Skip to content

Support EP24 for internode kernels #432

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
LyricZhao merged 2 commits into from
Sep 26, 2025
Merged

Support EP24 for internode kernels #432

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
3c12569
Support EP24 for internode kernels.
sphish Sep 26, 2025
ec7845a
Skip check for round_scale test
sphish Sep 26, 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
1 change: 1 addition & 0 deletions csrc/kernels/launch.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,6 +62,7 @@ cfg.dynamicSmemBytes = smem_size;
#define SWITCH_RDMA_RANKS(case_macro) \
switch (num_ranks / NUM_MAX_NVL_PEERS) { \
case 2: case_macro(2); \
case 3: case_macro(3); \
case 4: case_macro(4); \
case 6: case_macro(6); \
case 8: case_macro(8); \
Expand Down
2 changes: 1 addition & 1 deletion deep_ep/buffer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -234,7 +234,7 @@ def get_dispatch_config(num_ranks: int) -> Config:
4: Config(Buffer.num_sms, 6, 256, 6, 128),
8: Config(Buffer.num_sms, 6, 256, 6, 128),
16: Config(Buffer.num_sms, 36, 288, 20, 128),
24: Config(Buffer.num_sms, 8, 288, 32, 128),
24: Config(Buffer.num_sms, 32, 288, 8, 128),
32: Config(Buffer.num_sms, 32, 288, 8, 128),
48: Config(Buffer.num_sms, 32, 288, 8, 128),
64: Config(Buffer.num_sms, 32, 288, 8, 128),
Expand Down
2 changes: 1 addition & 1 deletion tests/test_internode.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,7 +92,7 @@ def test_main(args: argparse.Namespace, num_sms: int,
time.sleep(1)

# Config
rdma_buffer_size, nvl_buffer_size = 128, (720 if num_ranks in (48, 96, 144, 160) else 512)
rdma_buffer_size, nvl_buffer_size = 128, (720 if num_ranks in (24, 48, 96, 144, 160) else 512)
config = deep_ep.Config(num_sms, 8, nvl_buffer_size, 16, rdma_buffer_size)

# Test dispatch
Expand Down
3 changes: 2 additions & 1 deletion tests/test_low_latency.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -114,7 +114,8 @@ def test_main(num_tokens: int, hidden: int, num_experts: int, num_topk: int,
if do_check:
diff = calc_diff(current_x * topk_weights.masked_fill(topk_idx == -1, 0).sum(dim=1).view(-1, 1), combined_x)
assert torch.isnan(combined_x).sum().item() == 0
assert diff < (9e-4 if dispatch_use_fp8 else 1e-5), f'Error: {diff=}, {dispatch_use_fp8=}, {zero_copy=}'
if not round_scale:
assert diff < (9e-4 if dispatch_use_fp8 else 1e-5), f'Error: {diff=}, {dispatch_use_fp8=}, {zero_copy=}'
hash_value ^= hash_tensor(combined_x)

# noinspection PyShadowingNames
Expand Down