Skip to content

Add Python-side guardrail for HybridEP InfiniBand limit and rename seq_len #4094

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
janEbert merged 3 commits into from
May 8, 2026
Merged

Add Python-side guardrail for HybridEP InfiniBand limit and rename seq_len #4094

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
f7fbe84
fix(moe): add HybridEP hardware limit guardrail and rename seq_len
Shreyas-S-809 Apr 27, 2026
ff29c60
fix(moe): add HybridEP hardware limit guardrail
Shreyas-S-809 Apr 27, 2026
7548fb8
fix lexical issues
Shreyas-S-809 Apr 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
21 changes: 17 additions & 4 deletions megatron/core/transformer/moe/fused_a2a.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -278,7 +278,7 @@ def set_deepep_num_sms(num_sms):
def init_hybrid_ep_buffer(
group: torch.distributed.ProcessGroup,
hidden_dim: int,
seq_len: int,
num_tokens: int,
num_local_experts: int,
num_sms_dispatch_api: int,
num_sms_combine_api: int,
Expand Down Expand Up @@ -313,7 +313,7 @@ def init_hybrid_ep_buffer(
_hybrid_ep_buffer = HybridEPBuffer(
group=group,
hidden_dim=hidden_dim,
max_num_of_tokens_per_rank=seq_len,
max_num_of_tokens_per_rank=num_tokens,
num_local_experts=num_local_experts,
use_fp8=fp8_dispatch,
num_sms_dispatch_api=num_sms_dispatch_api,
Expand Down Expand Up @@ -351,12 +351,25 @@ def forward(
Forward pass of fused dispatch of the HybridEP backend
'''
if _hybrid_ep_buffer is None:
seq_len, hidden_dim = x.shape[-2:]
num_tokens, hidden_dim = x.shape[-2:]

# --- Hardware Limit Guardrail ---
# DeepEP calculates tx_depth = 3 * num_tokens + 1.
Comment thread
janEbert marked this conversation as resolved.
# InfiniBand strictly asserts tx_depth < 65536.
tx_depth = 3 * num_tokens + 1
if tx_depth >= 65536:
raise ValueError(
f"HybridEP RDMA Queue Pair depth ({tx_depth}) exceeds the InfiniBand "
f"hardware limit of 65535. This occurs because the total tokens per rank "
f"({num_tokens}) too high. Reduce sequence length or micro-batch size, "
f"or increase Tensor Parallelism (TP) / Context Parallelism (CP) to reduce "
f"the number of tokens processed per rank."
)
fp8_dispatch = False # Currently, we do not support fp8 dispatch
init_hybrid_ep_buffer(
group,
hidden_dim,
seq_len,
num_tokens,
num_local_experts,
num_sms_dispatch_api,
num_sms_combine_api,
Expand Down
Loading