DeepSeek/DeepEP
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepEP synced 2025-06-26 18:28:11 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Optimize intranode combine. (#247)

Browse Source 
* Increase the test round.

* Add warp synchronization.

* Shuffle the send warps.

* Add time elapsed into bench result.
This commit is contained in:
Shangyan Zhou 2025-06-24 09:10:23 +08:00 committed by GitHub
parent fbcf430006
commit 9eb2f84b3e
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 8 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
csrc/kernels/intranode.cu
View File

@ -618,8 +618,8 @@ combine(dtype_t* recv_x, float* recv_topk_weights,
const auto num_threads_per_rank = num_send_warps_per_rank * 32;
const auto send_thread_id = thread_id;
const auto send_warp_id = send_thread_id / 32;
const auto send_rank_id = thread_id / num_threads_per_rank;
const auto send_warp_id_in_rank = send_warp_id % num_send_warps_per_rank;
const auto send_rank_id = (responsible_channel + send_warp_id) % kNumRanks;
const auto send_warp_id_in_rank = send_warp_id / kNumRanks;
EP_STATIC_ASSERT(num_send_warps * 32 == kNumThreads, "Invalid warp count");
// Calculate pointers by the specific layout
@ -777,7 +777,7 @@ combine(dtype_t* recv_x, float* recv_topk_weights,
expected_head = ld_nc_global(send_head + token_idx * kNumRanks + lane_id);
auto start_time = clock64();
while (channel_tail_idx[lane_id] <= expected_head and expected_head >= 0) {
while (__any_sync(0xffffffff, channel_tail_idx[lane_id] <= expected_head and expected_head >= 0)) {
// Timeout check
if (clock64() - start_time > NUM_TIMEOUT_CYCLES) {
printf("DeepEP timeout for combine receivers, rank %d, responsible_channel = %d, expect = %d\n", rank, responsible_channel, expected_head);

8
tests/test_intranode.py
View File

@ -184,9 +184,9 @@ def test_main(num_sms: int, local_rank: int, num_ranks: int, rank: int, buffer:
best_time, best_results = t, (num_sms, nvl_chunk_size)
if local_rank == 0:
print(f'[tuning] SMs {num_sms}, NVL chunk {nvl_chunk_size if nvl_chunk_size else "default"}: '
f'{nvl_recv_bytes / 1e9 / t:.2f} GB/s (NVL) ', flush=True)
f'{nvl_recv_bytes / 1e9 / t:.2f} GB/s (NVL), avg_t: {t * 1e6:.2f} us', flush=True)
if local_rank == 0:
print(f'[tuning] Best dispatch ({"FP8" if isinstance(current_x, tuple) else "BF16"}): SMs {best_results[0]}, NVL chunk {best_results[1]}, {nvl_recv_bytes / 1e9 / best_time:.2f} GB/s (NVL)', flush=True)
print(f'[tuning] Best dispatch ({"FP8" if isinstance(current_x, tuple) else "BF16"}): SMs {best_results[0]}, NVL chunk {best_results[1]}, {nvl_recv_bytes / 1e9 / best_time:.2f} GB/s (NVL), t: {best_time * 1e6:.2f} us', flush=True)
print('', flush=True)
# Gather the best config from rank 0 and the first test setting
@ -215,12 +215,12 @@ def test_main(num_sms: int, local_rank: int, num_ranks: int, rank: int, buffer:
t = bench(lambda: buffer.combine(**tune_args))[0]
if local_rank == 0:
print(f'[tuning] SMs {num_sms}, NVL chunk {nvl_chunk_size if nvl_chunk_size else "default"}: '
f'{combine_bf16_nvl_send_bytes / 1e9 / t:.2f} GB/s (NVL) ', flush=True)
f'{combine_bf16_nvl_send_bytes / 1e9 / t:.2f} GB/s (NVL), avg_t: {t * 1e6:.2f} us', flush=True)
if t < best_time and nvl_chunk_size > 0:
best_time, best_results = t, (num_sms, nvl_chunk_size)
if local_rank == 0:
print(f'[tuning] Best combine: SMs {best_results[0]}, NVL chunk {best_results[1]}: {combine_bf16_nvl_send_bytes / 1e9 / best_time:.2f} GB/s (NVL)', flush=True)
print(f'[tuning] Best combine: SMs {best_results[0]}, NVL chunk {best_results[1]}: {combine_bf16_nvl_send_bytes / 1e9 / best_time:.2f} GB/s (NVL), t: {best_time * 1e6:.2f} us', flush=True)
print('', flush=True)

2
tests/utils.py
View File

@ -80,7 +80,7 @@ def create_grouped_scores(scores: torch.Tensor, group_idx: torch.Tensor, num_gro
return (scores * mask).view(num_tokens, num_experts)
def bench(fn, num_warmups: int = 20, num_tests: int = 30, post_fn=None):
def bench(fn, num_warmups: int = 50, num_tests: int = 50, post_fn=None):
# Flush L2 cache with 256 MB data
torch.cuda.synchronize()
cache = torch.empty(int(256e6 // 4), dtype=torch.int, device='cuda')