Skip to content
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

[SW-199650] Add HPU fp8 DynamicMOE Op #721

Merged
merged 4 commits into from
Jan 23, 2025
Merged

[SW-199650] Add HPU fp8 DynamicMOE Op #721

Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
8358d87
[SW-199650] Add HPU fp8 DynamicMOE Op
dudilester Jan 6, 2025
cd10ac3
Merge branch 'habana_main' into dev/dlester/dynamic_moe
dudilester Jan 22, 2025
0075169
Merge branch 'habana_main' into dev/dlester/dynamic_moe
michalkuligowski Jan 22, 2025
a7a871a
Merge branch 'habana_main' into dev/dlester/dynamic_moe
dudilester Jan 23, 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
23 changes: 6 additions & 17 deletions vllm/model_executor/layers/fused_moe/layer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -178,9 +178,7 @@ def forward_hpu(
'not supported on HPU')
assert topk_group is None, 'topk_group is not supported on HPU'
if layer is not None:
return layer.hpu_static_fused_moe(x, layer.w13_weight,
layer.w2_weight, router_logits,
top_k)
return layer.hpu_fused_moe(x, router_logits, top_k)

def forward_cpu(
self,
Expand Down Expand Up @@ -300,15 +298,11 @@ def __init__(
self.topk_group = topk_group
self.custom_routing_function = custom_routing_function
if is_hpu:
from vllm_hpu_extension.ops import DynamicFusedMOE, StaticFusedMOE
from vllm_hpu_extension.ops import DynamicFusedMOE
self.hpu_fused_moe = DynamicFusedMOE(self.num_experts)

from vllm.model_executor.layers.quantization.inc import INCConfig
selected_fused_moe = (StaticFusedMOE if isinstance(
quant_config, INCConfig) else DynamicFusedMOE)
self.hpu_static_fused_moe = selected_fused_moe(self.num_experts)
self.scoring_func = scoring_func
self.e_score_correction_bias = e_score_correction_bias

if self.scoring_func != "softmax" and not self.use_grouped_topk:
raise ValueError("Only softmax scoring function is supported for "
"non-grouped topk.")
Expand Down Expand Up @@ -404,10 +398,8 @@ def _load_w13(self,
expert_data.copy_(loaded_weight)

if is_hpu:
from vllm_hpu_extension.ops import StaticFusedMOE
if isinstance(self.hpu_static_fused_moe, StaticFusedMOE):
self.hpu_static_fused_moe.w13_list[expert_id].set_weight(
orig_exp_data)
self.hpu_fused_moe.MoeOp.w13_list[expert_id].set_weight(
orig_exp_data)

def _load_w2(self,
expert_data: torch.Tensor,
Expand All @@ -426,10 +418,7 @@ def _load_w2(self,
# w2, down_proj: Load into only logical weight of w2.
expert_data.copy_(loaded_weight)
if is_hpu:
from vllm_hpu_extension.ops import StaticFusedMOE
if isinstance(self.hpu_static_fused_moe, StaticFusedMOE):
self.hpu_static_fused_moe.w2_list[expert_id].set_weight(
expert_data)
self.hpu_fused_moe.MoeOp.w2_list[expert_id].set_weight(expert_data)

def _load_single_value(self, param: torch.nn.Parameter,
loaded_weight: torch.Tensor, expert_id: int):
Expand Down
Loading