[Performance][DSR1]: Fused RoPE+KVCache+q_concat for MLA (#40392)

Signed-off-by: Rohan138 <rohanpotdar138@gmail.com>
Signed-off-by: Rohan Potdar <66227218+Rohan138@users.noreply.github.com>
Co-authored-by: ElizaWszola <ewszola@redhat.com>
This commit is contained in:
Rohan Potdar
2026-05-11 14:10:50 +00:00
committed by GitHub
co-authored by ElizaWszola
parent 8415bf2cdb
commit a51376b3f0
12 changed files with 966 additions and 109 deletions
+10 -1
View File
@@ -136,8 +136,10 @@ class PassConfig:
"""Enable flashinfer allreduce fusion."""
fuse_minimax_qk_norm: bool = None # type: ignore[assignment]
"""Enable fused allreduce+RMSNorm for MiniMax QK norm."""
enable_qk_norm_rope_fusion: bool = False
enable_qk_norm_rope_fusion: bool = None # type: ignore[assignment]
"""Enable fused Q/K RMSNorm + RoPE pass."""
fuse_rope_kvcache_cat_mla: bool = None # type: ignore[assignment]
"""Enable fused MLA KV cache update with RoPE."""
# ROCm/AITER specific fusions
fuse_act_padding: bool = None # type: ignore[assignment]
@@ -228,6 +230,7 @@ class PassConfig:
"fuse_act_padding",
"fuse_mla_dual_rms_norm",
"fuse_rope_kvcache",
"fuse_rope_kvcache_cat_mla",
mode="wrap",
)
@classmethod
@@ -285,6 +288,12 @@ class PassConfig:
"The fusion will be disabled."
)
self.fuse_rope_kvcache = False
if self.fuse_rope_kvcache_cat_mla and not current_platform.is_cuda_alike():
logger.warning_once(
"MLA KV cache update with RoPE fusion enabled but the "
"current platform is not CUDA or ROCm. The fusion will be disabled."
)
self.fuse_rope_kvcache_cat_mla = False
def log_enabled_passes(self) -> None:
"""