Skip to content

[ROCM] enable aiter fused moe kernel for llama4 bf16 checkpoints #16674

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 4 commits into from
Apr 17, 2025
Merged

[ROCM] enable aiter fused moe kernel for llama4 bf16 checkpoints #16674

Changes from 2 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
c51662d
enable aiter fused moe bf16 for llama4
sijiac Apr 15, 2025
5f057ca
fix lint
sijiac Apr 15, 2025
8dbbb11
add assert
sijiac Apr 16, 2025
2ee3ec7
lint
sijiac Apr 16, 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
11 changes: 11 additions & 0 deletions vllm/model_executor/layers/fused_moe/rocm_aiter_fused_moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,7 @@ def rocm_aiter_fused_experts(
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
use_fp8_w8a8: bool = False,
apply_router_weight_on_input: bool = False,
w1_scale: Optional[torch.Tensor] = None,
w2_scale: Optional[torch.Tensor] = None,
block_shape: Optional[List[int]] = None,
Expand All @@ -39,6 +40,16 @@ def rocm_aiter_fused_experts(
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
per_token_group_quant_fp8)

if apply_router_weight_on_input:
_, topk = topk_weights.shape
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

can we assert topk_weights.dim() == 2?

assert (
topk == 1
), "Only support topk=1 when `apply_router_weight_on_input` is True"

hidden_states = hidden_states * topk_weights.to(hidden_states.dtype)
topk_ids = topk_ids.to(torch.int32)
topk_weights = torch.ones_like(topk_weights, dtype=torch.float32)
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Does AITER require fp32 weight?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

yes, the passed-in topk_weights must in fp32 dtype, otherwise, it will have numeric issues


if envs.VLLM_ROCM_USE_AITER_FP8_BLOCK_SCALED_MOE and use_fp8_w8a8:
assert w1_scale is not None
assert w2_scale is not None
Expand Down