Skip to content

[BugFix][quant] Fix per_channel_cast_fused expand tails mishandle bug #27

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
GenTang wants to merge 3 commits into
base: main
Choose a base branch
from
Open

[BugFix][quant] Fix per_channel_cast_fused expand tails mishandle bug #27

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
bff98f4
add test case to cover expand tails issue
GenTang Jul 28, 2026
8655a64
fix expand tails bug
GenTang Jul 28, 2026
bd36724
add store bounds guard
GenTang Jul 28, 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
10 changes: 6 additions & 4 deletions tests/quant/test_per_channel_cast_fused.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,7 @@
os.environ['TILELANG_PRINT_ON_COMPILATION'] = '0'


def generate_test_data(params):
def generate_test_data(params, alignment=128):
num_send_tokens = params['num_send_tokens']
num_topk = params['num_topk']
num_experts = params['num_experts']
Expand All @@ -28,7 +28,7 @@ def generate_test_data(params):
topk_idx = generate_topk_idx(params)
num_tokens = topk_idx.shape[0]
_, pos_to_token, _, token_topk_to_pos, _, _, _, _ = (
tile_kernels.moe.get_fused_mapping(topk_idx, num_experts, 0, 128)
tile_kernels.moe.get_fused_mapping(topk_idx, num_experts, 0, alignment)
)
x = torch.randn((num_tokens, hidden), dtype=torch.bfloat16, device='cuda')
x = tile_kernels.moe.expand_to_fused(x, token_topk_to_pos, pos_to_token)
Expand Down Expand Up @@ -74,8 +74,10 @@ def generate_test_params(is_benchmark: bool) -> list[dict]:


@pytest.mark.parametrize('params', generate_test_params(is_benchmark=False), ids=make_param_id)
def test_per_channel_cast_fused(params):
_, _, _, func, func_ref = generate_test_data(params)
@pytest.mark.parametrize('alignment', [16, 128])
# use alignment=16 to cover expand tails issues.
def test_per_channel_cast_fused(params, alignment):
_, _, _, func, func_ref = generate_test_data(params, alignment)

x_fp8, x_fp8_sf = func()
x_fp8_ref, x_fp8_sf_ref = func_ref()
Expand Down
14 changes: 10 additions & 4 deletions tile_kernels/quant/per_channel_cast_fused_kernel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -71,8 +71,12 @@ def per_channel_cast_fused_kernel(
T.assume(num_tokens_out % 128 == 0 or (with_expand and num_tokens_out % 16 == 0))
if with_expand:
tmp = T.alloc_var(T.int32)
if k_id < VEC_M:
tmp = pos_to_token[k_id + m_offset]
# Invalid tail rows remain -1 and follow the existing zero-fill path,
# so they do not participate in the amax reduction.
tmp = -1
row = k_id + m_offset
if k_id < VEC_M and row < num_tokens_out:
tmp = pos_to_token[row]
Comment thread
GenTang marked this conversation as resolved.

for i in T.serial(VEC_M):
pos_to_token_local[i] = T.shfl_sync(tmp, i)
Expand Down Expand Up @@ -128,8 +132,10 @@ def per_channel_cast_fused_kernel(
out_local[j] = in_local[j] * sf_invs_local[i] * amax_local[j]
else:
out_local[j] = in_local[j] * amax_local[j]
for j in T.vectorized(VEC_K):
out[i + m_offset, j + k_offset] = out_local[j]
out_row = i + m_offset
if out_row < num_tokens_out:
for j in T.vectorized(VEC_K):
out[out_row, j + k_offset] = out_local[j]

return per_channel_cast_fused_kernel

Expand Down