Skip to content

Fixing the shape to use in padding calculation #464

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
gshtras merged 5 commits into from
Mar 7, 2025
Merged

Fixing the shape to use in padding calculation #464

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
5267f4f
Fixing the shape to use in padding calculation
gshtras Mar 7, 2025
6dd8d46
Merge branch 'main' into moe_padding_shape_fix
gshtras Mar 7, 2025
8cd4763
Assertion on the int8 quantized moe
gshtras Mar 7, 2025
684f5d2
Merge branch 'moe_padding_shape_fix' of github.com:ROCm/vllm into moe…
gshtras Mar 7, 2025
15bd566
Properly testing for padding
gshtras Mar 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
34 changes: 17 additions & 17 deletions tests/kernels/test_moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,14 +54,6 @@ def test_fused_moe(
w2 = torch.randn((e, k, n), device="cuda", dtype=dtype) / 10

score = torch.randn((m, e), device="cuda", dtype=dtype)

# Pad the input if use padding
if envs.VLLM_MOE_PADDING:
w1 = F.pad(w1, (0, 128), "constant", 0)
torch.cuda.empty_cache()
w2 = F.pad(w2, (0, 128), "constant", 0)
torch.cuda.empty_cache()

if ep_size > 1:
local_e = e // ep_size
e_ids = torch.randint(0,
Expand All @@ -75,16 +67,7 @@ def test_fused_moe(
else:
e_map = None

triton_output = fused_moe(a,
w1,
w2,
score,
topk,
global_num_experts=e,
expert_map=e_map,
renormalize=False)
torch_output = torch_moe(a, w1, w2, score, topk, e_map)
torch.testing.assert_close(triton_output, torch_output, atol=2e-2, rtol=0)
iterative_output = iterative_moe(a,
w1,
w2,
Expand All @@ -93,6 +76,23 @@ def test_fused_moe(
global_num_experts=e,
expert_map=e_map,
renormalize=False)
# Pad the input if use padding
if envs.VLLM_MOE_PADDING:
w1 = F.pad(w1, (0, 128), "constant", 0)
torch.cuda.empty_cache()
w2 = F.pad(w2, (0, 128), "constant", 0)
torch.cuda.empty_cache()

triton_output = fused_moe(a,
w1,
w2,
score,
topk,
global_num_experts=e,
expert_map=e_map,
renormalize=False)

torch.testing.assert_close(triton_output, torch_output, atol=2e-2, rtol=0)
torch.testing.assert_close(iterative_output,
torch_output,
atol=1e-2,
Expand Down
4 changes: 3 additions & 1 deletion vllm/model_executor/layers/fused_moe/fused_moe.py
View file
Open in desktop
Comment thread
divakar-amd marked this conversation as resolved.
Original file line number Diff line number Diff line change
Expand Up @@ -719,6 +719,8 @@ def invoke_fused_moe_kernel(A: torch.Tensor,
block_shape is not None and block_shape[1] > 0:
assert B_scale is not None and B_scale.ndim == 3
assert B_zp is None or B_zp.ndim == 3
assert padding_size == 0, "MoE padding is not supported " \
"with GPTQ/AWQ quantization"

fused_moe_kernel_gptq_awq[grid](
A,
Expand Down Expand Up @@ -770,7 +772,7 @@ def invoke_fused_moe_kernel(A: torch.Tensor,
expert_ids,
num_tokens_post_padded,
B.shape[1],
A.shape[1] - padding_size,
B.shape[2] - padding_size,
EM,
topk_ids.numel(),
A.stride(0),
Expand Down