A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[CI] Fix broken compile tests due to unsupported SiluMul+Nvfp4Quant fusion (#23973)
Signed-off-by: Yong Hoon Shin <yhshin@meta.com> Co-authored-by: Roger Wang <hey@rogerw.io>
Yong Hoon Shin committed
Aug 30, 2025 at 00:14 UTC
9748c5198b492e22dc24d6eb455ec907369392f3
1 file changed
+6
-4
vllm/compilation/activation_quant_fusion.py
+6
-4
@@ -29,8 +29,9 @@ SILU_MUL_OP = torch.ops._C.silu_and_mul.default
29
FUSED_OPS: dict[QuantKey, OpOverload] = {
30
kFp8StaticTensorSym: torch.ops._C.silu_and_mul_quant.default, # noqa: E501
31
}
32
-if current_platform.is_cuda() and hasattr(torch.ops._C,
33
- "silu_and_mul_nvfp4_quant"):
32
+silu_and_mul_nvfp4_quant_supported = (current_platform.is_cuda() and hasattr(
33
+ torch.ops._C, "silu_and_mul_nvfp4_quant"))
34
+if silu_and_mul_nvfp4_quant_supported:
35
FUSED_OPS[
36
kNvfp4Quant] = torch.ops._C.silu_and_mul_nvfp4_quant.default # noqa: E501
37
@@ -171,8 +172,9 @@ class ActivationQuantFusionPass(VllmInductorPass):
172
pattern_silu_mul_fp8 = SiluMulFp8StaticQuantPattern()
173
pattern_silu_mul_fp8.register(self.patterns)
174
174
- pattern_silu_mul_nvfp4 = SiluMulNvfp4QuantPattern()
175
- pattern_silu_mul_nvfp4.register(self.patterns)
175
+ if silu_and_mul_nvfp4_quant_supported:
176
+ pattern_silu_mul_nvfp4 = SiluMulNvfp4QuantPattern()
177
+ pattern_silu_mul_nvfp4.register(self.patterns)
178
179
def __call__(self, graph: torch.fx.Graph):
180
self.begin()