Use aiter triton fused_add_rmsnorm_pad for gpt-oss (#30976)

Signed-off-by: Rohan138 <rohanpotdar138@gmail.com>
This commit is contained in:
Rohan Potdar
2026-01-28 20:47:47 +00:00
committed by GitHub
parent 3e440786af
commit 59bcc5b6f2
9 changed files with 327 additions and 11 deletions
+6 -2
View File
@@ -18,8 +18,9 @@ from .vllm_inductor_pass import VllmInductorPass
if rocm_aiter_ops.is_enabled():
from vllm.compilation.rocm_aiter_fusion import (
RocmAiterRMSNormFusionPass,
RocmAiterRMSNormQuantFusionPass,
RocmAiterSiluMulFp8GroupQuantFusionPass,
RocmAiterTritonAddRMSNormPadFusionPass,
)
if current_platform.is_cuda_alike():
@@ -123,13 +124,16 @@ class PostGradPassManager(CustomGraphPass): # type: ignore[misc]
self.passes += [RMSNormQuantFusionPass(config)]
if rocm_aiter_ops.is_enabled():
self.passes += [
RocmAiterRMSNormFusionPass(config),
RocmAiterRMSNormQuantFusionPass(config),
]
if self.pass_config.fuse_act_quant:
self.passes += [ActivationQuantFusionPass(config)]
if rocm_aiter_ops.is_enabled():
self.passes += [RocmAiterSiluMulFp8GroupQuantFusionPass(config)]
if self.pass_config.fuse_act_padding and rocm_aiter_ops.is_enabled():
self.passes += [RocmAiterTritonAddRMSNormPadFusionPass(config)]
if self.pass_config.fuse_attn_quant:
self.passes += [AttnFusionPass(config)]