ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-03 13:11:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

Revert " Fp8 block scale quantization for fmha fwd (#3330)" (#3633)

Browse Source 
This reverts commit dd0b4294af.
This commit is contained in:
Po Yen Chen
2026-01-23 13:21:19 +08:00
committed by GitHub
parent f30d04654e
commit de5a1d730d
14 changed files with 84 additions and 667 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
example/ck_tile/01_fmha/codegen/cpp_symbol_map.py
View File

@@ -77,13 +77,11 @@ def get_mask_cpp_check_expr(mask: str) -> str:
QSCALE_MAP = {
"no": "ck_tile::BlockAttentionQuantScaleEnum::NO_SCALE",
"pertensor": "ck_tile::BlockAttentionQuantScaleEnum::PERTENSOR",
"blockscale": "ck_tile::BlockAttentionQuantScaleEnum::BLOCKSCALE",
}
QSCALE_CHECK_MAP = {
"no": "quant_scale_enum::no_scale",
"pertensor": "quant_scale_enum::pertensor",
"blockscale": "quant_scale_enum::blockscale",
}
BIAS_MAP = {

7
example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py
View File

@@ -1024,7 +1024,7 @@ class KernelComponentFactoryGfx9(CompatibilityRuleFactoryGfx9):
# no need lse/dropout kernels
for logits, qscale, mask, bias, sink in itertools.product(
["t", "f"],
["no", "pertensor", "blockscale"],
["no", "pertensor"],
get_mask_map(mask_impl).keys(),
["no"],
["f", "t"],
@@ -1152,10 +1152,7 @@ class KernelComponentFactoryGfx12(CompatibilityRuleFactory):
elif dtype in cls._DT_FP8_FP8BF16 or dtype in cls._DT_FP8FP32:
# no need lse/dropout kernels
for logits, qscale, mask, bias in itertools.product(
["f"],
["no", "pertensor", "blockscale"],
get_mask_map(mask_impl).keys(),
["no"],
["f"], ["no", "pertensor"], get_mask_map(mask_impl).keys(), ["no"]
):
pipelines.append(FmhaFwdPipeline("qr", "row", "f", "f", "f", "f", logits, bias, "f", "f", qscale, mask, "f", "f", "f")) # fmt: skip
pipelines.append(FmhaFwdPipeline("qr", "row", "t", "t", "t", "t", logits, bias, "f", "f", qscale, mask, "f", "f", "f")) # fmt: skip