[GH-ISSUE #300] Why we don't give normal ep-128 test results ? #213

New issue

Closed

opened 2026-10-26 09:56:43 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 09:56:43 +00:00

Owner

Originally created by @wenzt on GitHub (Jul 14, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/300

When I test 128 ranks, ASSERT raised at:
EP_HOST_ASSERT(num_max_nvl_chunked_recv_tokens / num_rdma_ranks > std::max(num_max_rdma_chunked_send_tokens, num_max_nvl_chunked_send_tokens));

How to set parameters to get best perference ?

num_max_nvl_chunked_send_tokens
num_max_nvl_chunked_recv_tokens
num_max_rdma_chunked_send_tokens
num_max_rdma_chunked_recv_tokens

Originally created by @wenzt on GitHub (Jul 14, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/300 When I test 128 ranks, ASSERT raised at: `EP_HOST_ASSERT(num_max_nvl_chunked_recv_tokens / num_rdma_ranks > std::max(num_max_rdma_chunked_send_tokens, num_max_nvl_chunked_send_tokens));` How to set parameters to get best perference ? 1. num_max_nvl_chunked_send_tokens 2. num_max_nvl_chunked_recv_tokens 3. num_max_rdma_chunked_send_tokens 4. num_max_rdma_chunked_recv_tokens

bsakaguchi closed this issue

2026-10-26 09:56:43 +00:00

bsakaguchi referenced this issue

2026-10-26 10:01:06 +00:00

[PR #213] [MERGED] Add automatic warp count control for low-latency kernels #469