[Bugfix] Support CuTe DSL 4.8.0 block-scale API (#59480)
Signed-off-by: hjjq <50634613+hjjq@users.noreply.github.com> Co-authored-by: Codex <noreply@openai.com> Co-authored-by: Shang Wang <samshang.wang@mail.utoronto.ca>
Hanjie Qiu committed
Sep 30, 2026 at 22:30 UTC
bc505fc823b29739554e40c348d12b4f032f8369
1 file changed
+6
-1
vllm/cute_utils/_tcgen05.py
+6
-1
@@ -148,6 +148,11 @@ def mma_mxfp8(
148
loc=None,
149
ip=None,
150
) -> None:
151
+ # CuTe DSL 4.8 renamed the block-scale enum and keyword.
152
+ if hasattr(nvvm, "Tcgen05MMABlockScale"):
153
+ scale_kwargs = {"block_scale": nvvm.Tcgen05MMABlockScale.DEFAULT}
154
+ else:
155
+ scale_kwargs = {"scale_vec_size": nvvm.Tcgen05MMAScaleVecSize.X1}
156
with cute.arch.elect_one():
157
nvvm.tcgen05_mma_block_scale(
158
nvvm.Tcgen05MMAKind.MXF8F6F4,
@@ -159,9 +164,9 @@ def mma_mxfp8(
164
Boolean(enable_input_d).ir_value(loc=loc, ip=ip),
165
_make_tmem_llvm_ptr(sfa_tmem, loc=loc, ip=ip),
166
_make_tmem_llvm_ptr(sfb_tmem, loc=loc, ip=ip),
162
- scale_vec_size=nvvm.Tcgen05MMAScaleVecSize.X1,
167
loc=loc,
168
ip=ip,
169
+ **scale_kwargs,
170
)
171
172