[Kernel] Porting the TRTLLM minimax_allreduce_rms kernels (#37045)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
Jee Jee Li
2026-04-11 00:20:20 +08:00
committed by GitHub
parent 8f121f7879
commit ecd1ea1363
14 changed files with 1861 additions and 4 deletions
+2
View File
@@ -134,6 +134,8 @@ class PassConfig:
"""Enable async TP."""
fuse_allreduce_rms: bool = None # type: ignore[assignment]
"""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 fused Q/K RMSNorm + RoPE pass."""
+16
View File
@@ -1627,6 +1627,22 @@ class VllmConfig:
compile_range_end,
)
if compilation_config.pass_config.fuse_minimax_qk_norm:
from vllm.compilation.passes.fusion.minimax_qk_norm_fusion import (
MAX_TOKEN_NUM,
)
max_token_num = min(
MAX_TOKEN_NUM, self.scheduler_config.max_num_batched_tokens
)
if compile_range_end is not None and max_token_num < compile_range_end:
computed_compile_ranges_endpoints.append(max_token_num)
else:
logger.debug(
"Max num batched tokens below MiniMax QK norm fusion threshold, "
"MiniMax QK norm fusion enabled for all num_tokens."
)
if compilation_config.compile_ranges_endpoints is not None:
for x in compilation_config.compile_ranges_endpoints:
assert isinstance(x, int)