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 8, 03:45 UTC
[torch.compile] Turn on silu+fp4 quant fusion by default for O1+ (#34718)
Signed-off-by: Luka Govedič <lgovedic@redhat.com>
Luka Govedič committed
Feb 17, 2026 at 22:29 UTC
02e8f26ceaa3af0382b9de6b40825c4ad49ef5b7
1 file changed
+10
-5
vllm/config/vllm.py
+10
-5
@@ -95,11 +95,16 @@ def enable_norm_fusion(cfg: "VllmConfig") -> bool:
95
96
97
def enable_act_fusion(cfg: "VllmConfig") -> bool:
98
- """Enable if either SiLU+Mul or quant FP8 custom op is active;
99
- otherwise Inductor handles fusion."""
100
- return cfg.compilation_config.is_custom_op_enabled(
101
- "silu_and_mul"
102
- ) or cfg.compilation_config.is_custom_op_enabled("quant_fp8")
98
+ """
99
+ Enable if either SiLU+Mul or quant FP8 custom op is active;
100
+ otherwise Inductor handles fusion.
101
+ Also enable for FP4 models as FP4 quant is always custom so Inductor cannot fuse it.
102
+ """
103
+ return (
104
+ cfg.compilation_config.is_custom_op_enabled("silu_and_mul")
105
+ or cfg.compilation_config.is_custom_op_enabled("quant_fp8")
106
+ or (cfg.model_config is not None and cfg.model_config.is_nvfp4_quantized())
107
+ )
108
109
110
def enable_allreduce_rms_fusion(cfg: "VllmConfig") -> bool: