forked from ccf-ai-infra/TileOPs-Metax
da567b4239
Remove the `stage=` builder parameter, the read_only/read_shared/read probe branches, the _block_sum_reduce macro, and stage_kernel() from the fused SwiGLU + per-channel FP8-cast kernel. These existed only to attribute fused time to pipeline stages for profiling and had no effect on the production path; the unconditional kernel body is byte-identical to the former stage="full". lru_cache returns to default maxsize=32. Delete scripts/profile_quant_swiglu_stages.py and scripts/quant_swiglu_stage_profile.sh, the only consumers. Verified: tests/ops/test_quant_swiglu_channel_cast_transpose.py 12 passed (bit-exact fp8 contract preserved). Co-Authored-By: Claude <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| kernels | ||
| manifest | ||
| ops | ||
| perf | ||
| testing | ||
| trace | ||
| utils | ||
| __init__.py | ||