Expand description
Wall-clock phase counters for the bake profiler (CMF_BAKE_PROF=1 prints them). Written from worker threads, so atomics; nanoseconds.
Statics§
- ATTN_
CHAIN - Inside ATTN_FWD: the device chain call.
- ATTN_
FWD - BWD
- FFN_ACT
- FFN_
CHAIN - Inside FFN_FWD: the device chain call, the plane split + host activation, and the host (unfused) GEMM path.
- FFN_FWD
- FFN_
HOST - FFN_
POST - FFN_PRE
- Before the chain (residual add, post-norm, buffers) and after it (residual add, acts struct).
- GEMM
- GEMM_
CALLS - SHAPES
- (n, k, m) → (calls, ns). A handful of shapes dominate a training step; naming them decides which ones are worth a GPU submit and which are cheaper computed in place.
Functions§
- add
- gemm_
shape - shape_
report - take
- Read and reset, in seconds.