35
35
"--iree-codegen-gpu-native-math-precision=true" ,
36
36
"--iree-rocm-waves-per-eu=2" ,
37
37
"--iree-flow-inline-constants-max-byte-length=1" ,
38
- "--iree-preprocessing-pass-pipeline=builtin.module(iree-preprocessing-transpose-convolution-pipeline, util.func(iree-preprocessing-pad-to-intrinsics))" ,
38
+ "--iree-preprocessing-pass-pipeline=builtin.module(iree-preprocessing-transpose-convolution-pipeline, iree-global-opt-raise-special-ops, util.func(iree-preprocessing-pad-to-intrinsics, iree-linalg-ext-pad-attention{pad-to-multiple-of=0,128,0,32,0} ))" ,
39
39
],
40
40
"unet" : [
41
41
"--iree-flow-enable-aggressive-fusion" ,
@@ -275,7 +275,7 @@ def create_safe_name(hf_model_name, model_name_str):
275
275
276
276
277
277
def get_mfma_spec_path (target_chip , save_dir ):
278
- url = "https://raw.githubusercontent.com/iree-org/iree/main/build_tools/pkgci/external_test_suite/attention_and_matmul_spec .mlir"
278
+ url = "https://sharkpublic.blob.core.windows.net/sharkpublic/specs/latest/attention_and_matmul_spec_gfx942 .mlir"
279
279
attn_spec = urlopen (url ).read ().decode ("utf-8" )
280
280
spec_path = os .path .join (save_dir , "attention_and_matmul_spec_mfma.mlir" )
281
281
if os .path .exists (spec_path ):
@@ -287,9 +287,9 @@ def get_mfma_spec_path(target_chip, save_dir):
287
287
288
288
def get_wmma_spec_path (target_chip , save_dir ):
289
289
if target_chip == "gfx1100" :
290
- url = "https://github.com/iree-org/iree/raw/shared/tresleches-united/scripts/attention_gfx1100.spec .mlir"
290
+ url = "https://sharkpublic.blob.core.windows.net/sharkpublic/specs/latest/attention_and_matmul_spec_gfx1100 .mlir"
291
291
elif target_chip in ["gfx1103" , "gfx1150" ]:
292
- url = "https://github.com/iree-org/iree/raw/shared/tresleches-united/scripts/attention_gfx1103.spec .mlir"
292
+ url = "https://sharkpublic.blob.core.windows.net/sharkpublic/specs/latest/attention_and_matmul_spec_gfx1150 .mlir"
293
293
else :
294
294
return None
295
295
attn_spec = urlopen (url ).read ().decode ("utf-8" )
0 commit comments