Kealan Barbieri
Kealan Barbieri
# Description Address performance regressions in batched gemm with grouped attrs. Previously disabled 3d->2d reshape to avoid correctness issues, in some cases we can instead reshape grouped dims to preserve...
# Description Enable scaling for batched gemms that are not reshaped down to 2d. Fixes # MFDNN-13705 # Checklist ## General - [ ] Do all unit and benchdnn tests...
# Description Permit decompression of fp4 weights in GEMM, reusing strategies for int4. Fixes # MFDNN-13543 # Checklist ## General - [ ] Do all unit and benchdnn tests (`make...
# Description Expands cases where inputs are decompressed to include integer activations as well as weights. Integer weights are still required but in the presence of fpmath setting activations will...
# Description - update with latest KPIs from verbose logs - strip all layers with dim > 20k by commenting out This should resolve some remaining timeouts in nightly testing....
# Description Use strategy values rather than driver_info which may be uninitialized. # Checklist ## General - [ ] Do all unit and benchdnn tests (`make test` and `make test_benchdnn_*`)...
# Description Issue [MFDNN-14456](https://jira.devtools.intel.com/browse/MFDNN-14456) shows group dims being used with common scales. This behavior is disabled in benchdnn but not explicitly via API. This commit adds an explicit check to...
This reverts commit d5e144c9432aaeae4f77f214c355c5f580f2fb7a. # Description Revert change pending resolution of [GSD-11900](https://jira.devtools.intel.com/browse/GSD-11900). Currently the desired effect of this change is disabled at the driver level. Fixes # (github issue) #...
# Description Update fwks inputs, including adding `skip-impl=ref` to see if it helps testing timeouts. # Checklist ## General - [ ] Do all unit and benchdnn tests (`make test`...