mirror of
https://github.com/deepseek-ai/DeepEP
synced 2025-06-26 18:28:11 +00:00
Support BF16 for low-latency kernels
This commit is contained in:
@@ -128,7 +128,7 @@ struct LowLatencyLayout {
|
||||
|
||||
// Message sizes
|
||||
EP_HOST_ASSERT(num_scales * sizeof(float) <= hidden);
|
||||
size_t num_bytes_per_dispatch_msg = hidden + num_scales * sizeof(float) + sizeof(int4);
|
||||
size_t num_bytes_per_dispatch_msg = sizeof(int4) + std::max(hidden * sizeof(nv_bfloat16), hidden + num_scales * sizeof(float));
|
||||
size_t num_bytes_per_combine_msg = sizeof(int4) + hidden * sizeof(nv_bfloat16);
|
||||
|
||||
// Send buffer
|
||||
|
||||
Reference in New Issue
Block a user