current_platform.is_cuda():
# disable_ue8m0_cast is default False for SM100
fp8_gemm_nt(
(x, x_scale_tensor),
Expand Down
4 changes: 2 additions & 2 deletions fastdeploy/model_executor/layers/quantization/fp8_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,7 +65,7 @@ def load_deep_gemm():
"""

if current_platform.is_cuda():
if get_sm_version() == 100:
if get_sm_version() >= 100:
# SM100 should use PFCC DeepGemm
paddle.compat.enable_torch_proxy(scope={"deep_gemm"})
try:
Expand Down Expand Up @@ -245,7 +245,7 @@ def fused_stack_transpose_quant(expert_weight_list, use_ue8m0=False):
# Blackwell (SM100) GPUs require pow2_scale quantization.
# Guard with is_cuda() so non-CUDA environments do not call into
# paddle.device.cuda.* and cause a crash.
use_pow2_scale = current_platform.is_cuda() and get_sm_version() == 100
use_pow2_scale = current_platform.is_cuda() and get_sm_version() >= 100

w, scale = paddlefleet_ops.fuse_stack_transpose_fp8_quant(
expert_weight_list,
Expand Down
Loading