Fix FMHA fp8 hdim=64 incorrect result in MI200 (#3423)

* Fix incorrect result in hdim=64

* Add change log
This commit is contained in:
rocking
2025-12-18 00:16:54 +08:00
committed by GitHub
parent e67cd7edeb
commit 292f87aa03
2 changed files with 8 additions and 3 deletions

View File

@@ -1014,8 +1014,12 @@ class KernelComponentFactoryGfx9(CompatibilityRuleFactoryGfx9):
["no"],
["f", "t"],
):
pipelines.append(FmhaFwdPipeline("qr_async", "row", "t", "f", "t", "t", logits, bias, "f", "f", qscale, mask, "f", "f", sink)) # fmt: skip
pipelines.append(FmhaFwdPipeline("qr_async", "row", "t", "t", "t", "t", logits, bias, "f", "f", qscale, mask, "f", "f", sink)) # fmt: skip
if hdim == 64:
pipelines.append(FmhaFwdPipeline("qr", "row", "t", "f", "t", "t", logits, bias, "f", "f", qscale, mask, "f", "f", sink)) # fmt: skip
pipelines.append(FmhaFwdPipeline("qr", "row", "t", "t", "t", "t", logits, bias, "f", "f", qscale, mask, "f", "f", sink)) # fmt: skip
else:
pipelines.append(FmhaFwdPipeline("qr_async", "row", "t", "f", "t", "t", logits, bias, "f", "f", qscale, mask, "f", "f", sink)) # fmt: skip
pipelines.append(FmhaFwdPipeline("qr_async", "row", "t", "t", "t", "t", logits, bias, "f", "f", qscale, mask, "f", "f", sink)) # fmt: skip
elif dtype in ["fp8", "fp8fp16", "bf8"]:
# TODO
pass