Skip to content

[BugFix] Fix accuray problems with deepseek in situation of ep=1, etp>1 #863

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

Closed
wants to merge 3 commits into from
Closed

[BugFix] Fix accuray problems with deepseek in situation of ep=1, etp>1 #863

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
a990949
fix etp rank related accuracy problem
whx-sjtu May 14, 2025
f05ee46
fix: fix deepseek accuracy when ep_size=1
zzzzwwjj May 16, 2025
b23c5fe
fix ci problems
whx-sjtu May 20, 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
35 changes: 10 additions & 25 deletions vllm_ascend/ops/fused_moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -198,6 +198,7 @@ def fused_experts(
num_experts = w1.shape[0]
dtype = hidden_states.dtype
device = hidden_states.device
topk_weights = topk_weights.to(dtype)
# assert dtype in [torch.float32, torch.float16, torch.bfloat16
# ], "Only float32, float16, and bfloat16 are supported"

Expand Down Expand Up @@ -615,32 +616,16 @@ def __init__(
self.expert_map = None
self.activation = activation

if self.ep_size > 1:
# Create a tensor of size num_experts filled with -1
self.local_num_experts, self.expert_map = determine_expert_map(
self.ep_size,
get_ep_group().rank_in_group, self.global_num_experts)
if vllm_version_is("0.8.5") or vllm_version_is("0.8.5.post1"):
self.tp_rank = get_etp_group().rank_in_group
self.ep_rank = get_ep_group().rank_in_group
else:
self.moe_parallel_config.tp_rank = get_etp_group(
).rank_in_group
self.moe_parallel_config.ep_rank = get_ep_group().rank_in_group

# Create a tensor of size num_experts filled with -1
self.local_num_experts, self.expert_map = determine_expert_map(
self.ep_size,
get_ep_group().rank_in_group, self.global_num_experts)
if vllm_version_is("0.8.5") or vllm_version_is("0.8.5.post1"):
Copy link
Collaborator

Choose a reason for hiding this comment

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

Don't need anymore. #959

self.tp_rank = get_etp_group().rank_in_group
self.ep_rank = get_ep_group().rank_in_group
else:
# Adjust TP size for DP attention
# haven't test its functionality yet, may remove in the future
if vllm_version_is("0.8.5") or vllm_version_is("0.8.5.post1"):
self.tp_rank = self.tp_size * self.dp_rank
self.ep_rank = 0
self.tp_size = self.tp_size * self.dp_size
self.ep_size = 1
else:
self.moe_parallel_config.tp_rank = self.tp_size * self.dp_rank
self.moe_parallel_config.ep_rank = 0
self.moe_parallel_config.tp_size = self.tp_size * self.dp_size
self.moe_parallel_config.ep_size = 1
self.moe_parallel_config.tp_rank = get_etp_group().rank_in_group
self.moe_parallel_config.ep_rank = get_ep_group().rank_in_group

self.local_num_experts, self.expert_map = (self.global_num_experts,
None)
Expand Down
1 change: 1 addition & 0 deletions vllm_ascend/quantization/w8a8_dynamic.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -342,6 +342,7 @@ def fused_experts(hidden_states: torch.Tensor,
num_experts = w1.shape[0]
dtype = hidden_states.dtype
device = hidden_states.device
topk_weights = topk_weights.to(dtype)

if expert_map is not None:
# Generate token indices and flatten
Expand Down